01
Summary
Almost every AI rule favors whoever can already comply with it when it is written. The economist George Stigler argued this for regulation in general, and the Centre for European Policy Studies (CEPS) estimated, for the proposed EU AI Act, a compliance cost that weighs more heavily on small and medium-sized enterprises [3] [4]. Those rules come from treaties, laws, sector regulators and technical standards, and also from industry commitments and private contracts. Since 2025, Anthropic has excluded companies controlled from China, wherever they operate, and its policy also applies when the model is sold through a cloud [5] [6]. Microsoft cut off its services to a unit of Israel’s Ministry of Defense because it violated its terms [7]. SK hynix has already committed more than half of its 2027 and 2028 memory output in long-term contracts [8].
An open model run on one’s own hardware escapes the rules of the lab and the cloud, but not those of the chip, because running it requires getting chips, and their sale is subject to export controls and supply contracts. According to Girish Sastry, Lennart Heim and other authors, the chip is the input best suited to control, because it is a physical object that can be counted and tracked, and very few companies make it [9].
The AI chain runs from the machines that make chips to the person who uses an assistant, and in its first links one or a few suppliers decide. ASML is the only maker of the most advanced lithography machines, the ones that use light to print chip circuits onto silicon. TSMC makes the chips of NVIDIA and AMD and Microsoft’s own chip, and NVIDIA buys its memory from three companies [10] [11] [12] [13]. NVIDIA also coordinates investment along that chain. It asks its suppliers for the capacity it will need, it has committed to buying US$279 billion from them, and it backs its buyers’ credit [14] [15] (see “Who regulates AI” and “The chain around AI”).
Jensen Huang, NVIDIA’s founder and chief executive, resigned on November 11, and the company announced it on the 12th. Three days earlier, the company’s board had opened an investigation into reports that Huang had been the target of a foreign intelligence effort [16] [17]. The day after the resignation, three committees of the U.S. House of Representatives asked him for a closed-door transcribed interview on December 1, under threat of a subpoena. They asked for another interview with NVIDIA’s general counsel, about chip smuggling and NVIDIA’s “interference” in export policy [18]. Both agreed to appear voluntarily, on December 1 and 2 [19].
According to the New York Times and the Washington Post, Marco Rubio, U.S. Secretary of State and national security adviser, contributed to the fall. In the spring there was a dispute within the government over chip sales to China. In the middle of it, Rubio oversaw a memorandum on Jensen Huang’s relationship with Alice Huang, an executive at Megaspeed, a Singapore company that is an NVIDIA customer. U.S. agencies point to Alice Huang as passing information to the Chinese government [20] [21].
Behind the personal scandal lies China’s access to NVIDIA’s chips. Five days after the resignation, the U.S. Department of Commerce released the rule that requires a license from data centers in any country that provide large amounts of compute to Chinese companies and to companies from other countries under arms embargo. It is the same rule that eliminates the export control on weights [22]. There is no sign that the dispute between open and closed models caused the fall, but the fall does affect that dispute (see “Jensen Huang’s fall”).
NVIDIA is today the largest contributor of open models among the big tech companies, measured by published repositories, which are the online folders where each model is published with its files, according to Hugging Face’s March report, and Meta also contributes a significant share [23]. NVIDIA trains open models and agreed to buy Hugging Face, the site where they are published. Huang, moreover, declared himself a co-author of the July letter in favor of open weights. The company also funds closed models, and with larger amounts [24] [25].
With Huang gone, the open coalition loses its most visible voice before the U.S. government and Congress just as the only bill devoted expressly to supporting open models, the Open-Source AI Leadership Act (H.R. 10152), is pending. The House Committee on Energy and Commerce approved it in September, but it has not filed its report, the bill has not reached the floor, and it dies on January 3, when the Congress changes [26]. Other bills have also been introduced against Chinese models and against distillation, which consists of training a model on the responses of another (see “How the law treats each model”). The resignation also puts the chain’s coordination to the test. According to Huang himself, that coordination rested on the trust suppliers had in him, and SemiAnalysis, an independent semiconductor research firm, argues that without him they will not give NVIDIA the same priority [27] (see “NVIDIA orders the chain”).
Most of the leading Chinese labs publish their flagship models openly. With them, China dominates downloads and derivative models, the ones others build from an open model [28] [29]. The U.S. government sponsors both sides. It buys and evaluates closed models, and its action plan calls for an environment favorable to open ones and leaves the decision to publish to each developer [30]. Open models trail closed ones by four to eight months, depending on who measures and what [31] [32] [33] [34]. On Arena, a leaderboard where users vote for the better of two responses without knowing which model gave each one, the best closed model’s score exceeded the best open model’s by 0.5% in August 2024 and by 3.3% in March 2026. The gap is still small, but it grew [35]. In cybersecurity, by contrast, the gap narrowed [33].
China is also replacing NVIDIA’s chips. Huawei sold US$7.5 billion in AI chips in 2025 and expects to sell about US$12 billion this year. According to IDC, Huawei had 28% of the AI chips for data centers sold in China in the first half of 2026, against 34% for NVIDIA, and some Chinese labs are starting to train their models on those chips [36] [37] [38]. What holds China back now is chip manufacturing, which has replaced design as its bottleneck, because export controls deny it lithography machines and high-bandwidth memory, the kind AI chips use because it feeds the processor far more data per second than ordinary memory [39] [40].
The models that drive robots are published mostly in the United States: NVIDIA opens its own, some under its own licenses and others under a third party’s, while Google and Figure keep theirs. The bodies are made mostly in China, whose manufacturers made more than 97% of the humanoids shipped in the first half of 2026, according to Smart Analytics Global [41]. In July, the United States closed its market to new models of foreign-made mobile robots [42] (see “Robots”).
In quantum computing, the software is open and the machines are closed. The technology is advancing fast. These machines compute with qubits, units that, unlike a bit, can be in a combination of 0 and 1, and that fail easily. The United States and China already correct those errors in a way that adding qubits reduces them, and since 2025 calculations have been published that no ordinary computer can perform and that others can verify [43] [44]. A large quantum machine could break public-key cryptography, the kind that makes it possible to encrypt a connection or sign a file without first agreeing on a secret key. That cryptography protects connections and the signatures that prove a model or its data are the originals, and the U.S. government has already ordered its agencies to replace it by the end of 2031 at the latest [45]. That quantum computing improves AI has been shown in theory and on problems built to suit it; with real machines, experiments do not yet beat the best classical methods on AI tasks [46] [47] (see “Quantum computing”).
Alongside business and the rivalry between the United States and China, there is a dispute of ideas about how to develop AI. One position proposes concentrating it in a few labs that control it. The other proposes spreading it so that anyone can use it, study it and modify it. Arvind Narayanan and Sayash Kapoor, of Princeton, attribute that dispute to deep differences in worldview [48]. The same arguments are heard in California, in Beijing and in Paris, and they divide companies within the same country (see “Two ideas about how to develop AI”).
In business, the dispute repeats that of software. IBM, Meta, NVIDIA and Alibaba open their models because they get paid in another layer: consulting, advertising, chips or the cloud, which rents out the compute of its data centers to others. That is how Nathan Lambert explains NVIDIA’s bet, and how the Indian think tank Takshashila explains China’s [49] [50].
We believe that in intellectual property, history does not repeat itself. Software patents were neutralized with licenses that grant the user the use of the patents and withdraw it from anyone who sues over them, so suing stops being in the interest of whoever uses the software. Rights over training data cannot be neutralized that way, because their owners, such as authors and publishers, do not depend on the model’s license and lose nothing by suing. Where courts require paying for those rights, we believe the cost favors the large labs, open or closed, and whoever trains where the law is most permissive (see “Who gains from opening and who from closing”).
The money in AI stays today in the scarce links of the chain. For every dollar they sell, memory makers earn between 72 and 80 cents in operating profit, NVIDIA earns 66, TSMC 59, and the clouds of Amazon and Google between 36 and 39 [8] [51] [52] [14] [53] [54] [55].
The labs, by contrast, lose money. Anthropic is the only frontier lab that publishes its accounts, and frontier labs are the ones that make the most capable models. Anthropic still loses money on its operations, and it has committed to buying about US$900 billion in compute, largely from Google, Amazon and Microsoft, which are also its shareholders [56]. OpenAI, according to documents cited by the press, loses more on its operations than it takes in. The same happens to the open labs that publish their accounts, such as Zhipu and MiniMax [57] [58].
Open models are sustained as an expense that serves to sell something else, with capital, which is often from the state, or as a public good. Nathan Lambert describes NVIDIA funding open models to sustain its chip business, and Takshashila describes the Chinese state absorbing the cost of its own [49] [50]. In free software, according to Dries Buytaert, the product is the hosting, that is, running the program on one’s own servers and charging for its use, rather than the code. Something similar happens with open models: the business goes mostly to whoever serves the model, that is, whoever runs it on their servers and charges for each use, rather than to whoever publishes it [59] (see “Where the money stays”).
AI makes writing code cheaper and puts pressure on software sold by license or by user. Microsoft grew 35% in two years with fewer engineers and charges more and more for consumption of its cloud [60] [61]. Some business application companies, such as Intuit, Oracle, ServiceNow and Adobe, lost between a third and almost half of their stock market value in 2026, although the sector’s exchange-traded fund, which bundles many software companies into a single share, almost recovered what it lost [62]. Free software, for its part, gains users while it loses those who paid for it [63] (see “The software industry facing AI”).
Large companies, rather than choosing between open and closed models, prohibit their suppliers from training on their data, storing it or using it without written authorization [64] [65]. Some banned external tools. Banks built their own portals on closed models, and some companies with research capacity trained their own models from scratch. We believe each stance favors a different supplier. The portal favors the closed labs that run inside it. The ban favors the covert use of free versions. Zero retention, which is the commitment not to keep anything the client sends, and sovereignty, which is not depending on a foreign supplier, favor those who sell installable open models. In almost every case, whoever sells the compute wins, whether the cloud or NVIDIA (see “What large companies allow with their data”).
The law treats each model differently on four fronts. On exports, the Chinese proposal restricts the largest closed weights and exempts published ones. The United States, by contrast, eliminated its equivalent control with the November 16 rule. Once it is published on the 19th, that rule will control the compute that reaches companies in China and in other countries under arms embargo.
On evaluation and transparency, the European Union exempts openly licensed models from its documentation obligations, except the most powerful ones, which it considers to pose systemic risk. On liability, the new European product liability directive leaves out the developer who publishes outside a commercial activity, but holds liable as a manufacturer whoever substantially modifies the model and places it on the market. On market access, the United States removes DeepSeek’s models from federal systems, including third-party products built with them.
Three other fronts weigh on both kinds of model, though not equally. The rules on synthetic content require whoever provides the service to label it and leave almost untouched the open model that runs on the user’s hardware, although California and the United Kingdom already reach whoever hosts or supplies the weights. Competition authorities, for their part, see openness as a factor in favor of competition, but have not used it to decide any case. Meanwhile, two closed labs account for most business spending, according to a survey by Menlo Ventures, an Anthropic investor, and NVIDIA agreed to buy the main distribution site for open models [66].
The third of those fronts is the dispute over who regulates in the United States, the federal government or the states, which does not line up with the dispute between open and closed. The new Congress, which takes office in January with a Democratic majority, makes it harder for a federal law to override state laws, because, according to the law firm Covington & Burling, that majority, and perhaps some Republicans, will probably not accept it without enforceable federal rules to replace them [67]. Each front is developed in its own section.
The open licenses of Kimi, Qwen, Muse Spark Open and Mistral Large 4 are not entirely free in 2026, because they require a separate contract from large companies that sell third parties access to the model as a service, and leave free whoever uses it for themselves. According to the European Commission’s guidelines, a license like that hardly qualifies for the European exemption for open models, because those guidelines do not count as open a license that charges for commercial use or for use above a certain scale (see “Pre-deployment evaluation and transparency”).
The evidence on the risk of publishing weights points both ways. CAISI, the AI evaluation center of the U.S. Department of Commerce, tested two open models, DeepSeek’s and OpenAI’s. When a harmful request was worded to get around their safeguards, which are what the model is taught so that it refuses to help cause harm, both complied far more often than closed models. The safeguards of Chinese open models, moreover, give way to simple attacks. However, a laboratory trial with novices did not find that 2025 models helped them significantly (see “What open and closed mean”).
In cybersecurity there are already concrete cases. Since September, an open model, GLM-5.3, has turned known flaws in a program into working attacks almost as often as Anthropic’s best closed model: in an Anthropic test it succeeded in 50 of 410 attempts, against 56 for the closed model. Because its weights are public, anyone can retrain it to stop refusing those requests, and doing so cost Anthropic’s team about US$4,400 in compute. In November, the FBI documented that attackers prefer cheap Chinese models to automate intrusions [68] [69] (see “Cybersecurity”).
02
Who regulates AI
AI is regulated at seven levels, and a government decides at only four of them. The other three are technical standards, corporate self-regulation and the private rule of each contract.
Every rule has winners
A license like the one Sam Altman, OpenAI’s chief executive, proposed before the U.S. Senate in 2023 lets through whoever is already above the threshold and makes the climb more expensive for whoever comes behind. Altman proposed an agency that would license any effort above a certain scale of capabilities and could revoke those licenses. The easiest threshold to measure, he said, is compute, the amount of calculation used to train the model. He also asked that small companies, open-source models and researchers not be held back [70]. The next day, the economist Joshua Gans, of the University of Toronto, called that license a classic barrier to entry, because whoever comes later will have to obtain one or let themselves be bought by whoever already has it [71]. The public affairs study records the same accusation, made against Anthropic by David Sacks, the Trump administration’s AI lead until March 2026. According to Sacks, Anthropic uses fear to obtain rules that the large players can afford and newcomers cannot [72].
George Stigler argued in 1971 that, as a rule, regulation is “acquired by the industry” and is designed and operated “primarily for its benefit” [3]. Stigler put it forward as the central thesis of his article, not as a finding proven in every case. We do not take his thesis as proof. We use it to ask, at each level, who writes the rule, who already complies with it when it takes effect and whom it leaves outside its reach. The winners and losers we point out come from those answers and from the evidence at each level, and they are marked as what we believe.
The seven levels
| Level | Who writes the rule | Who gains, we believe | Who loses, we believe |
|---|---|---|---|
| International | Governments | The governments that sign, which keep defense and national security free | Those who hoped for a common limit on military and surveillance uses |
| Regional | The European Union | The labs that signed the code of practice, and small open models | Those who do not sign the code, and the open model with systemic risk |
| National and subnational | Congresses, governments and state legislatures | If a federal law preempts state laws, whoever publishes weights; if not, the states that legislate first | If there is preemption, whoever sues under a state law; if not, frontier developers, who comply with several laws at once |
| Sectoral | Supervisors and professional bodies | The clouds that already comply with those rules, and the suppliers that install the model inside the client | AI tools open to the public, and the lab that retains data for review |
| Technical standards | ISO and the standards bodies, with private auditors | Whoever gets certified first and uses it to sell | Whoever cannot pay for certification when asked for it |
| Self-regulation | Companies | Those who write the standard and decide who gets in | Those without the scale to get in, who arrive at a standard already written |
| Private | Each company, in its contracts | Whoever controls a scarce input: the model, the cloud, memory or the chip | Whoever depends on a single supplier of that input |
International. This level favors the governments that sign, because they commit to human rights principles and keep defense and national security free, which is where they clash most with AI companies. During the negotiation, civil society organizations such as Access Now, AlgorithmWatch and EDRi denounced that some states sought a blanket exemption for national security and defense, and the option of leaving companies out [73]. This level creates no direct obligations for companies, so it does not harm them either.
Those who hoped for a common limit on military and surveillance uses lose, because today that limit exists only in private rules. One of them is the one Microsoft applied to a unit of Israel’s Ministry of Defense. Another is Anthropic’s usage policy, whose prohibitions led the U.S. Department of War, the name the government has given the U.S. Department of Defense since September 2025, to designate the company a supply chain risk. That legal mechanism, provided for in defense procurement law and in federal procurement law, allows a supplier to be excluded from certain contracts [74], and the designation remains in litigation (see “A private rule weighs as much as a law”).
The Council of Europe Framework Convention is the only binding treaty on AI. The Council of Europe is an organization of 46 states devoted to human rights, democracy and the rule of law, and it is distinct from the European Union [75]. The convention obliges each State Party to take measures so that the AI used by its authorities, or by private actors acting on their behalf, respects those three values. It also requires assessing and mitigating the risks of systems, documenting their use and giving affected people a way to seek redress. Compliance is monitored through periodic reports from each state to a Conference of the Parties [76].
The convention was opened for signature on September 5, 2024. It can be signed by the Council’s member states, the European Union and the non-member states that took part in drafting it [77]. It has 21 signatories, among them the United States, the United Kingdom, Canada, Japan and Uruguay. Mexico took part in drafting it as an observer state and did not sign it. The convention has not entered into force because it needs five ratifications, three of them by Council member states, and it has only two, those of the European Union and of Bosnia and Herzegovina [78] [79]. It does not apply to national defense and does not require applying it to national security. It also leaves each state to decide whether it binds companies directly or takes “other measures” to comply [80]. The OECD Recommendation on AI, from 2019, sets out five principles for trustworthy AI, among them transparency, safety and accountability, and five policy recommendations for governments. It has 51 adherents and is not binding [81].
Regional. This level favors whoever already has the legal and security team needed to sign the code of practice, the voluntary document that spells out the EU AI Act’s obligations for general-purpose models, and to document their models. These are the large labs, as Miranda Cole and Dania Assas, of the U.S. law firm Perkins Coie, point out; in their view, small companies may lack those resources [82].
It also favors small open models, which the exemption for open models relieves. That exemption removes the obligation to document the model before the AI Office and before whoever integrates it, and the obligation to appoint a representative in the Union when the provider is outside it. It leaves them the copyright policy and the summary of training data [83]. Whoever does not sign the code loses, because it bears the burden of proof. The Act requires it to demonstrate compliance through adequate alternative means, which the European Commission assesses, while a signatory can rely on the code to demonstrate it [83]. The large open model also loses. Above the systemic risk threshold, the exemption stops applying. That threshold is presumed from 10^25 training operations, a 1 followed by 25 zeros. Each operation is an addition or a multiplication, and the total measures how much calculation it took to train the model, so the threshold reaches only the largest models. That model then complies with everything a closed one does: documentation, model evaluation, mitigation of its risks, reporting of serious incidents and protection against cyberattacks [83].
The EU AI Act, in force since August 2024, classifies AI systems by their risk. It bans a few practices, such as the social scoring of people. It imposes its heaviest obligations on high-risk systems, which are those used in areas such as employment, education, access to essential services, policing or justice. It regulates general-purpose models separately, which are those that competently perform a wide variety of tasks and can be integrated into many different systems. The obligations for these models have applied since August 2025 [84].
The Act imposes three obligations on every provider of a general-purpose model: documenting it before the European Commission’s AI Office and before whoever integrates it, having a copyright policy and publishing a summary of its training data. The AI Office is the Commission body that supervises those models, and it can request information from the provider, evaluate the model and impose fines [85]. The Act exempts openly licensed models from the two documentation obligations, unless they pose systemic risk (see “How the law treats each model”) [86]. Systemic risk is presumed when training the model consumes more than 10^25 calculation operations. The provider of such a model must also evaluate it, mitigate its risks, report serious incidents and protect it against cyberattacks [84].
The code of practice, published in July 2025, is a voluntary document that spells out those obligations. It has three chapters: transparency, copyright, and safety and security, and the last one is addressed only to models with systemic risk. It was signed by 22 companies, among them Amazon, Anthropic, Google, IBM, Microsoft, Mistral, OpenAI, Cohere and Aleph Alpha. xAI signed only the safety and security chapter, and Meta, DeepSeek, Alibaba and Zhipu are not on the list [87].
Signing the code does not create a presumption that the provider complies, but the AI Office can take adherence into account when monitoring compliance and as a mitigating factor when setting a fine. Whoever does not sign has to justify that its alternative means are sufficient, usually with an analysis that compares those means with the code and identifies the gaps [88]. The fine for a provider of general-purpose models reaches 3% of its worldwide turnover or €15 million, whichever is higher, and that article has no reduction for small and medium-sized enterprises. In the Act’s other fines, the lower of the two figures applies to an SME [86]. The code promises SMEs simplified ways to comply [88].
National and subnational. Each country regulates with its own laws, and in the United States each state also legislates. The underlying dispute is between the federal government and the states (see “How the law treats each model”). Google, the venture capital fund Andreessen Horowitz and OpenAI call for preempting state laws on model development, that is, having a federal law override them. Anthropic accepts this only with a federal law at least as strong as the one it proposes [89] [90] [91] [92]. Its proposal would require frontier developers to test their models, publish the results, submit them to independent evaluation and maintain a safety program, and would give the government the power to block the deployment of a dangerous model [92].
We believe that preempting state laws would favor whoever publishes weights only if the federal law includes the rule the White House is asking for. That rule would shield the developer from what a third party does illegally with its model [93]. That is precisely what an open developer cannot control, because, as Fei-Fei Li argued against California’s SB 1047, no one can foresee all the uses of a published model [94]. That bill would have required the developers of the largest models to adopt safeguards against catastrophic harms, and Governor Gavin Newsom vetoed it in September 2024 [95]. The bill covered models trained with more than 10^26 operations, ten times the European threshold, and with a compute cost above US$100 million, that is, only the largest and most expensive models. The safeguards also had to cover models derived from the original. The fine reached 10% of the cost of the training compute for the first violation causing harm or an imminent risk to public safety, and 30% for subsequent ones [96].
The draft Great American AI Act, published in June 2026 by Representatives Jay Obernolte, a Republican, and Lori Trahan, a Democrat, is a discussion draft that has not been introduced as a bill. It would bar for three years state laws that regulate the development of a model. It leaves state laws on distribution and use in place, so it is not enough to favor whoever publishes weights. Because the draft defines deploying as making a model available for use or copying, publishing the weights would still be covered by those laws, just as California’s AB 853 already covers whoever hosts them [97] [98]. From January 1, 2027, that 2025 law bars a platform that hosts generative systems from knowingly offering one that does not label its content [98].
Without preemption, we believe the states that legislate first win, such as California and New York, because their laws become the de facto rule for anyone who sells nationwide, since a developer does not make a version of its model for each state. Frontier developers bear several laws at once, because they are the only ones above the compute threshold of California’s SB 53 and New York’s RAISE Act, which cover models trained with more than 10^26 operations [99] [100] [101].
SB 53, signed in September 2025, requires those developers to publish a transparency report and to report critical incidents. Those with revenue above US$500 million must also publish a safety framework, and only they can be fined, up to US$1 million per violation [99]. The RAISE Act, signed in December 2025 and amended in March 2026, takes effect on January 1, 2027, and follows the same scheme. It requires reporting critical incidents within 72 hours, and its fines reach US$1 million for the first violation and US$3 million for subsequent ones [100] [101]. Anthropic, which already published a safety framework and therefore already met that obligation, supported California’s SB 53 [102].
Sectoral. The rules of each sector say where data can be kept and who is responsible for it, without naming a model. The EU’s Digital Operational Resilience Act (DORA) for the financial sector, applicable since January 2025, requires banks, insurers and other financial entities to manage the risk of their IT systems, to report serious incidents and to oversee their technology providers. It also places the providers designated as critical under direct European supervision [103]. Among the clauses it imposes, it requires the contract to set the countries where the data of European financial entities is processed [104].
The U.S. health law HIPAA makes anyone who processes patients’ health information on behalf of a hospital, a doctor or an insurer a “business associate”. That status is subject to the federal privacy and security rules for that information [105]. The associate may use that information only as its contract allows, and the U.S. Department of Health and Human Services can fine it [106]. The U.S. Department of Defense, for its part, requires that the cloud where a contractor stores, processes or transmits defense information have the security level of federal systems [107]. That level is the Moderate baseline of FedRAMP, the program the federal government uses to assess the security of the cloud services it contracts [108].
The Solicitors Regulation Authority, which regulates solicitors in England and Wales, warns that entering a client’s data into an AI tool open to the public, such as ChatGPT, probably breaches confidentiality and may cause the permanent loss of legal professional privilege, the protection that prevents a client from being compelled to disclose what they discussed with their lawyer [109]. The model risk guidance of U.S. bank supervisors, revised in April 2026, guides banks with more than US$30 billion in assets on how to control the quantitative models they use to estimate and decide. It sets no enforceable requirements and leaves generative AI out [110].
We believe that whoever already has those contracts and certifications signed when the client decides wins. They are the three large clouds, which already host the data of those sectors, and the suppliers that install the model inside the client, such as Mistral, Cohere, IBM and SAP [111] [112] [113] [114]. The public tools, which are used without those contracts, lose, as does the lab that requires retaining conversations for review, because that retention clashes with the clause that prohibits the supplier from keeping the client’s data [115] (see “The rules of each sector”).
Technical standards. We believe that whoever gets certified first and uses it to sell wins. Whoever cannot pay for certification when a buyer or a supervisor starts asking for it loses. CEPS estimated how much it costs to set up from scratch the quality management system that the EU AI Act requires of high-risk providers: between €193,000 and €330,000, plus €71,400 a year [4]. That cost weighs above all on small and medium-sized enterprises [4].
The ISO/IEC 42001 standard, from December 2023, sets the requirements for an AI management system [116]. That system is the set of policies, objectives and processes with which an organization develops, provides or uses AI responsibly. It is published by ISO, a non-governmental organization that develops voluntary international standards, and the International Electrotechnical Commission (IEC) [117]. Since July 2025, ISO/IEC 42006 has regulated the bodies that certify against it. It sets requirements for them in addition to those of the general standard for certifying management systems, ISO/IEC 17021-1, so that their auditors have specialized knowledge of AI and audits are carried out with the same criteria [118]. The standard sells for 225 Swiss francs. Certification also requires implementing the management system and passing an audit by an accredited body. Amazon Web Services (AWS), Amazon’s cloud, was certified in November 2024 and announced itself as the first major cloud to be certified. Anthropic was certified in January 2025, and Microsoft and Google have their main AI services certified [119] [120] [117] [121]. The first certificates, however, did not go to the large companies. Before them came Synthesia, a video startup, in September 2024, and KPMG Australia, in October [122] [123].
In the European Union, a high-risk system or a model that complies with a harmonized standard, which is a European technical standard whose reference is published in the Official Journal of the European Union, is presumed to conform with the Act. The authority assumes it complies unless the contrary is shown, so complying with the technical standard amounts in practice to complying with the law. Those standards are drafted by the European standardization organizations at the Commission’s request, not by the legislator [86].
Self-regulation. Whoever writes the voluntary standard wins, because it sets as the standard what it already does, decides who enters the forum and, when the law arrives, sees a document it already had turned into an obligation. Kris Shrishak, of the Irish Council for Civil Liberties, describes this for European technical standards. Companies bring their internal policies to the drafting, and the more the standard resembles that policy, the easier it is for them to comply [124]. Whoever lacks the scale to enter the forum and later arrives at a standard already written loses.
Seven companies in July 2023 and eight more in September signed voluntary commitments with the White House, whose uneven fulfillment the public affairs study measures [125] [126] [127]. They committed, among other things, to test the safety of their systems before releasing them, partly with independent experts, to protect unreleased weights, to develop mechanisms to label AI-generated content and to report on the capabilities and limits of their systems [125].
Five days after the first commitments, Anthropic, Google, Microsoft and OpenAI founded the Frontier Model Forum, which now includes Amazon and Meta. The forum is a nonprofit industry association. It aims to advance research on the safety of frontier models, identify best practices and share information with governments [128]. The forum admits only those that demonstrate the capacity to develop and deploy frontier models at scale and that fund it for at least three years [128] [129].
C2PA, the content provenance standard, takes its name from the Coalition for Content Provenance and Authenticity, which develops it. The standard attaches information about its origin to a photo, a video or an audio file, for example whether it came from a camera, whether it was edited or whether an AI produced it [130]. The coalition has Adobe, Amazon, Google, Meta, Microsoft and OpenAI on its steering committee [131], and Google uses it in its “About this image” feature and in its ad systems [130]. Anthropic published its Responsible Scaling Policy in September 2023, and Google DeepMind published its Frontier Safety Framework in May 2024 [132] [133]. Both are internal policies with a similar scheme. The company defines levels of dangerous capability, evaluates its models to know when they approach them and commits to applying stricter safety measures at that point. Anthropic acknowledges that its policy may force it to pause the training of more powerful models [132] [133]. California’s SB 53 later required frontier developers with revenue above US$500 million to publish a safety framework (see “Pre-deployment evaluation and transparency”) [99].
Private. Each company writes its own rules: the license of each model, the labs’ usage policies, the clouds’ terms, the clients’ clauses and the long-term contracts for manufacturing and memory capacity. Those rules favor whoever controls a scarce input and harm whoever depends on a single supplier of that input. The U.S. Federal Trade Commission (FTC) warned of this in January 2025 about the contracts between the large clouds and the labs, when it published a report on the partnerships of Microsoft with OpenAI and of Amazon and Google with Anthropic. According to the statement by Lina Khan, then chair of the FTC, that accompanied the report, those contracts can tie labs to one supplier and deprive startups of key inputs [134] (see “A private rule weighs as much as a law”).
A private rule weighs as much as a law
In the supply chain, a private rule weighs as much as a country’s law wherever the law does not reach or has not yet acted. Gabriela Ramos, former Assistant Director-General of UNESCO, wrote in April that the ethical limits of AI were left to private companies because public systems failed [135]. When a government acts, the public rule prevails over the private one. That happened in June, when the U.S. Department of Commerce ordered Anthropic to cut off foreigners’ access to two of its models, and with the Cyberspace Administration of China, the agency that registers and evaluates the AI services offered to the public in that country, which in September 2025 ordered ByteDance and Alibaba to cancel their orders for a chip NVIDIA had designed for China [136] (see “Why China stopped buying from NVIDIA”).
A private rule reaches foreign companies wherever they operate. Since September 2025, Anthropic has not sold to companies controlled from regions where it does not offer its products, “like China”, and counts among them those that are more than 50% owned, directly or indirectly, by companies headquartered there [5]. Its policy in force since November 12 applies the criterion to all excluded regions, among them Russia, Iran, Cuba and Venezuela, and also to models sold through cloud platforms and resellers [6]. No law required it to do so. OpenAI and Google restrict by the user’s location, and we did not find an ownership criterion in their terms [137] [138].
A private rule also reaches governments. In September 2025, Microsoft cut off part of its cloud and AI services to a unit of Israel’s Ministry of Defense, because its terms prohibit using its technology for the mass surveillance of civilians. The company explained it this way: “Microsoft is not a government or a country. We are a company” [7]. Its code of conduct prohibits facial recognition by or for state or local police in the United States [139]. The U.S. Department of War designated Anthropic a supply chain risk because the company did not remove two prohibitions from its usage policy, the one on fully autonomous weapons and the one on domestic mass surveillance. The designation remains in litigation, with opposing rulings from the U.S. Court of Appeals for the District of Columbia Circuit and a federal court in California (see the public affairs study) [140]. OpenAI, for its part, requires reviewing and approving any use of its models for national security, law enforcement or intelligence purposes [141].
A private rule is enforced without a trial. A lab cuts off access when it decides a client is in breach, as OpenAI did with Cursor after SpaceX bought that company and Anthropic did with xAI, two cases from the public affairs study [142]. OpenAI’s policy also prohibits using its models or their outputs to prepare training data for models that replace its own [141]. It is a barrier against distillation, the technique of training a model on the responses of another. The company itself wrote it, before any of the bills against distillation. As Louis Tompros, of Harvard Law School, explained, copyright does not protect what a model generates, so what remains against distillation is mainly the contract [143]. Henderson and Lemley, however, doubt that a judge would enforce it [2] (see “How the law treats each model”).
A private rule also allocates what is scarce. SK hynix, one of NVIDIA’s three memory suppliers, expects demand to exceed supply until 2030 and already has long-term contracts for more than half of its 2027 and 2028 output [8]. As of July 26, NVIDIA had supply and capacity commitments of US$279 billion, more than double the figure three months earlier, and in October those commitments reached US$390 billion [14] [144]. Anthropic’s compute commitment is described in “Where the money stays”. According to the market research firm TrendForce, the U.S. clouds that signed multiyear contracts locked in their memory prices, and since the third quarter of 2026 the increases have fallen on those without a contract [145]. We believe that whoever signs those contracts first decides who will have chips and memory in 2028, with an effect similar to that of an export control and without any government having decided it.
03
The chain around AI
AI is produced along a chain that runs from the machines that make chips to the person who uses an assistant. In the first links, one or a few suppliers decide, and in the last ones thousands of companies compete.
| Link | Who dominates | Public rule | Private rule |
|---|---|---|---|
| Chipmaking equipment | ASML is the only maker of EUV lithography machines, the ones that print the most advanced circuits with extreme ultraviolet light, of a very short wavelength that allows finer lines. Canon and Nikon compete only in the previous generation, and ASML’s optics have a single supplier, Carl Zeiss SMT [10] | Dutch and U.S. export licenses for EUV machines and for part of the previous-generation ones [10]. China is developing its own machine (see the public affairs study) [146] | Zeiss’s exclusivity |
| Manufacturing and packaging | Packaging is the final step, which mounts the chip and connects it with its memory in a single piece. TSMC makes NVIDIA’s chips (NVIDIA also uses Samsung), all of AMD’s chips at 7 nanometers or less, that is, from the most advanced manufacturing generations, in which circuits are smallest, and Microsoft’s Maia 200 [11] [12] [13]. High-performance chips went from 43% of TSMC’s revenue in 2023 to 58% in 2025 and to 68% in the third quarter of 2026, and its ten largest customers account for 78% of its revenue [147] [53] | Export controls (see “How the law treats each model”) | Committed capacity: NVIDIA’s supply and capacity commitments went from US$119 billion in April to US$279 billion in July and to US$390 billion in October [14] [144] |
| High-bandwidth memory (HBM) | SK hynix, Samsung and Micron, NVIDIA’s three suppliers [11]. All three report more demand than capacity [8] [51] [52] | Export controls. Memory enters China through smuggling (see the public affairs study) [146] | Long-term contracts for more than half of SK hynix’s output for 2027 and 2028 [8] |
| Chip design | NVIDIA took in US$215.9 billion in its fiscal 2026, almost 90% from data centers, and two direct customers accounted for 22% and 14% [11]. The clouds design their own, Google’s TPUs, Amazon’s Trainium and Microsoft’s Maia, and NVIDIA counts those clouds as competitors [11] [148] [149] [13] | Export controls. Under the November 16 rule, in force since the 19th, a license is required to sell them to five countries at risk of diversion and to provide their compute to companies in China and in other countries under arms embargo [22] | Anthropic will use up to one million Google TPUs and is deploying Claude on more than one million Amazon Trainium2 chips [150] [151] |
| Data centers and energy | The clouds and the data center operators | Permits, power grids and Chinese controls on rare earths (see the public affairs study) [152] | Turbine production, sold out for years (see the public affairs study) [152] |
| Cloud | Amazon, Microsoft and Google, which are also shareholders in the labs (see “Where the money stays”) | Competition authorities and the European rule that sets a deadline for switching clouds (see the public affairs study) [153] | Terms of service, such as those Microsoft applied to Israel, and AWS’s responsible AI policy, which also covers the third-party models it sells [7] [154] |
| Training data | Publishers, platforms and authors. Wikipedia sells access to its data to Amazon, Meta, Microsoft, Mistral and Perplexity, as well as Google; Reddit charges licenses for its content and News Corp is seeking more [155] [156] [157] | Copyright (see “Owning the input”) | Licenses with no public price |
| Labs | Closed and open (see “Who is on each side”) | The EU AI Act, export controls and restrictions on the rival’s model (see “How the law treats each model”) | The model’s license and the lab’s use policy [6] [141] |
| Distribution | The model marketplaces inside the clouds: Amazon’s Bedrock offers “hundreds” of models from at least 13 providers, Microsoft’s Foundry more than 11,000 and Google’s Model Garden more than 200 [158] [159] [160]. Hugging Face, which NVIDIA agreed to buy [24] | Competition (see “How the law treats each model”) | The lab’s policy travels with the model to the cloud [6] |
| Applications, integrators and user companies | Thousands of software and services companies (see “The software industry facing AI”) | The rules of each sector (see “The rules of each sector”) | The clauses large companies impose on their suppliers (see “What they require of their suppliers”) |
| End users | Individuals and employees | Data protection and consumer protection | Each service’s terms of use and internal policies, which at some companies ban the free versions (see “Ban first, then contract”) |
The dispute between open and closed is fought in the links of the labs and of distribution. We believe the upstream links win with any outcome, because models of both kinds are trained and run on their chips, although they do not win equally. SemiAnalysis, an independent research firm covering the semiconductor and AI chain [161], argues that no one needs open models more than NVIDIA, because those models give countries, companies and new clouds fresh reasons to buy GPUs, the graphics processors used to train and run almost all models, beyond the handful of frontier labs. According to the same firm, six of NVIDIA’s seven largest customers, all except SpaceX, are trying to design their own chips to take away its margin, that is, to stop paying it the profit it earns on each chip [27] (see “What Huang says”). The firm counts Anthropic, OpenAI and the large clouds among those seven [27].
ASML, TSMC and the memory makers sell to both sides. ASML also invested in a lab: in 2025 it led Mistral’s funding round with €1.3 billion, for which it obtained about 11% of Mistral’s capital, counting all the shares the company could eventually issue [10]. Later, in 2026, Samsung led Mistral’s 3 billion euro round [162].
An open model escapes the lab’s use policy and the cloud’s terms, because it is downloaded and run on one’s own hardware. That exit has a limit in the largest open models, those with more than 2 trillion parameters. They do not fit on an eight-GPU server and in practice are used through cloud providers, which apply their own terms to them (see “Who pays for open”).
No open model, however, escapes the upstream links. It still needs chips made by TSMC with ASML machines, and memory produced by three companies. Governments control the sale of those chips and that memory, and their production capacity is already allocated by contract, so a new buyer only gets what those contracts leave free.
Governments exercise that control through export controls, which require an official license before certain goods and technologies are sold to specific countries, buyers or uses. In the United States they are administered by the Bureau of Industry and Security of the U.S. Department of Commerce, through the Export Administration Regulations [163].
NVIDIA orders the chain
NVIDIA no longer sells a single chip but a computer the size of a cabinet. Vera Rubin, the platform it began shipping between August and October 2026, brings together seven of its own chips. Among them are the Rubin GPU, the Vera CPU, the general-purpose processor that directs the rest of the machine, and a processor for inference, which is the work of running an already trained model so that it responds. That processor comes from Groq’s technology [164] [14]. The platform is sold in racks, the cabinets that group dozens of chips as a single machine. It is also sold in pods, sets of 40 racks with 1,152 GPUs and almost 20,000 NVIDIA silicon dies, the pieces of silicon that make up each chip [165]. Each rack carries between 1.3 and 1.5 million components from some 200 suppliers, and NVIDIA expects to build some 200 pods a week, Huang said in March [15]. In May the company counted hundreds of partners in its chain, 150 of them in Taiwan alone, spread across more than 350 factories in 30 countries [166].
Huang calls designing the chip, memory, network, system and software at the same time “extreme co-design”, and that way of working changes NVIDIA’s relationship with its suppliers. Huang justifies it by the slowdown of Moore’s law, the observation that the number of transistors on a chip, the tiny switches it computes with, doubles every two years with little increase in cost [167]. While that law held, the chip improved on its own. He says that with co-design NVIDIA multiplied compute a million times in ten years, when Moore’s law would have multiplied it by a hundred [15]. Integration, which used to be done in the data center, is now done in the chain’s factories. Huang said he explains to each supplier what is coming and why, and asks them to invest “several billion dollars”. He said suppliers do so “because they trust me” [15].
About three years ago, Huang said, he convinced several heads of the memory companies to invest in HBM, which at the time was used almost only by supercomputers. Later he convinced them to invest in the low-power memory used in phones, for use in data centers. The three companies had, he said, the best year in their 45-year history. When podcast host Lex Fridman asked him, in the March interview, whether he worries about bottlenecks in the chain, he said no: “I told ’em what I needed”. Of TSMC he said that in three decades of business NVIDIA has never signed a contract with it, and he attributed that to the trust between the two companies [15].
NVIDIA has an edge over the clouds’ chips because of its installed base, that is, the number of machines in use that already run on its technology. “The install base is, in fact, the single most important part of an architecture”, Huang said, because programmers write for the platform that reaches the most people. That is why NVIDIA put CUDA, its programming platform, on its GeForce gaming cards, even though that decision raised its cost by about 50% when its gross margin, what it kept from each sale after paying for manufacturing, was 35% [15]. Huang said gamers did not pay more for that feature and that the bet took a decade to pay off [15].
In 2026 NVIDIA also began financing the chain. Its supply and capacity commitments went from US$119 billion in April to US$279 billion in July. They are agreements with its suppliers, above all for memory and fabs, to secure components for the coming years [14]. In the same report, the company wrote that AI-specialized clouds and labs have large demand for compute but cannot obtain long-term infrastructure contracts or investment-grade financing, the rating that agencies give to the borrowers with the lowest risk of default [168]. That is why, according to the report, NVIDIA provides guarantees, such as the one for the Ohio campus (see “NVIDIA between the two sides”) [14].
NVIDIA’s sales are concentrated. One direct customer, that is, one that buys from it without intermediaries, such as a server maker, a distributor or a cloud, accounted for 16% of revenue in the second quarter of its fiscal 2027, from May to July, and over the half year three customers accounted for 16%, 15% and 13%. An AI research and deployment company, which NVIDIA does not name, also contributed “a meaningful amount” of revenue by buying cloud services from companies that are, in turn, NVIDIA customers [14].
NVIDIA coordinates investment across the whole chain. It tells suppliers what capacity to build, commits to buying it and backs the credit of those who will use it. SemiAnalysis calls it the chain’s “central bank”, because it tied up most of the supply of wafers, the thin silicon discs on which chips are made [169], and of memory, and its guarantees let new clouds and labs obtain credit without depending on the large clouds [27]. The guarantees work because NVIDIA ensures a minimum income to the buyer of its chips, answers to the data center owner or signs the lease itself. With that backing, an operator without an investment-grade rating gets credit at the price of one that has it [27].
That coordination rested, according to Huang himself, on the trust that the suppliers’ chief executives placed in him. His resignation tests whether that trust belonged to the company or to the person, just as the chain is building the 2027 and 2028 capacity that NVIDIA committed to buy. SemiAnalysis argues that it belonged to the person, because Huang won those deals chief executive to chief executive [27]. Suppliers in Taiwan and Korea told Reuters they doubt his successor will keep the same relationship, although none announced a change. Armina Rosenberg, cofounder of Minotaur Capital, which holds NVIDIA shares, believes instead that customers’ 2027 spending plans do not depend on who runs the company, because the chips NVIDIA will deliver in one or two years are already designed and in production [170].
04
Jensen Huang’s fall
- In May 2026, before Huang joined Trump’s trip to Beijing during a stopover, Marco Rubio, the U.S. Secretary of State, warned Trump about intelligence reports that Huang had a relationship with a woman suspected of spying for China, according to the Washington Post [171]. In October, the New York Times identified that woman as Alice Huang, an executive at Megaspeed. Megaspeed is an NVIDIA customer in Singapore, under investigation for helping Chinese companies evade export controls [172]. On May 14, with Trump visiting Beijing, the U.S. Department of Commerce authorized some ten Chinese companies, among them Alibaba and ByteDance, to buy NVIDIA’s H200 chip, according to CNBC [173]. The H200 is not NVIDIA’s newest chip, and its export to China was banned until the Trump administration allowed it in 2025 for approved customers [174]. CNBC notes that some of those authorizations predate Rubio’s memo, a classified document about Alice Huang described below [173].
- On October 29, the New York Times reported that U.S. agencies had concluded that Alice Huang was passing information to the Chinese government. The public affairs study covers that report. It also covers the investigation into Megaspeed that, according to the New York Times, the Department of Commerce suspended [172].
- On November 4, the chairs of three committees of the U.S. House of Representatives asked Huang for a closed-door session on November 18, in room 2172 of the Rayburn building. The three committees are the Committee on Foreign Affairs, the Select Committee on the Chinese Communist Party and the Permanent Select Committee on Intelligence. Under the House resolution that created the Select Committee, it can only investigate and recommend policy, with no authority to process legislation [175]. The session would cover NVIDIA’s security protocols against espionage and Huang’s contacts with Alice Huang, whom the letter calls an intelligence asset of the People’s Republic of China “according to public reports”. The chairs asked him to accept the date or propose another by the 11th [176].
- On November 8, the New York Times reported that Rubio had overseen, in the spring, a classified memo based on surveillance of Alice Huang. The memo circulated among officials with access to classified information before Trump’s trip to Beijing [20]. Earlier, the Financial Times had revealed Alice Huang’s relationship with Jensen Huang, according to the press that summarized it [172].
- On November 8, NVIDIA’s board opened an independent investigation “in light of recent reporting” that Huang had been “the target of a foreign intelligence effort”. A committee of independent directors, with outside counsel, oversees it. The board said it had no reason to believe that national security or NVIDIA’s secrets had been compromised, and that it had confidence in Huang [17]. That same day, Huang wrote to employees that he had been unfaithful to his wife, apologized and said he supported the investigation [177].
- On November 9, the New York Times reported that Tim Teter, NVIDIA’s general counsel, wrote in August 2025 to the Department of Commerce agents investigating Megaspeed. He told them that Jeff Kessler, Under Secretary for Industry and Security at the U.S. Department of Commerce, had authorized him to contact them to discuss the case. Some officials saw it as a breach of procedure that could taint the case. According to the paper, an under secretary usually stays out of the details of an investigation, because he may have to rule on it as a neutral arbiter, and NVIDIA had not been ruled out as a subject of the investigation. The case then passed to the U.S. Department of Justice, which asked the Department of Commerce to pause its work. The Department of Commerce says it acted by the rules, and NVIDIA says it consults officials to answer their questions [178].
- On November 10, the independent directors retained Joon H. Kim, former acting U.S. Attorney for the Southern District of New York, and the firm Wachtell, Lipton, Rosen & Katz [179].
- On November 11, the board accepted Huang’s resignation as president, chief executive officer and director, effective immediately. Colette Kress, the chief financial officer, became interim chief executive officer while keeping the finance role. Harvey C. Jones, a director since 1993, became the first chair of the board in NVIDIA’s history [180]. The filing with the U.S. Securities and Exchange Commission (SEC), submitted on the 12th, gives no reason for the resignation, does not mention the investigation or China and says nothing about a future role for Huang [16]. That filing is Form 8-K, which a listed company files within four business days of a material event. When a chief executive resigns, the form only requires reporting the fact and its date [181].
- On November 11, Axios reported that U.S. Secretary of the Treasury Scott Bessent had confronted U.S. Secretary of Commerce Howard Lutnick on Thursday the 5th, at a meeting in the White House Situation Room, for not having added Megaspeed to the Entity List, and that Lutnick replied he would review it. According to Axios, White House and intelligence officials had been asking him to act against Megaspeed since the fall of 2025. CNN confirmed the meeting with two sources. The Department of Commerce denied that the conversation took place [182] [183].
- On November 12, NVIDIA’s stock opened 13.1% lower, was down as much as 14.4% and closed down 4.3% [184].
- Also on the 12th, the three committees replaced the November 18 session with a closed-door transcribed interview on December 1. The letter is addressed to Huang as NVIDIA’s founder. It asks him to confirm the interview in writing by 8:00 p.m. Eastern time on November 13, or face a compulsory subpoena. A compulsory subpoena is the order by which a House committee requires testimony, and anyone who knowingly disobeys it is exposed to criminal prosecution for contempt of Congress [185]. The committees also asked for an interview with Tim Teter, NVIDIA’s general counsel, on December 2. That interview would cover chip smuggling, military use of NVIDIA’s products by adversaries and NVIDIA’s “interference” in export policy and in investigations, among them the one into Megaspeed. They asked Jeff Kessler, Under Secretary for Industry and Security at the U.S. Department of Commerce, to appear. The statement ties the case to four bills on chips and remote compute that the committees ask to include in the defense act [18]. That act is passed every year, authorizes the spending of the U.S. Department of Defense and sets its policies [186].
- The four bills in the statement of the 12th are the AI OVERWATCH Act, which would require a license to export advanced chips to companies in China and in five other countries and would require notifying Congress 30 days before approving each one; the Chip Security Act, which would require those chips to be able to verify their location before export; the MATCH Act (see “The fab”), and the Remote Access Security Act (see “Restrictions on the rival’s model”) [187]. According to TechCrunch, Brian Mast, chair of the House Committee on Foreign Affairs, said that day that Huang would agree to appear before Thanksgiving, November 26 [188].
- On November 13, on CNBC, Bessent blamed the Department of Commerce for not having added Megaspeed, the Singapore company where Alice Huang worked, to its Entity List earlier [189]. The Entity List gathers the foreign companies and persons to which a product subject to U.S. regulations cannot be exported without a specific license. The Department of Commerce adds to it those it considers, based on specific facts, a risk to the national security or foreign policy of the United States [190]. In the same interview, Bessent did not deny the meeting: he said he does not repeat what is discussed in the Situation Room. According to CNN, Trump’s advisers are discussing whether to remove Lutnick, and the White House said the day before that the president maintains “complete confidence” in him [191].
- Also on the 13th, according to CNBC, Cantor Fitzgerald, the firm Lutnick ran and that his sons now run, reported in its quarterly filings to the SEC 49% fewer NVIDIA shares in March 2026 than in September 2025: it went from some 3.5 million to 1.78 million, with a value that fell from US$652 million to US$310 million. Those filings do not say why the figure fell or whether the firm reduced its exposure to NVIDIA, because Cantor also held options, almost all of them puts, which protect it if the stock falls. Cantor says those shares hedge its clients’ trades and that on its own account it has only had minor positions. CNBC makes clear that there is no evidence those changes responded to the intelligence warnings [173].
- On November 13, the Washington Post reported that Rubio oversaw the memo and recast in more categorical terms the CIA report on which it was based. According to the paper, CIA officials were not convinced that what was intercepted proved deliberate transfers to the Chinese government. The paper also reported that the memo circulated in the spring among officials with access to classified information. It added that Huang had obtained Trump’s permission to sell chips to China, and that this earned him the opposition of officials and lawmakers who wanted to block that technology’s path [21].
- On November 15, John Moolenaar, chair of the Select Committee on the Chinese Communist Party, told CBS News that Huang and NVIDIA representatives agreed to a voluntary closed-door interview, and that the dates were still being negotiated [192].
- On November 16, the Department of Commerce released an interim final rule, written “by and at the direction of” the U.S. Department of the Treasury and the National Security Council. An interim final rule takes effect without prior consultation, and the public can comment on it afterward, in this case for 60 days. The rule requires a license to sell advanced chips to Indonesia, Malaysia, Singapore, Thailand and Brazil, except when the buyer, or its ultimate parent, is headquartered in the United States and the seller reports each month to the Bureau of Industry and Security who takes part in the sale, where the chips are going, which ones and how many they are, and the shipping date. According to the rule, in those five countries the risk that chips are diverted to unauthorized buyers is higher.
- The same rule of the 16th also requires a license for a data center in any country to provide more than 12 million units of TPP to companies headquartered, or with an ultimate parent headquartered, in Macau or in the countries of Country Group D:5. That group is the U.S. arms embargo list, which includes China, Russia and Iran. TPP is the measure of power by which chips are classified: it multiplies the operations per second the chip performs by the bit length of the numbers it computes with [22]. By our calculation with NVIDIA’s specifications, 12 million units are equivalent to about 760 H100 chips [193]. According to the rule, that volume of compute makes it possible to train advanced models, and the requirement seeks to prevent those companies from doing so with U.S. chips installed in other countries.
- The rule of the 16th exempts compute used only in commercial recommendation systems, such as those of online stores and social networks, provided it is reported to the Bureau of Industry and Security; a system that generates text, images or other content does not count as such even if it also recommends. Those who already provide that service have 90 days to apply for the license. The rule applies from its publication in the Federal Register, on November 19 [22]. The Federal Register is the official journal of the federal government, where agency rules are published every business day [194].
- Also on the 16th, the U.S. Department of Commerce and the U.S. Department of Justice issued a joint statement on Megaspeed. According to the statement, the Department of Commerce opened an investigation into that company in 2024 for possible export control violations, after detecting suspicious activity. In May 2025 it asked NVIDIA for its records of direct or indirect sales to Megaspeed. In June, NVIDIA put Megaspeed on its list of customers it does not sell to. Also in June, Bureau of Industry and Security agents, with Singapore authorities, inspected Megaspeed in person and interviewed its chief executive, Alice Huang. Since then the criminal investigation has been handled by the Department of Justice. The statement does not mention espionage or charges [195]. According to CNBC, the Department of Justice has not concluded whether Alice Huang was spying for China. Singapore did not investigate her for espionage, but in February 2026 it barred her from returning to the country for having knowingly given inaccurate information about her role in Megaspeed [196]. Chris McGuire, of the Council on Foreign Relations, a foreign policy think tank based in New York [197], said the statement does not explain why the Department of Commerce did not cut off Megaspeed’s chips [196].
- On November 17, NVIDIA reported its third-quarter results, the first without Huang. The company had revenue of US$115.844 billion, twice as much as a year earlier [198]. Opening the call with analysts, Jones said the board does not intend to give Huang his job back. He said it hours after Huang wrote that it would be an honor to return [199]. On the same call, Kress said NVIDIA is ending H200 sales in China. The company also announced a board committee that will oversee national security, export control compliance and the risks of global trade [200].
- On November 17, the chairs of the three committees announced that Huang and Tim Teter voluntarily agreed to the transcribed interviews, on December 1 and 2, after Thanksgiving [19].
The press attributes the resignation to two relationships: the one Huang had with Alice Huang and a second one, with an NVIDIA executive, reported in a New York Times story. Yahoo Finance’s live blog states, without citing a source, that the board decided to remove him. The same blog picks up two reports from other outlets: according to Reuters, Huang had a third relationship, with a second NVIDIA employee, and according to Bloomberg, federal prosecutors in Manhattan are examining whether companies linked to Alice Huang helped smuggle products from Super Micro, a U.S. server maker [201] [202]. No company document confirms those reports, and the SEC filing speaks of a resignation [16] [180].
Behind the personal scandal, the case is about China’s access to NVIDIA’s chips. That is the focus of the interviews with the general counsel and with the Under Secretary of the U.S. Department of Commerce [18]. According to the Wall Street Journal, while Alice Huang was obtaining NVIDIA chips for companies in China, her husband, Kang Wei, was planning data centers in that country [203] [204].
According to records reviewed by the Wall Street Journal, Megaspeed, the Singapore company where she was an executive, arranged, from March 2024 to June 2025, the shipment to companies in China of top-tier NVIDIA chips worth about US$12 billion [203]. According to The Information, Alice Huang, her sister and her husband built a network of companies in China and Singapore to obtain NVIDIA chips. Through that network, chips worth at least US$7.5 billion have reached Chinese companies since 2023. The Information relies on corporate records and on two people familiar with the network [205]. According to U.S. investigators cited by CBS News, Alice Huang used industry contacts and shell companies to obtain 240,000 chips whose export to China is banned [206]. According to Bloomberg, the account NVIDIA gave the government of what it sold to Megaspeed through its partners fell short by 50,000 chips [207].
Kang Wei, for his part, planned from 2021 to 2024 investments in data centers in China worth more than US$9 billion, through Century Longteng, a company based in Shanghai. The plans included a US$4.5 billion campus in Wuhu, with state support for electricity and infrastructure, and another of US$5 billion in Liangshan, according to the Wall Street Journal [204].
According to The Information, two companies in Alice Huang’s network invested in DeepSeek’s second funding round, from July to October 2026. They are Century Litong and its parent, Jiangsu Lettall Electronics, and they invested through vehicles, entities created solely to channel an investment, so that their names do not appear in the shareholder register. Corporate records show 95 million yuan from Lettall in a vehicle of Monolith Management, DeepSeek’s third-largest shareholder. According to one of the three people with direct knowledge cited by The Information, the investment sought to have those companies supply DeepSeek with Blackwell chips, an NVIDIA generation newer than the H200’s [174], whose sale to China is banned. DeepSeek employees told investors that Litong could buy NVIDIA chips in large quantities on the black market [208].
Megaspeed says it operates within the law and designs its supply chain to comply with U.S. export controls. It also says that Alice Huang left the company on August 30, 2025, and that it hired outside lawyers to review reports that, according to the company, have no verifiable evidence. Megaspeed has not been charged with violating export rules [203]. The company denies having diverted chips [207].
The Wall Street Journal found no confirmation that Megaspeed’s chips reached Kang Wei’s centers, or that Alice Huang bought them on the orders of Chinese authorities. Neither of them responded to the paper [204]. The Information found no records that Alice Huang holds shares or positions in Century Litong or in Lettall. It links the two companies to her for two reasons: Century Litong was founded by a company ultimately controlled by her sister, Huang Yu, and there are transactions between companies of Huang Yu and of Kang Wei. What has been published includes no response from DeepSeek, the two companies or Monolith [208]. China’s Ministry of Foreign Affairs called the espionage accusations groundless, speculative and made to defame China. Its spokesperson Mao Ning asked the United States to stop fabricating spy stories as a pretext for a technology blockade [209].
According to Reuters, some 600,000 NVIDIA H200 chips are waiting in Hong Kong for the Chinese government to authorize their import [174]. According to the news agency, the Chinese government is withholding the permit for large imports because it does not want to depend further on U.S. technology [174]. The Chinese market already weighs little in NVIDIA’s accounts. Together with Hong Kong, it contributed 9.1% of its revenue in fiscal 2026. In February the company said it was effectively excluded from Chinese data centers [11]. In the third quarter it still shipped Hopper products to China, the NVIDIA chip generation that includes the H100, the H200 and the H20, worth US$2.4 billion, and it expects no compute revenue from China for the fourth quarter [144] (see “NVIDIA between the two sides”). The November 16 rule closes, from the 19th, the path of rented compute in third countries. Until then, a Chinese company could remotely use NVIDIA chips installed in data centers in other countries, such as those Megaspeed was building in Southeast Asia for Chinese clients, without violating the controls [178]. We believe what is now being decided is whether the U.S. government will again authorize sales to China and how credibly NVIDIA, without Huang, will defend its position on exports.
There is another reading of the fall, one that touches the subject of this study. According to it, the espionage story would be a pretext. The underlying conflict would pit closed models, backed by capital and national security, against open ones, defended by NVIDIA and China. NVIDIA would have committed US$26 billion over five years to open models, in line with China’s strategy, and the scandal would strike the main defender of those models right after the July 24 letter in favor of open weights [210] (see “The open camp in the United States”).
The two camps exist, NVIDIA and China are on the open side and the blow comes at the worst moment for that side. But that reading exaggerates how sharply defined the camps are and is wrong to call the espionage story a pretext.
It also misdescribes the amount, which was already committed and was not announced to compete with China. The figure comes from WIRED, which published in March that NVIDIA will spend US$26 billion over five years on open-weight models. WIRED relied on NVIDIA’s SEC filing and on what its executives confirmed [211]. The instrument is a cloud purchase commitment, the obligation to pay cloud providers for compute capacity, which NVIDIA disclosed in November 2025 for its research and its DGX Cloud service, payable through 2031 or later. In August 2026 that commitment already totaled US$29 billion, and the company allocated it to its open models and its autonomous driving software [212] [14]. According to Hugging Face, the platform where the AI community publishes and shares models, data and applications [213], NVIDIA is the largest contributor of open models, and Constellation Research, a technology research and advisory firm based in Silicon Valley [214], calls it the leading open model company in the United States [23] [215].
The camps are real because they respond to opposing businesses. Whoever lives off selling access to their models gains if closed dominates. Whoever sells chips, cloud or services around the models gains if models circulate. What is at stake is who gets paid for AI, and that is why the issue is of the first order (see “Who is on each side” and “NVIDIA between the two sides”).
But the camps are not as Manichaean as that reading paints them, because almost no one plays on a single side. Open labs also make use of closed ones. Anthropic reported in February that DeepSeek, Moonshot and MiniMax made more than 16 million queries to Claude, with some 24,000 fraudulent accounts, to improve their models [216]. In the other direction, OpenAI, Amazon and Google, which keeps Gemini closed, signed the letter in favor of open weights. The U.S. government, which buys closed models, promotes open ones in the AI action plan the White House published in July 2025 [217] [30].
The espionage story has a documented origin in press reports on the conclusions of intelligence agencies and in the congressional letter that took them up, and it centers on China’s access to NVIDIA’s chips [172] [176]. There is no public sign that the dispute over models caused it.
Even so, the two readings are not mutually exclusive, because chips and open models are part of the same NVIDIA bet. The company gains the more people run models on their own hardware, because that hardware carries its chips, and many of the most used open models are Chinese. When it announced the purchase of Hugging Face, NVIDIA itself warned that others are lobbying the U.S. government to restrict open source models and that many of the most popular ones in the world originated in China [24]. The pressure to block chips from reaching China and the pressure against Chinese open models hit NVIDIA in the same bet, and Huang’s fall leaves it with less of a voice in both debates. TechCrunch wrote that Huang’s departure strengthens those in the U.S. government and Congress who call for limits on Chinese AI companies and puts his campaign against chip controls under scrutiny. Nathan Lambert, author of the Interconnects newsletter on AI models [218], called it very bad for the cause of open models [219].
The open coalition loses its most visible spokesperson just as it is being decided which Congress will debate the laws that affect it. Bills not passed before January 3 expire that day, when the Congress elected on November 3 is seated, with a Democratic majority in both chambers (see “Who regulates”). That is why Huang’s fall weighs in this study beyond the scandal.
05
What open and closed mean
Almost all frontier open models publish the weights and the code to run them, but not the data or the training code. The weights, or parameters, are the numbers the model learns during training and that determine each of its answers [220]. Anyone who downloads the parameters of an open-weights model can run it without going through its developer. Open source, in the strict sense, requires more. The Open Source Initiative is the nonprofit organization that, since 1998, has stewarded the open source definition for software and approves the licenses that meet it [221]. The definition the Open Source Initiative published for AI on October 28, 2024, requires three components: enough information about the training data to rebuild an equivalent system, the complete code to train and run it, and the parameters. All three must be delivered under terms that allow the system to be used for any purpose without asking permission [222]. When the organization itself validated its definition, OLMo, Pythia, Amber, CrystalCoder and T5 met it. Llama 2, Grok, Phi-2 and Mixtral did not, because they lacked components [223].
Among frontier models, the exceptions are few. Olmo 3, from the Allen Institute, a nonprofit AI research institute based in Seattle [224], updated as Olmo 3.1, publishes data, code, weights and intermediate checkpoints, which are copies of the weights saved midway through training [225]. NVIDIA’s Nemotron 3 publishes the code, the training recipes and the part of the data it is able to redistribute [226].
Open model licenses are of three types.
- Permissive. The MIT and Apache 2.0 licenses allow use, modification and sale without asking permission. MIT only asks that the copyright notice and the license text be kept. Apache 2.0 also asks that changes be noted and grants a license to the contributors’ patents, which is lost if the user sues over those patents [227]. They are used by DeepSeek, whose R1 license expressly authorizes distillation, which consists of training another model on its answers; Google, in Gemma 4 since April 2026; OpenAI, in gpt-oss, which it accompanies with a usage policy; Microsoft, in Phi-4, and IBM, in Granite [228] [229] [230] [231] [232].
- Restricted. The Llama 4 license requires asking Meta’s permission above 700 million monthly users. It also requires complying with a use policy that excludes individuals and companies in the European Union from its multimodal models, those that understand images, audio or video in addition to text [233] [234]. That license keeps the restrictions of the Llama 3.x license, which for the Open Source Initiative “is not Open Source” [235]. Tencent’s community license excludes the European Union, the United Kingdom and South Korea, that is, it only grants use rights outside those territories [236]. Grok 2’s license allows fine-tuning it, that is, training it a little more on one’s own data to specialize it, but bars using it to train other general models [237].
- Permissive with a gate for the big players. In 2026, several licenses of the main open models adopted a design that grants broad use and reserves a separate contract for the large provider of model services. The licenses measure that provider by its revenue or by its business. The business is model as a service, which the Kimi K3 license defines as giving a third party access to the model’s outputs or fine-tuning, so that the third party controls what it asks of the model or the data it trains it with [238].
| Model | Developer | Condition for the large service provider |
|---|---|---|
| Kimi K3 and K3.1 Pro | Moonshot (China) | Separate contract for all commercial use if the licensee’s group runs a model-as-a-service business and earns more than US$20 million in 12 months [238] [239] |
| Qwen3.8 with 2.4 trillion parameters | Alibaba (China) | Separate license if the group runs a model-as-a-service or AI work assistant business and earns more than US$50 million in 12 months [240] |
| GLM-5.3 | Z.ai (China) | Security review by Z.ai for all commercial use, including internal use, if the group runs a model-as-a-service business and earns more than US$10 billion in 12 months [241] |
| MiniMax M3 | MiniMax (China) | Below US$20 million in annual revenue, commercial use with notice and the legend “Built with MiniMax M3”. Above that, written authorization. Bans all military use [242] |
| Muse Spark Open | Meta (U.S.) | Notice and separate contract for anyone offering it as a service above US$20 million; the download on Hugging Face requires a request that Meta reviews [243] |
| Mistral Large 4 | Mistral (France) | Separate commercial license for any model-as-a-service business, with no revenue threshold [244] |
That separate contract is one of the ways open developers get paid today. With Kimi, Qwen and Muse, the company that runs the model on its own hardware for internal use pays nothing. The large reseller that offers it through an API, the connection through which other programs query someone else’s model and pay per use, does pay. Z.ai’s license is the only one that reserves a security review for its developer, and the only one that does not exempt the large group’s internal use.
Control. The provider of a closed model can monitor use, filter requests, correct the model and withdraw it. Whoever publishes an open one can no longer do any of that. As the public affairs study sums up, an open model cannot be recalled, and its safeguards come off with fine-tuning [245]. Safeguards are what the model is taught so that it refuses to help cause harm, and the fine-tuning that removes them, which is the fine-tuning described above, is cheap. In 2023, undoing the safety training of the 70-billion-parameter Llama 2 cost less than US$200 on a single GPU, the type of chip used to train and run almost all models [246].
Closed models share that weakness. Fine-tuning GPT-3.5 through OpenAI’s API on ten harmful examples, for less than US$0.20, also broke its guardrails. The difference is that the closed provider can monitor or cut off that fine-tuning [247].
According to Kyle O’Brien, of EleutherAI, a nonprofit AI research lab [248], and other researchers, one promising defense is to filter dangerous topics out before training. Their 6.9-billion-parameter models, trained that way, withstood up to 10,000 steps of malicious fine-tuning, that is, 10,000 rounds of retraining on harmful examples, although they can still use dangerous information they receive during the conversation, for example from a search engine [249]. Moonshot says it did something similar with Kimi K3.1 Pro: before opening it, it removed cybersecurity data from two stages of its training, the mid-training stage and the reinforcement learning stage, in which the model learns from the rewards its best answers receive [239].
Data. An open model runs on the user’s own hardware and sends no information to a third party. The U.S. action plan supports open models because many companies and governments have sensitive data they cannot send to a closed provider [30]. The advantage applies above all to mid-sized models, because frontier open models require one’s own data center or someone else’s. In someone else’s, the information once again passes through a third party.
Risk. An open model cannot be withdrawn after publication. The 2025 International AI Safety Report warns that there is no practical way to revoke access even if flaws are discovered later. That report synthesizes research on the capabilities and risks of advanced AI, without recommending policies. It was agreed by 30 countries at the 2023 Bletchley summit and is chaired by Yoshua Bengio [250]. The National Telecommunications and Information Administration (NTIA), of the U.S. Department of Commerce, warns that the release of weights cannot be reversed [251] [220]. Copies of an open model can lose their safeguards but not gain new ones, because the developer cannot correct copies that have already been downloaded.
The risk of an open model is usually assessed as marginal risk, that is, by what it adds compared with what already exists. Kapoor, Bommasani and others proposed that framework in 2024. The 2025 International AI Safety Report described it as a high-level consensus, and warned that those increments, taken together, can become large even if each seems acceptable [252] [251]. The 2026 edition presents it as one approach among others, complex in practice. It warns that the stolen weights of a closed model would pose risks similar to those of an open model. It also records that as of December 2025 there had been no confirmed, publicly documented theft of weights [253]. The NTIA concluded in July 2024 that the evidence was not enough to justify restrictions or to rule them out in the future, and recommended monitoring [220].
The 2026 evidence points both ways. FAR.AI, a nonprofit that researches AI safety [254], found in May that DeepSeek V4 Pro’s safeguards against chemical, biological, radiological, nuclear, cyber and terrorist requests gave way between 98% and 100% of the time against low-skill attackers [255]. The UK AI Security Institute (AISI) and CAISI concluded in July, before Kimi K3’s weights were published, that the model’s safeguards allow it to help develop cyberattacks [256]. CAISI is the center of the U.S. National Institute of Standards and Technology (NIST) that serves as the government’s main point of contact with industry for testing models, and it evaluates the capabilities that may threaten national security [257]. Since October 2026 NIST has called it CAISSI, the Center for Advancing Innovation and Standards for Super Intelligence, because Executive Order 14434, of September 29, ordered agencies to say “Super Intelligence” in place of “Artificial Intelligence”. This study keeps the acronym CAISI, which is the one its evaluations use [258] [259].
In contrast, the first randomized trial with novices in a real laboratory, in which a lottery decided who worked with a model and who worked without one in order to compare their results, did not find that the models significantly helped them complete the biology procedures. That trial was done with models from mid-2025 [260].
06
Cybersecurity
Cybersecurity is today the field where the dispute between open and closed models is already measured in real cases. In October, Fortune gathered attacks and defenses carried out with models of both kinds [261]. Since September there has been an open model that, according to Anthropic, develops complete cyberattacks almost at the pace of Claude Mythos Preview, the model that company presented in April [68]. In November, the FBI documented that attackers prefer cheap Chinese models to automate their intrusions [69]. The closed labs, for their part, reserve their most cyber-capable models for verified defenders [262]. Meanwhile, AI already finds serious flaws faster than many systems can be patched [263].
The open model that develops attacks
We believe that GLM-5.3, from Z.ai, changed the calculation of marginal risk described above, which measures what an open model adds compared with what already exists (see “What open and closed mean”), because what this model adds is no longer compared with other open models, but with the closed ones that only verified defenders receive. CAISI, the AI evaluation center of the U.S. Department of Commerce, called it in September the most cyber-capable open model published to date. On an index that summarizes its cybersecurity tests, it placed the model about four months behind a U.S. frontier that includes restricted-access models [264]. The evaluation was published when the weights were already public, because Z.ai had released them two weeks after the August 14 launch. That is why it describes a risk that was already circulating and could not condition the release. CAISI used four benchmarks, fixed sets of tasks used to measure and compare models, in which the model works as an agent, that is, it runs commands and programs on its own, and must find flaws in programs and turn them into attacks [264].
Anthropic, which competes with Z.ai, tested it against Chrome’s JavaScript engine, the part of the browser that runs the programs on web pages. In that test, GLM-5.3 achieved a complete attack, from the known flaw to the program that exploits it, in 50 of 410 attempts, and Claude Mythos Preview, which Anthropic reserved for defenders and has since surpassed with Mythos 5.1, achieved it in 56. A model anyone can download performed almost as well as a restricted one. GLM-5.3’s safeguards gave way between 64% and 100% of the time to simple techniques, such as telling it that it was taking part in a simulated attack exercise, which did not work against the Claude models with their safeguards. Because the weights are public, anyone can modify the model. Removing its refusals, the answers with which it turns down a request, cost Anthropic’s team about US$4,400 in compute. By the company’s estimate, an experienced team would spend about US$1,200. A few days after the launch, several developers published copies without refusals [68].
Anthropic acknowledges that those capabilities also serve defenders, and asks governments to evaluate GLM-5.3’s successors [68]. According to Epoch AI, an organization that studies the trajectory of AI and compares models with its own capabilities index, in October the open frontier in cybersecurity was where the closed frontier had been six or seven months earlier [265].
Copies without refusals circulate in large numbers. The most common technique for making them is abliteration, which locates the direction in the model’s computations that produces the refusal and removes it, without retraining the model. According to a 10a Labs study published on arXiv, the open-access repository of scientific papers, 3,471 original models without refusals were published on Hugging Face between January 2024 and March 2026, and each was republished 2.4 times on average, in compressed versions. In the first quarter of 2026, a single abliteration tool, Heretic, produced 54% of new models. Of 1,643 GitHub applications that use those models, 411, one in four, were openly malicious. The study concludes that removing a model from its creator does not reduce how widely it circulates, because the copies are already in other hands [266].
What attackers use
Attackers use both kinds of model, and they choose by cost as much as by the lack of safeguards. Georgetown University’s Center for Security and Emerging Technology (CSET) is a research center that studies how emerging technologies affect national security [267]. For Jessica Ji, a researcher at that center, open models are harder to set up, but cheaper [261]. Whoever offers a model as a service, open or closed, can see that use and cut it off. By contrast, an open model installed on a compromised machine passes through no one, and no provider can detect or cut off what the attacker does with it.
An open frontier model now fits on a personal computer. Colibri, an open-source program under the Apache 2.0 license, released in July 2026, keeps in memory the part of the model that is always used and reads the “experts”, the specialized parts the model activates depending on the task, from disk only when it needs them. With GLM-5.2, which has 744 billion parameters, it generates 0.05 to 0.1 tokens per second on a laptop with 25 GB of memory and 1.83 on a computer with 128 GB, according to its repository [268]. Tokens are the pieces into which the model cuts text. We believe that slowness does not hinder an attacker in no hurry, and that the cost of hardware stops being a filter on who uses an open model without anyone seeing it.
On November 13 the FBI published a FLASH alert on ARTEX. With that kind of notice, which can be shared without restriction, the FBI describes the tactics of an attack and the indicators for detecting it. ARTEX is an open-source tool that uses AI agents to automate the phases of an intrusion, from reconnaissance to data extraction. Criminals have used it since at least August against banks, governments and companies in defense, chips, energy and health, among other sectors. Some intrusions reached privileged access, that is, administrator permissions, in less than three hours, and a campaign against financial institutions in South Korea took out data on about 67,000 people. ARTEX does not come with a model of its own. It connects to Claude and GPT, and also to DeepSeek, GLM, Grok and Kimi. According to the FBI, attackers connect ARTEX mostly to Chinese models, especially the DeepSeek Flash models, because of their cost and the “lack of observable guardrails” when the task is framed as a penetration test, the simulated attack an organization commissions to measure its own defenses. Among the indicators of attack the FBI publishes are connections to the APIs of Anthropic, DeepSeek, Zhipu, which is Z.ai’s former name, and xAI, so attackers also use models as a service [69].
In September, Google observed a China-linked group that installs an open model on the infrastructure it compromises, to avoid the monitoring of commercial APIs [269]. In November it observed attack frameworks, programs like ARTEX that automate an intrusion, that connect commercial and open models to more than a hundred tools [270].
AI finds the flaws
The same capability serves both to attack and to defend, and the advantage goes to whoever receives it first, because they can find and fix the flaws before the other side exploits them. That is the argument of Anthropic, which competes in this market, to justify giving its most capable models to defenders first [271]. The company gave Claude Mythos Preview only to verified defenders, who used it to find more than 10,000 flaws in critical software [68]. It did so through Project Glasswing, which it announced on April 7, 2026. Twelve founding organizations, counting Anthropic itself and including Amazon Web Services, Apple, Google, Microsoft and the Linux Foundation, received the model, and so did more than 40 organizations that maintain critical software. Anthropic committed up to US$100 million in usage credits for all of them [271].
On November 9, Anthropic launched Mythos 5.2, also only for verified organizations: those admitted to its cybersecurity and life sciences verification programs, and Claude Enterprise customers through Claude Security. Fable 5.2, which is available to the general public, has classifiers, automatic filters that review each request and block high-risk cybersecurity tasks, among them penetration testing, attack generation and searching for flaws in compiled programs, those distributed already translated into machine language and without their source code, as well as most biology research tasks. A blocked request is answered by an earlier Anthropic model, Opus 4.8 if it concerns cybersecurity and Opus 5 if it concerns biology. Fable and Mythos are the same model; Mythos 5.2 lacks some of those cybersecurity and biology safeguards [262]. The British government, for its part, says its pilots with frontier AI found more than a thousand flaws in state systems, and that its central vulnerability monitoring service cut the time to fix flaws from fifty days to eight [272].
On October 21, Wiz Research, the research team of the cloud security company Wiz, disclosed PHP2Shell, a group of critical flaws in the core of PHP, the central part of the language, registered in the global catalog of known flaws under two numbers, which are present in almost all of its versions since 5.4. PHP is the programming language that WordPress and Wikipedia run on, and according to Wiz it powers about 70% of the world’s websites. The flaws allow a remote attacker, without identifying themselves, to run their own code on the server, that is, to give orders to the machine that serves the site, or to read its memory, which may hold passwords and user data. They had been in the code for more than a decade, but exploiting them required a chain of steps at the limit of what a human researcher could achieve. Wiz says that frontier models, among them Gemini 4 Argon and GPT-5.6 Sol, reduced the effort needed to turn those flaws into an attack [263]. That same day, the U.S. Cybersecurity and Infrastructure Security Agency (CISA) called for applying the patch immediately and disconnecting systems that could not be updated [273].
The number of known flaws is also growing. As of November 13, more than 100,000 registered flaws had been published in 2026, more than double the total for all of 2025. Each registered flaw receives a unique identifier from the CVE program, assigned by the organizations that program authorizes. VulnCheck, which is one of them and keeps the count, attributes the rise mainly to AI-driven searching [274]. Google also observed attackers who use AI to compare the code before and after each patch, and so turn newly fixed flaws into attacks, without human intervention [270].
The week of the outages
From November 12 to 17, the United States saw outages at Microsoft 365, the air traffic of the entire country, Workday, Chase, United, Apple’s services and NSANet, which is the classified network of the National Security Agency (NSA). None of those outages was attributed to an attack. Microsoft attributed its own to a change in its network, and said it had no connection to attacks or to AI or agents [275]. The Federal Aviation Administration (FAA) attributed the air traffic outage to a communications failure linked to the installation of a router [276] [277]. Chase spoke of a software failure in its equipment [278], and Apple, of an internal update [279].
The link between those outages and AI is indirect. According to the Director of National Intelligence, Jay Clayton, the NSA’s network went down because some systems did not restart when the patch for the PHP flaw was applied to them, and there are no signs of a coordinated attack or of action by a state [280]. The U.S. Secretary of Homeland Security, Markwayne Mullin, suggested without evidence that AI could overwhelm old systems [281]. We believe the NSA outage shows a new cost of defense: when AI finds serious flaws in software that runs everywhere, the urgent patch can also bring systems down. Clayton asked the government and industry to attend to cybersecurity before the launch of each new frontier model [280].
Those who serve open models are targets too
We believe the risk of theft also reaches the companies that serve open models and store their customers’ fine-tuned models. Fireworks, which serves open models through an API, reported two intrusions. The first lasted from August 3 to October 7, and the second, from October 23 to 30. In the second, the intruders took the account and billing data of all accounts created before October 30. In both they took data, model files or secrets from a small number of customers. The secrets are the keys and credentials that customers store on the platform. In November, Fireworks linked the evidence to a criminal group that resells illicit access to models [282]. Google had already warned that model weights and companies’ compute quotas, the cloud capacity they have contracted and that an intruder can use or resell, are targets of espionage, extortion and theft [269].
How governments respond
Almost the entire official response rests on evaluating models before they are launched, with early access for governments. With an open model, that control ends on the day its weights are published, because, as the NTIA warned in 2024, any restriction is more effective before the weights spread [220].
The British AI minister, Kanishka Narayan, said on October 12 that increasingly capable models “are distributed openly”, and that this makes it impossible for any country to be an island, that is, to protect itself only with its own rules. Narayan asked the labs not to launch a system if they cannot control its risks. His government seeks to keep, together with that of the United States, its AISI’s access to the most powerful models before they come out [272]. With that access, the institute’s scientists test each model before launch, to find its risks and review its safeguards [272]. On October 26, that country’s AISI published a conclusion on open biology models. Filtering the training data, which is the defense O'Brien and others called promising (see “What open and closed mean”), raises the cost of recovering a capability, but does not eliminate it. A fine-tuning with the removed data brought the capability back for about US$80 in compute in the largest model tested [283].
On November 6, Japan approved an emergency package. Under it, the Japanese AI safety institute will quickly evaluate the attack capability of the most powerful models, and the government will seek to have frontier companies hand over flaws to it before publishing them, that is, the information on flaws not yet disclosed [284].
In the United States, the guidance that the National Institute of Standards and Technology (NIST), the NSA and CISA published on October 16 asks frontier developers, on a voluntary basis, to keep weights and checkpoints so that an incident can be investigated. The document brings together ten practices for preventing, detecting and handling incidents that occur during the training and evaluation of models. The guidance does not distinguish between open and closed models [285]. Bill H.R. 10981, introduced in the House of Representatives on November 10, would make the NSA’s prior evaluation, which is voluntary today, mandatory (see “Pre-deployment evaluation and transparency”) [286]. California responded through the courts. Its attorney general sued OpenAI on November 16 for losing control of its agents in tests, among them the attack on Hugging Face (see “Liability for harm”) [287].
In China, the cybersecurity standards committee scheduled a national standard on October 13, which will be a security guide for the open release of AI models and has a deadline of December 2027 [288]. That committee, known as TC260, is the National Cybersecurity Standardization Technical Committee, and the standard is in a plan of recommended national standards, that is, compliance with it is not mandatory [288].
China also controls the outflow of technology with a catalog of technologies whose export is prohibited or restricted, managed by the Ministry of Commerce and the Ministry of Science and Technology. Technology the catalog prohibits cannot be exported, and technology it restricts needs a license [289]. The proposed revision of the catalog was open for comment from October 16 to November 15, and it leaves out technologies that are already in the public domain because they were published as open source or in academic papers [290]. As a result, that control does not reach a model whose weights have already been published. The proposal prohibits exporting the autonomous AI generation of malicious code capable of attacking critical infrastructure, and restricts the export of AI-assisted flaw searching [290].
07
Who is on each side
The closed labs of the United States
Anthropic does not publish weights. Its chief executive, Dario Amodei, wrote on July 27, 2026 that Anthropic “has never advocated for a ban on open-weights models”, and that those models can pose a greater risk than closed ones, whatever country they come from. His main concern is that authoritarian governments will build more powerful models than those of the United States. The secondary one is that models will be used in cyber or biological attacks, or will have serious alignment problems, that is, that they will not act according to the intentions of those who make and use them. Amodei cites the British AISI, according to which publishing the weights of a model with dangerous capabilities creates a “persistent and irreversible” risk of misuse. He proposes three measures: not selling advanced chips or chipmaking equipment to China, going after industrial-scale distillation, which consists of training a model on another model’s answers, and subjecting every sufficiently capable model, open or closed, to mandatory safety testing [291]. Anthropic did not sign the July 24 letter in favor of open weights (see “The open camp in the United States”) [217].
OpenAI published gpt-oss on August 5, 2025. They are two models under the Apache 2.0 license and a usage policy, and they were its first open-weights language models since GPT-2, from 2019 [292] [230]. Before publishing them, the company deliberately fine-tuned them to do harm in biology and cybersecurity. It concluded that, even so, they stayed below o3, one of its own models that does not reach its high-capability threshold, the level set by the preparedness framework with which the company itself classifies the risk of its models, and it said those results weighed in the decision to publish them. This is an evaluation by the developer itself [293]. Since then OpenAI has not published any new general model, only derived or specialized models. Among them are gpt-oss-safeguard, which classifies content against safety policies and comes in 120-billion and 20-billion-parameter versions, and a personal-data filter [294]. OpenAI signed the July letter when it was expanded [217].
Google keeps Gemini closed. On April 2, 2026 it published Gemma 4, the first version of that family under the Apache 2.0 license. The company says Gemma models have passed 400 million downloads [229]. Google also signed the letter [217].
xAI published Grok 1 under the Apache license and Grok 2 under a license of its own. Elon Musk promised the weights of Grok 3, which to date are not on Hugging Face [295] [237]. Since February 2026, xAI has belonged to SpaceX [296]. SpaceX signed the letter, and xAI does not appear in it under its own name [217].
In practice, the closed labs keep their main model and open smaller versions. The gap between open and closed models is short. On Epoch AI’s capabilities index, since January 2026 the best open models have trailed closed ones by four months on average, or six under a stricter criterion, which requires the open model’s score to exceed the closed one’s and not merely that the difference be so small it can no longer be told apart from the margin of error [31]. The British AISI estimated that gap at four to eight months in December 2025 [32]. In cybersecurity it put it at four to seven months in July 2026, against six to ten a year earlier [33]. The institute, which reports to the British science and technology ministry, measures that gap with 70 attack tasks and with simulated networks. It looks for the closed model that performs the same as each open model, and counts the months between the launches of the two [33].
CAISI calculated in May that DeepSeek V4 Pro was about eight months behind the U.S. frontier [34]. Arena ranks models with the votes of users who compare two models’ answers to the same question without knowing which is which. On that leaderboard, according to the AI Index, the AI data report published by Stanford University’s HAI institute, the best closed model’s score exceeds the best open model’s by 3.3%, against 0.5% in August 2024, so on that measure the gap grew [35].
The open camp in the United States
Meta was the main promoter of open models and is now hybrid, because it publishes some models and keeps others. In 2024, Mark Zuckerberg wrote that selling access to models “isn’t our business model”, and that closing them to slow China “is just not going to work” and would only disadvantage the United States and its allies [297]. In July 2025 he warned that it would be necessary to be careful about what is opened [298]. Llama has had no new version since April 2025 [299]. Muse Spark, the first model from its superintelligence lab, came out closed on April 8, 2026, with API access only, according to Epoch AI’s records and the press [31] [300].
On August 10, Zuckerberg announced that Meta would publish some open models again, and that restricting that ecosystem would be a mistake. He wrote: “I do not believe restricting access to foreign open source models is an effective solution” [301]. That same day, Meta published Muse Glimmer, of about 30 billion parameters and under the Apache 2.0 license [302]. On October 15 it published Muse Spark Open, of 816 billion parameters and under a license of its own [303].
NVIDIA trains the Nemotron family and is, according to Hugging Face, the largest contributor of open models [23]. This study covers it separately below (see “NVIDIA between the two sides”). Microsoft publishes Phi under the MIT license, and IBM publishes Granite under Apache 2.0 [231] [232]. Reflection, which counts NVIDIA among its investors, published Beam in October, of 501 billion parameters, under Apache 2.0. The public affairs study compares its results with those of the Chinese models [304] [305].
The open coalition was formalized on July 24, 2026 with the letter “Open Weights and American AI Leadership”. The first version, published by Microsoft, had 25 signatories, among them Meta, Microsoft, NVIDIA, IBM, Palantir, CrowdStrike, Dell, Mistral, Hugging Face, Andreessen Horowitz and Y Combinator [306]. The July 29 version already had 235, including Amazon, AMD, Google, Intel, OpenAI and SpaceX [217], and on August 3 Microsoft counted more than 270 [307]. The signatory counts are those of each published version, and the header of every version keeps the July 24 date.
The letter holds that U.S. leadership will be judged by whether it builds “a strong, open ecosystem”. It acknowledges that open weights carry real risks, because once published they escape the control of whoever made them, and it asks to “keep the frontier plural by avoiding premature restrictions on open models”. It also warns against confusing legitimate distillation with misappropriation, and asks to address the latter with specific rules [306].
According to Fortune, Huang shared the letter in his first post on X, saying that the world needs both closed and open frontier models [308]. In September, Huang wrote that he was a co-author of the letter [25]. The letter came out eight days after the launch of Kimi K3, the open model from the Chinese company Moonshot, on July 16. In those days, according to Fortune, the U.S. government was debating whether to restrict Chinese open models like that one, and Michael Kratsios, Director of the White House Office of Science and Technology Policy, accused Moonshot of distilling U.S. models [308].
China
Most of the leading Chinese labs publish their flagship models openly. DeepSeek uses the MIT license for V3, R1, V4 and V4.1 [228] [309]. Alibaba publishes most Qwen models under Apache 2.0. Its largest model, Qwen3.8, of 2.4 trillion parameters, and others such as Qwen3.8-Flash-Next carry licenses of its own, which require a contract from the large reseller, the group that offers models as a service and exceeds a certain revenue (see “What open and closed mean”) [240]. Moonshot published Kimi K2 under a modified MIT license, and Kimi K3, of 2.8 trillion parameters, under a license of its own. On November 4 it published Kimi K3.1 Pro, of the same size and with the same condition, which is a separate contract with Moonshot for the large reseller [238] [239]. Z.ai published GLM-5 under the MIT license, and GLM-5.3 with the security review already described, which its license reserves to the company with respect to the largest groups [241]. Tencent published Hy3 under Apache 2.0 [310]. The exception is ByteDance, which reserves its main model, Seed 2.0, for its own service, which it offers through Volcano Engine and Doubao. Even so, ByteDance published smaller models under Apache 2.0, such as Seed-OSS-36B [311] [312].
That bet dominates adoption:
- Chinese models accounted for 41% of downloads on Hugging Face over the past year, according to the platform itself. Hugging Face warns that automated systems inflate the downloads of small models, so the figure does not equal real use [28].
- In March 2026, Chinese models had 1.15 billion cumulative downloads, against 723 million for U.S. models. In February they accounted for 70% of new derivative models. Both figures come from the ATOM report, by Nathan Lambert and Florian Brand, which tracks the adoption of about 1,500 open models and assigns country by the publisher’s headquarters [29].
- OpenRouter is a platform that gives access to hundreds of models. Use there is measured in tokens, the pieces into which the model cuts text, the size of a short word or part of a long one. On that platform, open models reached about a third of tokens by late 2025, and Chinese open models averaged 13% of weekly volume over the year [313]. The public affairs study records a week in July 2026 when the Chinese share passed half. It also records U.S. companies that switched to Kimi and GLM, or that build on them [304].
- In the week of September 14, 2026, Chinese models accounted for between 57% and 67% of OpenRouter tokens, against 6% to 13% in February. On Vercel, a platform where developers publish websites and applications and connect AI models to them, they reached 55% in August, against 11% in January. The data come from CNBC, which adds that U.S. frontier models still take more of the spending [314].
- Lindy, a U.S. startup, moved all of its traffic from Anthropic’s Claude models to DeepSeek in June 2026. Its chief executive said that “you could see that cost curve go down, like, crash to the ground” [315]. Brian Chesky, Airbnb’s chief executive, said in October 2025 that its customer service agent relies heavily on Alibaba’s Qwen, because it is “very good” and “fast and cheap” [316].
According to Stanford’s AI Index, in March 2026 the best U.S. model led the best Chinese model by 2.7% on the Arena leaderboard, and that gap stayed in single digits over the past year [317]. The report contradicts itself about who was ahead. Its text says that the two countries’ models traded first place several times since early 2025, but the only case it gives is a tie: DeepSeek-R1 briefly matched the best U.S. model in February 2025. The chart in the same report shows the scores, and in it the U.S. model stays on top for the whole period, with 1,503 points against 1,464 in 2026 [317]. We believe the two versions are explained by Arena’s method. Until November 2025, Arena gave each model the best rank compatible with its margin of error, so two models with different scores could both appear in first place [318]. A Chinese model could thus share first place without ever exceeding the U.S. model’s score. The contradiction is recorded because anyone who consults the source will find both versions and because it changes the reading of the data: according to the scores, the U.S. lead is small but constant, and the ties for first place do not indicate that a Chinese model was ever ahead.
Chinese developers charge for the service, not for the weights. Alibaba reported 45% growth in its cloud and AI compute services. The company says that, in the long run, Qwen’s open strategy, with more than 3 billion downloads, turns adoption into business opportunities for its cloud [319]. DeepSeek published in 2025 that, at its prices at the time, its service would have a theoretical margin of 545% over cost, that is, it would charge about six and a half times what it costs to serve, although its actual revenue was lower because its website and app are free [320]. Opening models also has a cost, because other providers can serve the same model and take the traffic. The public affairs study shows that, twenty days after Z.ai released the weights of its model, the share of that model’s traffic that went through Z.ai itself had fallen from 88% to 22% [321].
U.S. government evaluations find weak safeguards in Chinese models. CAISI found in September 2025 that DeepSeek R1-0528 complied with 94% of overtly harmful requests when common jailbreak techniques were applied, against 8% for the U.S. reference models. Those techniques are ways of wording a request so that the model ignores its safeguards. It also found that DeepSeek’s models echoed Chinese Communist Party narratives four times as often, on average, as the U.S. reference models [322]. The evaluation responds to the mandate in the White House AI action plan of July 2025 (see “The state sponsors”), and it compared three DeepSeek models with four U.S. models on 19 benchmarks [322].
The same test included gpt-oss, OpenAI’s open model. For writing phishing emails, those that pose as a person or a company to deceive the recipient, gpt-oss proved as easy to hijack as R1-0528, that is, to divert from the user’s task with malicious instructions hidden in the content it reads while working as an agent: it gave way in 48% of cases, against 3% for the closed models. In the other tasks it resisted better than R1-0528. It gave way in 27% of cases against 37% for stealing credentials, and in 28% against 49% for downloading malicious programs. Faced with jailbreak techniques it complied with between 6% and 10% of requests, against more than 90% for the DeepSeek models. Even so, both open models fell far behind the closed ones, which did not exceed 4% [322].
In September 2026, CAISI called GLM-5.3 the most cyber-capable open model published to date, about four months behind a U.S. frontier that includes restricted-access models [264]. Of GLM-5.2 it said that its safeguards can be circumvented when the model is run on one’s own hardware [323].
Independent studies confirm the censorship. On questions asked in Chinese to 2023 and 2024 models, Chinese models refuse far more political questions than the rest: DeepSeek refuses 36%, and non-Chinese models, between 0% and 2.8% [324]. Bias, moreover, passes to other models through training data. According to Aleph Alpha, a German company that competes with Chinese providers, between 17% and 18% of the answers of NVIDIA’s Nemotron Cascade 2 repeat those narratives. The cause, according to the company, is about 3,500 rows with Party slogans among 9.3 million fine-tuning examples, generated mostly with DeepSeek and Qwen [325].
Europe and the rest
Outside the two powers, several governments that fund their own models publish them openly in the name of sovereignty, that is, so as not to depend on a foreign provider that can change prices, terms or access.
- Switzerland. EPFL, ETH Zurich and the national supercomputing center published Apertus in September 2025, in 8-billion and 70-billion-parameter versions. They published it with weights, checkpoints, code and data documentation, and with a license that allows commercial use [326]. The model is part of the Swiss AI Initiative, led by those two public universities, and it was trained on Alps, the national center’s supercomputer, with more than 10 million GPU hours, the equivalent of 10,000 of those chips working nonstop for about six weeks [326]. Version 1.5, which is multimodal and carries the Apache 2.0 license, came out in 2026 [327].
- United Arab Emirates. The Technology Innovation Institute, the applied research arm of the Abu Dhabi government’s Advanced Technology Research Council, publishes the Falcon family. In 2026 it released on Hugging Face Falcon Perception, a 300-million-parameter vision model, “strengthening sovereign AI capabilities” [328]. MBZUAI, an Abu Dhabi university devoted entirely to AI, and the company G42 published K2 Think in September 2025, of 32 billion parameters, with data, weights and code [329].
- India. The IndiaAI Mission, which the government launched in 2024, funds twelve organizations and consortia to build their own foundation models, the general models on which others build applications, and the support covers the cost of compute. The models of Sarvam and BharatGen are published on AIKosh, the government’s public platform, a shared library of models anyone can use. The government presents that publication as a contribution to open source [330].
- Japan. On July 14, 2026, the cabinet approved the revision of the Basic AI Plan (the first plan dates from December 23, 2025). The plan is the instrument that Japan’s 2025 AI promotion law requires the government to adopt. The revision proposes a trustworthy ecosystem that includes open-source and open-weights models, an “open AI sovereignty” and avoiding excessive dependence on any one country or company. The government will open-source Gennai, its generative AI platform, which 180,000 civil servants already use, so that municipalities can adopt it too [331].
- South Korea. The Ministry of Science’s sovereign foundation model program selected five teams in August 2025. The government gives them GPUs, training data and support for hiring foreign researchers, and it planned a first evaluation to cut them to four. The teams committed to open policies, so that other companies can use their models for commercial purposes [332]. In the first evaluation, in January 2026, Naver was dropped because it used an external encoder with its weights, without retraining them. The encoder is the component that converts video and audio into data the model can process. The ministry set as a minimum condition training from initialized weights, that is, emptying the weights and forming them again with one’s own data, although it allows the use of verified open source [333].
- Russia. Its foundation model law, in force since September 1, 2026, distinguishes two categories, which apply from March 1, 2027 [334]: sovereign models, developed entirely in the country, and national ones, which may use foreign open components. The law also sets the basic definitions of AI and of large foundation models, imposes obligations on their developers and creates a general liability for violations in the sector [334]. Sber, for its part, published GigaChat 3.1 Ultra openly. Both facts are in the public affairs study [335]. In September 2026, Sber also published GigaChat 3.5, of 432 billion parameters, under the MIT license [336].
- European Union. In February 2025 the European Commission gave its STEP seal, with 37.4 million euros, to OpenEuroLLM, a family of open-source models for all of the Union’s official languages [337]. The STEP seal is the mark of excellence of the Strategic Technologies for Europe Platform, and it makes it easier for a project to obtain other European funds. Of the 37.4 million in OpenEuroLLM’s budget, 20.6 million come from the Digital Europe program [337]. Mistral, the largest European model lab, publishes its small models openly, under Apache 2.0. In September 2026 it closed a 3 billion euro round led by Samsung, and it presents itself as the only company that builds open-weights models, infrastructure and products at the same time [162]. Its Large 4 model carries a license that requires a contract to resell it as a service. Others, such as Medium and Codestral, are reserved for the company’s own service [244].
The Korean case shows that a government that demands independence may not accept a national model built on someone else’s weights, even if those weights are public. The Korean ministry justifies it on national security grounds, because it wants to be able to develop and control the model on its own under any circumstances [333]. That is a limit of openness as the basis for a sovereign model. Russia allows foreign open components, but puts them in another category, that of national model.
08
The state sponsors
The United States sponsors both sides
The AI action plan that the White House published in July 2025 devotes a section to “encourage open-source and open-weight AI”. It holds that the United States needs leading open models, founded on American values, and that those models can become global standards, which gives them geostrategic value. It adds that the decision to publish a model open or closed “is fundamentally up to the developer”, and that the government should create a supportive environment for open models [30]. The same plan tasks CAISI with evaluating Chinese models for their alignment with Communist Party talking points and for their censorship [30].
The NSTM-4 memorandum from the Office of Science and Technology Policy, dated April 23, 2026, accuses entities “principally based in China” of deliberate industrial-scale distillation campaigns. It says there is “nothing open” about supposedly open models born from malicious exploitation. In the same text it reaffirms the commitment to open development, from open-source frameworks to open-weights models, and does not propose restricting them [338].
Executive Order 14409, of June 2, 2026, orders the federal government’s cyber defense to be strengthened with the help of AI. It also tasks the NSA with deciding, through a classified procedure, which models count as covered frontier models because of their cyber capability. For those models it creates voluntary government access up to 30 days before they are provided to other trusted partners, the organizations chosen to receive them before the public. The order forbids reading that section as a license, prior approval or mandatory permission to develop, publish or distribute models, so it does not prevent publishing the weights of any of them [339].
Order 14319, of July 2025, only allows federal agencies to buy language models that meet two principles, truth-seeking and ideological neutrality [340]. The Office of Management and Budget’s procurement guidance of December 2025 applies those principles to every language model an agency acquires. It exempts models acquired under a free and open license from its neutrality requirements, although it asks agencies to review them before use [341]. Order 14319, finally, asks to avoid, where possible, requiring disclosure of weights, that is, an agency forcing the provider to hand them over to verify compliance [340].
Federal contracts and voluntary pre-launch national security testing, by contrast, are carried out with the closed labs. In May 2026, CAISI signed agreements of that kind with Google DeepMind, Microsoft and xAI, adding to those it already had with Anthropic and OpenAI [342]. The public affairs study records that the White House told the labs those tests leave out Chinese open models [343]. Mark MacCarthy, in Tech Policy Press, reads the exclusion as covering all open models, U.S. ones included, and proposes extending the review to them [344]. CAISI does evaluate open models after they are published, such as the DeepSeek and Z.ai models cited above [34] [264].
The 2027 defense act, the annual law that authorizes military spending, treats weights as something to be protected. Its two versions, the one the U.S. House of Representatives passed on September 14 and the one reported by the Senate Committee on Armed Services, order the U.S. Department of Defense to build and test high-security data centers. Those centers must protect model weights against theft, sabotage or unauthorized access by adversary states. The Senate version also calls for rules requiring large AI contractors to report incidents that compromise their models or their weights [345]. For the Department of Defense, the weights of the models it contracts are an asset to be guarded instead of published. The House version counts them among the sensitive assets that must be protected from theft, sabotage or unauthorized access [345].
In Congress, the only bill devoted to supporting open models is the Open-Source AI Leadership Act, H.R. 10152. The House Committee on Energy and Commerce approved it on September 16 by 26 votes to 21. Since then its report, the step that follows committee approval, has not been filed, it has not had a floor vote, and the bill dies if it is not passed before January 3, 2027, when the 119th Congress ends [26].
The bill directs the U.S. Department of Commerce to support the adoption of American open models. It also directs it to assess and publish the risks of open models from foreign adversaries, which are China, Russia, Iran and North Korea and the entities under their control [346], including the risk that they were trained using other models without authorization. It further provides that nothing in the act authorizes it “to ban, restrict, or otherwise make unavailable an open artificial intelligence model” [347]. Some Democrats objected that the bill gives the administration authority without clear limits. Democrat Lori Trahan said the biggest threat to the open ecosystem is the government’s own regulatory uncertainty [348]. Another bill, H.R. 8094, on foundation model transparency, exempts fully open-source models from the rules it tasks the Federal Trade Commission with writing, without defining what counts as fully open source [349]. Those rules would require providers of the most used or most powerful models to publish the origin of their training data, foreseeable risks and their test results. The bill was introduced in March 2026 and referred to the Committee on Energy and Commerce [349].
China sponsors only the open side
In China, open source is written state policy. The State Council’s opinion on the “AI+” action, of August 2025, devotes one of its sections to it, “promote the prosperity of the open-source ecosystem”. With that opinion, the central government orders AI to be brought into six fields, among them science, industry, consumption and public administration. It sets as a goal that the use of AI agents and of intelligent terminals, the phones, computers, cars and other devices with built-in AI, exceed 70% in 2027 and 90% in 2030 [350]. In that section it calls for supporting open AI communities and the opening of models, tools and datasets. It encourages universities to count open-source contributions as student credits and faculty achievements, and calls for developing projects “with international influence”. In its global cooperation section, the opinion also calls for making the technology open and accessible to the Global South, that is, to developing countries [350]. The 15th Five-Year Plan, of March 2026, which is the country’s economic and social development program for the period from 2026 to 2030, calls for building the open-source system and improving its mechanisms [351].
The “AI+ Manufacturing” plan, signed in December 2025 by the Ministry of Industry and Information Technology and seven other bodies, sets for 2027 the goal of a “world-leading” open ecosystem, and orders the deployment of open projects for models, data and agents. Among its other goals for that year are 1,000 high-level industrial agents and three to five general models applied in depth in manufacturing. It proposes developing open-source licenses suited to AI projects and “a new order of open-source rules”. Its annexed guide encourages leading companies to offer their solutions by opening general models and tools [352].
The licenses that Z.ai, Alibaba and Moonshot adopted in 2026, with a security review or with a contract for the large reseller, coincide with that call [241] [240] [238]. However, Meta and Mistral adopted the same design. That is why we believe those licenses respond above all to a common commercial reason, which is to charge the reseller that takes the traffic (see “Who is on each side”).
Abroad, the Global AI Governance Action Plan, of July 2025, is a 13-point text that the Chinese government published at the World Artificial Intelligence Conference. The plan calls on governments, organizations and companies to act, without creating obligations. It proposes cross-border open-source communities, with technical security standards and with an open flow of “non-sensitive” technical resources [353]. The agreement creating the World AI Cooperation Organization, headquartered in Shanghai, was signed by 29 countries on July 16, 2026. The organization is an intergovernmental body that China promoted to advance international cooperation and the global governance of AI. The official note does not mention open source. Xinhua attributes to “the relevant parties” a program that includes jointly building an open ecosystem, but we have not found the text of the agreement [354] [355]. On November 3, the National Development and Reform Commission said the organization already had 38 founders and that work was under way to make it operational. It also said that China will promote AI open source “unwaveringly” [356].
The Chinese government defends its open models in the U.S. debate. The public affairs study records the response of its Ministry of Commerce to the distillation accusations [357]. In that response, the ministry warned that it would take countermeasures if the United States acted against its companies, and repeated that China “encourages open source” [358]. In July, the ministry’s spokesperson had pointed out that nearly 200 U.S. startups were asking their government not to cut off access to Chinese open models, because that would reduce the competitiveness of U.S. companies [359].
We found no public program in China aimed at closed models. ByteDance, which keeps Seed 2.0 to itself and publishes smaller models openly, does so by its own choice [311] [312]. The rules that bind Chinese models apply to the service offered to the public and not to the weights, as shown below (see “Pre-deployment evaluation and transparency”).
09
NVIDIA between the two sides
NVIDIA sells the chips that both camps use, although, according to SemiAnalysis, a market of open models suits it better, because it multiplies the number of buyers of chips to run them (see “What Huang says”) [27]. Data centers were 89.7% of its revenue in fiscal 2026, which ended in January, and China, together with Hong Kong, was 9.1%. The company itself says that in China it is effectively excluded from the data-center market, because it cannot offer a competitive product there that both the U.S. government and the Chinese government approve [11].
For the third quarter of its fiscal 2027, which runs from August to October, NVIDIA assumed no compute revenue from China [360]. Even so, it shipped to China data-center products of the Hopper generation, the one before Blackwell [11], for US$2.4 billion [144]. For the fourth quarter it again assumes no compute revenue from China [144]. On November 17 it said it is stopping sales in China of the H200, the chip for which the United States had granted it licenses to sell to Chinese customers [200].
NVIDIA funds the closed side. In 2025 it announced its intention to invest up to US$100 billion in OpenAI as its systems were deployed [361]. In March 2026 it replaced that intention with a US$30 billion equity investment, according to Jensen Huang, then chief executive of NVIDIA, and the press [362]. In August 2026 it gave residual value guarantees on the leases of a campus in Ohio, with a cumulative cap of US$105 billion. If the tenant stops paying or becomes insolvent, NVIDIA covers the difference between a guaranteed minimum value for each lease and whatever is recovered by leasing or selling it again. Each guarantee starts with its lease, from 2028, and ends after 20 years or earlier, for example if OpenAI obtains a satisfactory credit rating. The tenant is an OpenAI subsidiary, and OpenAI is obliged to reimburse NVIDIA for whatever it ends up paying [363].
It also invests in other labs. In 2025 it agreed to invest up to US$10 billion in Anthropic [212]. Anthropic’s IPO prospectus, dated November 10, also provides for NVIDIA to buy about US$10 billion of its shares in a concurrent private placement, that is, a direct sale of shares outside the public offering [56]. NVIDIA also took part in xAI’s January 2026 round [364]. As of July 26 it held US$99 billion in equity investments and US$25 billion committed to its ecosystem, which includes model developers [14].
Business Insider wrote that Huang’s fall removes a threat to Anthropic’s IPO [365]. Anthropic’s prospectus describes a proprietary orchestration layer with which it trains and serves Claude on four families of chips: Amazon Web Services’ Trainium, Google’s TPUs, and NVIDIA’s and AMD’s GPUs. According to the company, that layer reduces its dependence on any single vendor [56]. Michael Cembalest, of J.P. Morgan Asset Management, calculated that NVIDIA has 28% of the compute agreements that frontier labs announced for 2026 onward, when in 2024 and 2025 it had between 80% and 90% of AI chip sales. For Cembalest, that shift by the labs toward their own chips or other vendors’ explains why Huang defends open models so insistently [366]. We believe Huang’s fall favors the closed labs on two fronts: it weakens the main sponsor of the open side, and it comes when they depend less on his chips.
NVIDIA also funds the open side.
- It trains the Nemotron 3 family of models. Nano, with 30 billion parameters, was published in December 2025, and Super, with 120 billion, in March 2026, both under NVIDIA’s own license; Ultra, with 550 billion, was published in June 2026 and uses OpenMDW, a third-party open license [367] [368] [369]. OpenMDW is a license that the Linux Foundation and the PyTorch Foundation introduced in 2025 to publish a complete model, with its weights, its code and its documentation. It allows anyone to use, modify and redistribute the model without restrictions [370]. In October 2026 NVIDIA began publishing the Nemotron 3.5 family [371].
- Its multi-year cloud commitments are contracts to use the capacity of cloud providers for several years. NVIDIA devotes them to research and to DGX Cloud, its cloud computing service, and they totaled US$26 billion in October 2025 [212]. By August 2026 they were US$29 billion [14]. The company wrote that those commitments provide the infrastructure for researching and developing its open models, such as Nemotron, Cosmos and GR00T, and its autonomous driving software [14].
- On September 2, 2026 it signed the purchase of Hugging Face for about US$11.9 billion. It committed to keeping the platform open [24]. The agreement includes a retention program of up to about US$1 billion in shares, which are given to Hugging Face employees so that they stay after the purchase, and closing is expected in the first half of 2027 [24]. That closing requires regulatory approvals. According to Reuters, the U.S. Department of Justice is investigating NVIDIA’s conduct and, within that investigation, is asking whether the purchase would let it “capture a fundamental distribution point for AI” [372]. In September 2026, according to Reuters, citing the New York Times, the same department had begun investigating whether NVIDIA structured its deal with Groq, an inference technology company, to avoid antitrust review. In the United States, the Hart-Scott-Rodino Act requires large acquisitions to be notified to the Federal Trade Commission and the Department of Justice before they close, and a waiting period so that they can be reviewed [373]. The Groq deal was a technology license combined with hiring executives, not a purchase (see “What Huang foresees for compute”). MLex anticipated that this investigation would toughen the review of the Hugging Face purchase [374] [375].
- It invested in Mistral and took part in Reflection’s round, according to TechCrunch [376] [377].
In August 2026 NVIDIA wrote in a filing with the Securities and Exchange Commission that demand for open models “promotes use of our products worldwide” [14]. In its filing on the Hugging Face purchase it warned that others are lobbying the U.S. government to adopt measures that restrict or disadvantage open-source models, and that many of the world’s most popular open models originated in China [24]. Its annual report lists two more risks. One is that regulation could prevent it from supporting applications built on Chinese models, such as DeepSeek and Qwen. The other is that the Chinese government could discourage its customers from buying from it [11].
The Cyberspace Administration of China summoned NVIDIA in July 2025 to explain the risks of “vulnerabilities and backdoors,” hidden access to the chip, in the H20, the chip that NVIDIA adapted to export controls to sell it in China [378]. In September 2025, the State Administration for Market Regulation, China’s competition authority, concluded on a preliminary basis that NVIDIA had violated the Anti-Monopoly Law and the conditions under which its purchase of Mellanox was approved, without specifying which, and decided to keep investigating [379]. Mellanox is a maker of networking equipment for data centers. The authority approved its purchase in April 2020 with conditions. NVIDIA had to keep supplying the Chinese market with its accelerators, the chips specialized in AI calculations, and Mellanox equipment on fair, reasonable and nondiscriminatory terms, and could not require them to be bought together [379]. According to NVIDIA, the finding is that it discriminated against Chinese customers by complying with U.S. controls and offering China degraded products [14]. In November 2025 the White House announced that China would end its antitrust investigations into U.S. semiconductor companies, without naming NVIDIA. It announced this as part of the trade truce between the two countries (see “The fab”) [380]. In August 2026 NVIDIA was still describing the Chinese authority’s finding as a current risk [14].
The United States authorized the sale of small quantities of the H200 to specific Chinese customers, but the Chinese government restricted those purchases. In the first half of its fiscal 2027, NVIDIA recorded a charge, an accounting loss, of US$400 million for excess inventory and purchase commitments, because demand for the H200 fell. After the charge it made only a fraction of the authorized shipments, and those shipments were less than 1% of its data-center revenue for the quarter. The H20 had already cost it US$4.5 billion in 2025, when the United States required a license to sell it to China [14]. The U.S. government collects 25% on those sales. That 25% is a tariff on importing the chip into the United States, through which the licenses require it to pass for inspection before leaving for China [11]. NVIDIA has not been able to pass that tariff on to its customers [14]. Meanwhile, China is building its own chip supply chain (see “China makes its own chips”).
What Huang says
Huang does not set open against closed. “Proprietary versus open is not a thing. It’s proprietary and open,” he said in March 2026, at a session on open models at NVIDIA’s annual conference [381]. In Washington, in October 2025, he put it this way: “We have amazing proprietary models. We need also amazing open source models. Our country depends on it” [382].
Huang treats open as shared infrastructure. Without open source, he said in December 2025, “startups can’t thrive, university researchers can’t do research, you can’t teach AI, scientists can’t use AI” [383]. When he announced the purchase of Hugging Face, he wrote that open weights broaden access to AI and help spread its leadership among companies, institutions and communities [25]. For him, AI is what electricity and the internet once were, “the essential infrastructure of our time” [384].
In March 2026 he gave three reasons for NVIDIA to train its own open models. The first is co-design: researching model architectures, that is, how the computation is organized inside each one, such as that of Nemotron 3, which combines two different computing designs, shows NVIDIA what computers future models will need. The second is for AI to reach every industry, every country and every researcher, because “if everything is proprietary, it’s hard to do research.” The third is that AI is not only language, and someone has to take models of biology, chemistry, climate and physics to the frontier. NVIDIA does not make cars or discover drugs, and it wants the companies that do to have good models. In the same answer, Huang accepted that the best models, sold as products, “should be proprietary” [15].
From there comes his idea of sovereignty, under which each country should produce its own AI and not only consume it. “Every country needs to own the production of their own intelligence,” he said in Dubai in 2024, because AI “codifies your culture, your society’s intelligence, your common sense, your history” [385]. In 2025 he called on every country to use the large U.S. models and also build its own, trained on its language, its history and its values: “It’s not sensible that one Western company will be able to capture and somehow deeply appreciate the values of every country and every religion and every background and every society around the world” [386]. In Berlin, in October 2026, he said that in the AI economy consuming it is not enough and that it has to be produced [387].
With China, Huang separates competition from enmity. He calls China “our competitor and adversary, not our enemy,” because of the interdependence between the two countries [386]. Of its researchers he said: “These are not Chinese AI researchers, they’re world-class AI researchers,” and he noted that they abound at Anthropic, OpenAI and DeepMind [388]. In November 2025 he told the Financial Times that “China is going to win the AI race.” That same day he clarified in writing that China is “nanoseconds behind America” and that it is vital for the United States to win “by racing ahead and winning developers worldwide” [389].
In March he estimated that “50% of the world’s AI researchers are Chinese, plus or minus.” He explained the number of Chinese companies by the competition among provinces and cities, whose mayors compete with one another. He explained their openness by a culture of “family first, friends second, and company third,” in which engineers at rival companies are siblings, friends and schoolmates. “They’re essentially open source all the time” [15]. In April 2026 he said that the ecosystem “needs open models,” that “a lot of that is coming out of China” and that “we ought to not suffocate that.” In the same interview he warned that the day DeepSeek comes out first for Huawei’s chips will be “a horrible outcome for our nation” [390]. In September 2026 he said of China’s leaders that they want China to prosper, “just as we want America to prosper” [391].
On safety, Huang rejects fear and acknowledges the problem. Comparing AI to enriched uranium strikes him as “lunacy” [390]. “Scaring people is unnecessary. It is irresponsible,” he said in September 2026, and he argued that existing product liability laws are enough [391]. Those laws, which in the United States are state laws, make those who make and sell a product answer for the harm it causes [392]. Ten days later, when launching an open safety platform for agents, he said that AI’s potential “will only be realized if we solve AI safety,” and that this safety requires engineering at every layer [393]. Agents are programs that chain tasks together on their own, and an agent can access sensitive information, run code and communicate with the outside. Huang proposes that, to be safe, an agent should have only two of those three capabilities at any given moment [15].
The five ideas form a single stance. If AI is infrastructure, no one should hoard it, each country should produce its own, researchers from the two countries should keep working together and safety is built with engineering, not with exclusion.
Huang does not separate that stance from the business either. In July 2026, when the launch of Kimi K3, the model of the Chinese company Moonshot, knocked down NVIDIA’s stock on fears that cheap models would reduce the need for infrastructure, he told Axios that “free AI should be very good for chips” and for data centers, and that a world with a single model would have “a single point of attack, a single point of failure” [394]. The stance calls for sovereignty in models and in data, but not in chips, which are the layer where NVIDIA dominates [395]. Lionel Laurent called it “sovereignty-as-a-service” in Bloomberg Opinion [395]. A market of a few closed labs concentrates NVIDIA’s customers in companies that design their own chips, such as the large clouds (see “The chain around AI”). A market of open models, by contrast, multiplies the buyers of GPUs to run them, as SemiAnalysis and Ben Thompson, author of the Stratechery newsletter, argue [27] [396]. We believe that explains NVIDIA’s place better than an alliance with China.
What Huang foresees for compute
In March 2026, eight months before resigning (see “Jensen Huang’s fall”), Huang laid out his vision of compute at NVIDIA’s annual conference and on Lex Fridman’s podcast [397] [15]. The company’s product plan and what NVIDIA asks of its suppliers rest on that vision [397] [15].
For Huang, the computer changed its function. For decades it served to store and retrieve information, like a warehouse, and now it generates tokens that are sold, like a factory. “Warehouses don’t make much money,” he said, and factories do. The unit of compute went from the GPU to the computer, then to the cluster, which is a set of connected computers, and then to the AI factory, and the next one will be “planetary scale.” Demand grows because of four scaling laws: that of pretraining, the first stage, in which the model learns from enormous volumes of text; that of post-training, the later tuning; that of reasoning when answering, and that of agents. According to each one, more compute yields more capability: when training on more data, when fine-tuning with synthetic data, the data another model generates, when reasoning more in each answer and when adding agents. “Intelligence is gonna scale by one thing, and that’s compute” [15].
The product plan follows that idea at an annual pace. NVIDIA shipped the GB200 systems in 2024 and GB300 in 2025, and the Vera Rubin platform (see “NVIDIA orders the chain”) from the second half of 2026 [165]. Mixture-of-experts models are those that activate only part of themselves in each answer. Compared with Blackwell, Vera Rubin trains them with a quarter of the GPUs, and delivers up to ten times more inference performance per watt, at a tenth of the cost per token [164]. Next come Vera Rubin Ultra, with Kyber racks that double to 144 the GPUs connected as one, and the Feynman architecture, which NVIDIA presented without a date, with a new CPU, Rosa, and a next-generation inference processor, LP40. NVIDIA is also designing Space-1 Vera Rubin, for data centers in orbit [165] [397].
Huang explained that the plan gets ahead of the models because their architectures change every six months and those of hardware every three years, so NVIDIA has to foresee today the models of two or three years from now. Grace Blackwell was designed for mixture-of-experts models, and Vera Rubin, for agents [15].
For fast inference, NVIDIA turned to Groq. In December 2025 it licensed that company’s technology on a nonexclusive basis and took its founder, Jonathan Ross, and its president, Sunny Madra. Groq did not publish the price and still operates its cloud [398]. The rack that uses that technology, LPX, brings together 256 of its processors and, according to NVIDIA, unlocks “a new tier of ultra-premium” inference [164].
The price of the token and that of the computer move in opposite directions. “Our computer price is going up,” Huang said, but the cost of the token is “coming down an order of magnitude every year,” that is, to a tenth. Tokens are segmented “like iPhones,” into free, mid-range and premium, and the idea that somebody will pay “$1000 per million tokens is just around the corner” [15]. In prices, existing models get cheaper and frontier models are charged ever more (see “The price of the model”).
The market Huang foresees is much larger than the current one. At the conference he said he sees at least a trillion dollars, that is, a million million, in revenue between 2025 and 2027 [397]. With Fridman he said he was “absolutely certain” that the share of world output spent on compute will be “100 times more than the past,” and that NVIDIA can be a company with US$3 trillion in revenue “in the near future” [15]. His reason is that NVIDIA “is not in the market share business”: almost everything it sells did not exist before, and the burden of producing it is shared by 200 companies in its supply chain [15]. In its fiscal 2026, NVIDIA took in US$215.9 billion [11].
Energy, for Huang, is a limit, but not the only one. The power grid is designed for the worst case and operates most of the time at about 60% of its peak. Data centers could use that idle capacity if they accepted lowering their consumption when the grid asks. Today, instead, they demand “six nines” of availability, that is, 99.9999%, which is equivalent to about 32 seconds without service per year, and that is why they do not accept that the grid lower their consumption [15]. In 2025 the United States added 53 gigawatts (GW) of electric capacity, its largest increase since 2002, and China added 315 GW of solar power alone [399] [400]. A GW is roughly the capacity of a large nuclear reactor. China is building 41 nuclear reactors, and the United States, none [401]. The public affairs study covers the energy limit in depth [152].
Elon Musk, chief executive of SpaceX, which owns xAI, wrote on X on January 31, 2026 that China’s electricity generation “is still growing super fast”, with solar as the largest incremental contributor, and “will exceed America by a factor of 3X either this year or next”. He wrote it while quoting another account’s post, which gave China 33.2% of world generation in 2025 and the United States about 14.2% [402]. According to Ember, an energy think tank, China consumed 10,573 terawatt-hours (TWh) in 2025, a third of world demand and more than double the United States’ 4,536 TWh. A TWh is what a thousand large nuclear reactors produce in one hour. Chinese demand grew 5% and U.S. demand 3% [403]. At those rates, China would consume about 2.4 times as much as the United States in 2027, and to triple it, it would have to grow about 17% a year.
The vision rests on a buyer that loses money today, the lab. Ed Zitron warns that data-center demand depends on those labs paying their compute commitments [404]. Each link in the chain invests because it expects two things: that token use grows faster than its price falls in existing models, and that the frontier keeps being charged ever more (see “The price of the model”). The one who ultimately buys that capacity, through the cloud, is the lab, and the two frontier labs with known accounts lose money on their operations (see “Where the money stays”). NVIDIA already backs part of that credit, with the guarantee on the Ohio lease (see “NVIDIA between the two sides”).
The vision also explains NVIDIA’s bet on open, because more models and more users mean more tokens, as Nathan Lambert writes [405]. The product plan belongs to the company, but the agreements with suppliers were won by Huang in person, according to SemiAnalysis, which therefore argues that without him suppliers will not give NVIDIA the same priority (see “NVIDIA orders the chain”) [27].
10
China makes its own chips
China’s chip supply chain today has designers whose sales are doubling, a fab working almost at full capacity, memory that falls short and software that Huawei opened to attract developers. In April 2026, the month in which DeepSeek presented V4 adapted to Huawei’s chips, Jensen Huang, then chief executive of NVIDIA, warned: “The day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation” [38]. Eric Xu, Huawei’s rotating chairman, repeated in September 2025, as he had said at Huawei Connect 2024, that mainland China “will lag behind in semiconductor manufacturing process nodes for a relatively long time.” For that reason, he said, Huawei designs its systems with the manufacturing processes within its reach [406]. A node is a generation of the manufacturing process and is named with a measure in nanometers, which are millionths of a millimeter. The smaller the elements printed on the chip, the more transistors fit on it and the more it can do [169].
Why China stopped buying from NVIDIA
NVIDIA wrote in August 2026 that its exclusion from Chinese data centers “helped our competitors build larger developer and customer ecosystems to challenge us worldwide” [14]. Huang had said it in October 2025 with a figure: “we went from 95% market share to 0%” [407].
The exclusion came from both governments. In September 2025, according to the Financial Times, the Cyberspace Administration of China ordered ByteDance and Alibaba to cancel their orders of the RTX Pro 6000D, the chip NVIDIA had designed for China, and to stop testing it [136]. In November, according to Reuters, Chinese government guidance ordered that new publicly funded data centers use only domestic chips and that those less than 30% complete remove foreign chips [408]. That same month, according to the Financial Times, Gansu, Guizhou and Inner Mongolia cut by up to 50% the electricity bills of the largest data centers of ByteDance, Alibaba and Tencent, which had been spending more on electricity since they switched to domestic chips [409]. For its part, the United States has warned since May 2025 that using Huawei’s Ascend chips in any country “risks violating U.S. export controls.” The reason it gives is that those chips were likely designed or made with U.S. technology without the required license [410].
The exclusion continues. Since February 2026, the United States has given NVIDIA licenses to ship small quantities of the H200 to specific Chinese customers [11], and on May 14 it authorized about ten Chinese companies as buyers (see “Jensen Huang’s fall”). However, according to Reuters, in November about 600,000 of those chips were waiting in Hong Kong, because the Chinese government has not authorized importing them in volume and, according to The Information, cited by Reuters, asked companies in July to prioritize Chinese processors to run models [174]. Several large internet companies obtained about 10,000 each, and a server with eight H200s cost three times as much at the end of October as a year earlier [174].
According to IDC, cited by Reuters, 2,008,000 AI accelerators, the chips specialized in AI calculations, such as GPUs, were sold in China in the first half of 2026. NVIDIA shipped 678,000, 44% fewer than a year earlier, and fell to 34% of the market. Huawei shipped 572,000 and reached 28%. They were followed by Alibaba’s T-Head, with 7%; Cambricon, with 5%; Baidu’s Kunlunxin, with 4%, and Hygon, with 3% [37].
We take the IDC figure only as a signal of the trend. IDC does not say which NVIDIA models it counts or through which channel they arrived, and the figure does not match what NVIDIA itself reports about its sales in China. IDC also measures AI chips for data centers, so the gap with what Huang says depends on which chips and which channels each one counts; according to Reuters, the figure could include smuggled B200 and B300 chips, of the Blackwell generation, whose sale to China is banned. By contrast, Morgan Stanley estimates that Cambricon will have about 14% of the market in 2026, behind only Huawei [411], and Huawei said in September 2026 that it had already overtaken NVIDIA in China [412].
Who designs the chips
Huawei is the largest designer. According to the Financial Times, cited by Reuters, it sold US$7.5 billion in AI chips in 2025 and, based on the orders it already has, expects about US$12 billion in 2026 [36]. According to Reuters, Huawei plans to ship about 750,000 Ascend 950PR chips this year, at about 50,000 yuan per card, and 70,000 for the version with faster memory, and ByteDance and Alibaba plan to place orders. The same report says the 950PR barely improves on the compute of the Ascend 910C [413]. Huawei’s Atlas 350 card, presented in March, uses the 950PR, which is Huawei’s first chip with its own HBM memory [414]. HBM is the high-bandwidth memory that AI chips use (see “The chain around AI”). According to the company, that card delivers 2.8 times the compute of the H20, the chip NVIDIA sold in China, measured in FP4, a low-precision number format. With less precise numbers a chip performs many more operations per second, so the figure can only be compared with other measurements in the same format [414].
Cambricon multiplied its sales more than fivefold in 2025, to 6.5 billion yuan. That year it posted the first annual profit in its history, of 2.1 billion yuan, after losing 455.8 million in 2024 [415]. In the third quarter of 2026 it doubled its sales again, to 3.53 billion yuan (about US$529 million), but analysts expected 6.26 billion. According to Bloomberg, cited by The Next Web, the gap is because Cambricon cannot make all the chips it is asked for [411].
T-Head, Alibaba’s chip division, presented the Zhenwu 810E in January 2026, whose performance places it between the A800 and the H20, the two chips NVIDIA adapted to export controls to sell them in China. Alibaba is preparing a separate stock market listing for T-Head [416].
The stock market funds the smaller designers. Moore Threads, founded in 2020 by a former NVIDIA executive in China, rose 425% on its first day in Shanghai, in December 2025, after raising 8 billion yuan [417]. MetaX doubled its sales in 2025, to 1.6 billion yuan, with a loss of 830 million excluding extraordinary items [418]. Biren rose almost 76% in its Hong Kong debut, in January 2026, after raising US$717 million [419]. Enflame, which still has no profits, went public in Shanghai in September 2026 to raise 6.1 billion yuan. Its largest customer, Tencent, is also one of its main shareholders [420]. Kunlunxin, Baidu’s chip division, applied in January 2026 to list in Hong Kong [421].
That source of capital contracted in the fall. According to Reuters, the STAR50 index of the Shanghai stock exchange, which groups the leading AI and semiconductor companies, fell 40% in less than four months from its July high and lost some US$480 billion in value. CXMT, the Chinese memory maker, lost more than a quarter of its value in a month, and MetaX and Moore Threads were worth half what they were in July. Reuters found no obvious trigger. Traders attributed the fall to high prices, tax bills and the threat that the United States or other countries would ban Chinese equipment in data centers. Zhongji Innolight and Eoptolink, Chinese makers of optical components, were trading about 60% below their June and July highs, in part because the United States is moving toward barring them from supplying its data centers [422]. The U.S. Federal Communications Commission has sought since August to prohibit new models of Chinese optical transceivers, and in November it was pressing ahead despite lobbying by Amazon, Google, Meta and Microsoft, among others [423].
Even after the fall, the price of the index’s shares was equivalent to about 100 years of its companies’ earnings, down from nearly 250 at the peak, against 30 for the Nasdaq 100. Hong Hao, of Lotus Asset Management, said the fall will stop many companies that were seeking to go public, and Reuters expects it to slow fundraising by tech companies [422]. We believe the fall weighs more on manufacturing than on design, because making memory and chips requires long-term capital, and that a market that stops providing that capital leaves more weight to the Chinese state.
How many chips China makes
Huawei’s figures differ depending on who calculates them. In June 2025, the Bureau of Industry and Security of the U.S. Department of Commerce estimated before the U.S. House of Representatives that Huawei could produce 200,000 Ascend chips or fewer that year [424]. SemiAnalysis estimated in September 2025 that Huawei would ship 805,000 that year, 653,000 of them 910Cs. It also estimated that Huawei had received from TSMC, in violation of export controls, more than 2.9 million Ascend chips, a reserve that would run out in nine months [39]. Huawei has been on the U.S. Entity List (see “Jensen Huang’s fall”) since May 2019. Since May 2020, in addition, a foreign fab that uses U.S. equipment needs a license to produce chips designed by Huawei [425]. Epoch AI estimates that Huawei produced about 800,000 Ascend chips in 2025 and will produce about 1.5 million in 2026, against about 5.9 million for NVIDIA. It also estimates that, in compute, Huawei’s 2026 output will be less than 4% of NVIDIA’s [426].
The fab
SMIC, China’s largest contract chipmaker, makes China’s most advanced chips without EUV lithography machines, which the controls deny it (see “The chain around AI”). SMIC has been on the U.S. Entity List since December 18, 2020. Because of that listing, the United States as a general rule denies every license application to sell it what is useful only for making chips of 10 nanometers or less, which are the most advanced [427].
TechInsights, which analyzes chips already made, found SMIC’s N+3 process in the Kirin 9030, Huawei’s phone processor. That process is an evolution of its 7-nanometer class, “significantly less scaled” than the 5-nanometer processes of TSMC and Samsung, that is, with larger elements and, therefore, fewer transistors in the same space [428]. In March 2026, TechInsights concluded that this process “will not be a viable solution” for data-center AI chips. The reason is that manufacturing defects grow with the size of the chip, and an AI chip is much larger than a phone chip [40]. In the third quarter of 2026 SMIC sold US$3.19 billion, 33.9% more than a year earlier, with 97.3% of its capacity in use [429]. According to Nikkei, cited by TrendForce, China wants to go from fewer than 20,000 wafers a month in the 7- and 5-nanometer classes to 100,000 in one or two years, with SMIC, Hua Hong and fabs linked to Huawei [430] [431].
Hua Hong, China’s second contract chipmaker after SMIC, is the second option. According to Reuters, cited by TrendForce, its subsidiary Huali began developing a 7-nanometer process in 2025 with support from SiCarrier, a company backed by Huawei, and Huawei plans to shift part of the production of its AI chips to it. According to Reuters, which reported it on April 28, 2026, the U.S. Department of Commerce sent letters to Lam Research, Applied Materials and KLA, U.S. equipment makers, to bar them from selling equipment to two Hua Hong fabs, one of them Fab 6 in Shanghai, where Huali is developing its 7-nanometer process [432].
China is the largest buyer of chipmaking equipment. According to SEMI, it spent US$49.3 billion in 2025, out of a world total of US$135.1 billion [433]. Its lithography maker, SMEE, showed in November 2026 machines for processes of 110 nanometers or more, with more than 90% Chinese components. Those processes are far older than the 7-nanometer class in which SMIC makes China’s most advanced chips. According to Reuters, another Chinese company, Aishengna, began producing immersion machines, which print with deep ultraviolet light through a layer of water and are the generation before EUV, which still require testing [434]. The public affairs study describes the Chinese prototype EUV machine in Shenzhen, which does not yet make chips [146].
The United States controls that equipment. In December 2024 it placed under control 24 types of chipmaking equipment and three types of software, as well as HBM memory, and added 140 entities to its restricted list, the Entity List (see “Jensen Huang’s fall”) [435]. On April 22, 2026, the Committee on Foreign Affairs of the U.S. House of Representatives approved the MATCH bill, which calls for identifying the Chinese fabs of advanced chips and the equipment that the United States and its allies must deny them [436]. The bill is by Representative Michael Baumgartner and assigns that review to the U.S. Secretaries of Commerce and of State. Committee approval is the step before the vote on the House floor [436].
Part of the controls is suspended. The trade truce is the agreement that Donald Trump and Xi Jinping reached when they met in South Korea, and that the White House announced in November 2025. In it, the United States cut its tariffs on China by ten points, and China committed to buying U.S. soybeans and to closing its investigations into semiconductor companies [380]. With that truce, the United States suspended for one year the rule that extends its controls to companies in which those on its list own 50% or more, alone or together, and China suspended its expanded controls on rare earths. In November 2026 the two governments extended those suspensions until January 10, 2027 [437] [438].
The suspended U.S. rule is known as the affiliates rule. The Bureau of Industry and Security issued it in September 2025 for the Entity List and for the Military End-User List. Before it, an affiliate not named on those lists fell outside the restriction [439]. The rule seeks to prevent diversion, that is, a listed company receiving through an unnamed affiliate what it cannot buy directly [440]. The suspension has been in effect since November 10, 2025. While it lasts, only the companies the lists name are restricted [441]. The suspended Chinese controls date from October 9, 2025. They required a license from China’s Ministry of Commerce to export from any country certain products made abroad, when rare earths of Chinese origin were 0.1% or more of their value [442].
Memory
Memory is the main limit. AI chips use HBM memory, whose export to China the United States has controlled since December 2024 (see “The fab”). SemiAnalysis said it in September 2025: “We believe HBM production is the bottleneck” [39]. Epoch AI estimates that, even if Huawei smuggled in ten times the HBM that China produces, Huawei’s output in 2028 would be only 16% of NVIDIA’s [426].
CXMT, the Chinese memory maker, raised 58 billion yuan (US$8.71 billion) in July 2026, in the largest IPO in Asia that year. In the third quarter it sold 118.71 billion yuan, 613.2% more than a year earlier, and earned 54.4 billion. According to sources cited by Reuters, CXMT will begin limited production of HBM3E, one of the most recent generations of HBM memory, in two months, with manufacturing yields, the proportion of chips that come out usable, of 20% to 30%, against about 70% for SK hynix [443]. The public affairs study describes the smuggling of Korean HBM through Hong Kong [146].
Software
Huawei opens its software to sell its hardware. Xu put it this way: “Our monetization strategy for AI is focused on hardware.” Huawei committed to opening, before December 31, 2025, the code of CANN, its equivalent of CUDA, and that of its development tools [406]. CUDA is the platform used to program NVIDIA’s chips (see “NVIDIA orders the chain”).
Chinese labs tune their models to those chips. In August 2025, DeepSeek trained V3.1 with UE8M0 FP8, a format that stores each number with few digits so that the chip calculates faster and uses less memory. A model like that runs better on chips that handle that format, and according to the company the format “is designed for the next generation of domestically produced chips to be released soon” [444] [445]. Huawei said its chips were used in part of the training of DeepSeek V4 [38]. Zhipu trained an image model, GLM-Image, entirely on Ascend chips in January 2026 [446]. According to the Chinese press, DeepSeek will train V5 on them [447].
Scale instead of quality
Huawei makes up in quantity for what each chip lacks. Xu said that, without access to advanced manufacturing processes, its bet is to combine chips [406]. A CloudMatrix 384 system brings together 384 Ascend 910C chips, each with a third of the performance of a Blackwell. According to SemiAnalysis, that system delivers almost twice the compute of NVIDIA’s GB200 NVL72, a rack of 72 Blackwell GPUs. With more than five times as many chips, Huawei achieves twice the compute, but with 2.6 times fewer operations per watt, that is, with far more electricity spent for the same work [448]. China can pay for that extra electricity, because of the electric capacity it adds each year (see “What Huang foresees for compute”).
In the same speech, Xu put dates on the next chips. The Ascend 950PR was set for the first quarter of 2026, with its own HBM memory, cheaper than HBM3E. The 950DT was set for the fourth quarter of 2026; the 960, with twice the compute, for the fourth quarter of 2027, and the 970, for the fourth quarter of 2028. For the fourth quarter of 2026 he also announced the Atlas 950, a system of 8,192 950DT chips that, according to Huawei, will deliver 6.7 times the compute of the NVL144 that NVIDIA planned for the second half of the year [406].
At Huawei Connect 2026, on September 17, Huawei moved the Ascend 960DT up by three quarters, to the first quarter of 2027, and presented the Atlas 960E SuperPoD, with 4,096 chips, much smaller than the 15,488-chip Atlas 960 it had announced earlier. Its executives told Bloomberg that they already have a larger share in China than NVIDIA, but that they cannot produce all the chips they are asked for. Huawei raised the price of the 950DT by 60% because of the shortage of components [412]. Epoch AI estimates that the Ascend 950 performs at half the level of the 2022 H100, and that Huawei will remain three or four years behind NVIDIA in each chip at least until 2029, so it will hardly catch up by 2030. Epoch AI revised its report on September 24 with the new schedule [426].
NVIDIA’s chips keep arriving by other routes. According to TechCrunch, Will DePue, a former OpenAI researcher, found in the weights of Kimi K3 and DeepSeek V4 traces of training on Blackwell chips, whose sale to China is banned. According to that analysis, the weights retain traces of the configuration of the chips on which the model was trained. Those chips may have been in a data center outside China [449]. That is the route closed by the rule that the U.S. Department of Commerce made public on November 16. That rule requires data centers in any country to obtain a license to provide large amounts of compute to Chinese companies (see “Export controls”).
China’s bottleneck today is in making the chips: in memory and in the machines, which are the two links where the controls still weigh. Gregory Allen, of the Center for Strategic and International Studies (CSIS), a Washington think tank, acknowledges that Huawei designs its chips well, and SemiAnalysis pointed to HBM memory production as the bottleneck [450] [39]. Huawei opens its software so that its hardware gets used, as Eric Xu said [406]. It is the zero-price strategy, which consists of giving away one product to sell the one that goes with it, and which, according to Alicia García-Herrero and Bertin Martens, of Bruegel, a European economic think tank, NVIDIA follows with CUDA [451] [452], and which NVIDIA also applies to its models (see “Cheapen someone else’s layer”).
For now, the distance between Chinese and U.S. models is shorter than the distance between their chips. Stanford’s AI Index measures the first at 2.7%, the lead of the best U.S. model over the best Chinese one in the Arena ranking (see “China”) [317], and Epoch measures the second in years. Lennart Heim, of RAND, a U.S. nonprofit research organization [453], already noted in 2025 that Chinese models were closing the gap even though the United States has about ten times more compute [454]. We believe that weighing on that distance is the fact that some Chinese models were trained on NVIDIA chips outside China; the November 16 rule targets that route. And we believe that the closure imposed by the two governments hastens the arrival of top-tier Chinese models trained and served on Chinese chips. That is what Huang feared, because it takes away from NVIDIA in China the developer base that, according to Huang himself, is its advantage (see “NVIDIA orders the chain”).
11
Robots
In robots, the line between open and closed does not match the border between the United States and China. The models that move robots are published mostly in the United States, and the bodies are made mostly in China. Of the roughly 13,000 humanoids shipped in 2025, the three manufacturers that shipped the most are Chinese, according to the consultancy Omdia [455], and in the first half of 2026 Chinese manufacturers accounted for more than 97% of shipments, according to Smart Analytics Global [41]. Jensen Huang, then chief executive of NVIDIA, acknowledged it in March 2026: he said Chinese microelectronics, motors, rare earths and magnets are “the best in the world” and that the U.S. robotics industry depends on that chain [456].
Open and closed in robots
The most widely used robot models are vision-language-action models: they take in what the camera sees and an instruction in words, and return the robot’s movements. On the closed side are Google DeepMind and Figure. Google distributes Gemini Robotics On-Device 2, its model that runs inside the robot, only among trusted testers [457]. Figure calls Helix 02, the model its Figure 03 robot uses at BMW’s plant in Spartanburg, South Carolina, “proprietary” [458].
NVIDIA opens its robot models, some under licenses it writes itself and others under a third-party license. In April 2026 it presented GR00T N1.7, a 3-billion-parameter model, as open and licensed for commercial use; the two previous versions, from 2025, were licensed only for noncommercial use [459] [460]. Its own technical page warns that the license file accompanying the N1.7 weights is still the noncommercial one [459]. Its world model, which simulates the physical world to train robots without touching a real one, is called Cosmos 3 and came out in May under OpenMDW, a Linux Foundation license that covers the weights, the data and the code [461]. On November 9, NVIDIA announced GR00T 2 as open, without yet publishing its license or its release date [462].
Physical Intelligence, a San Francisco company, publishes its earlier models under the Apache 2.0 license, which allows commercial use, and gives access to its most recent models through its own interface. Its openpi repository contains the π₀, π₀-FAST and π₀.₅ models, trained on more than 10,000 hours of robot data [463]. Since October it has offered that interface as a trial, with no published price: the customer uploads its data, the company fine-tunes the model on its own compute and serves it back [464]. Hugging Face, which NVIDIA agreed to buy, publishes SmolVLA, with 450 million parameters, also under Apache 2.0 [465].
Chinese labs publish weights, but not always for commercial use. In January 2026 Unitree published the weights of UnifoLM-VLA-0, built on Alibaba’s Qwen2.5-VL, under a license that prohibits commercial use [466]. The Beijing Academy of Artificial Intelligence published RoboBrain 2.5, built on Qwen3-VL, under the Apache 2.0 license [467]. We believe robots repeat what happens with language models, but with the roles reversed: the one that opens the most is a U.S. company that sells the chip, and the one that makes the body opens the model without letting anyone else sell it.
The body is made in China
In 2025, AgiBot shipped 5,168 humanoids, Unitree 4,200 and UBTech about 1,000, according to Omdia [455]. Unitree says in its prospectus that it sold more than 5,500 [468]. For the first half of 2026, IDC estimates about 25,000 humanoids shipped, 432% more than a year earlier, with AgiBot in the lead with more than 8,600, and attributes more than 19,000 to China, about 77.9% [469]. The consultancies do not agree. Smart Analytics Global counts 19,100 in total, compared with about 5,100 a year earlier, and attributes more than 97% to Chinese manufacturers [41].
Unitree makes money from its robots. According to its prospectus, in 2025 it took in about 1.7 billion yuan, 335% more than in 2024, and humanoids contributed about 51% of its sales. The gross margin, which is what remains of each yuan of sales after paying what it costs to make the product, was about 60% for the company, and the humanoid margin fell from 87.67% in 2023 to 63.18% in 2025 [468] [470]. It went public on August 19, 2026 on Shanghai’s STAR Market, the Chinese market for technology companies. It raised about US$900 million, and its shares closed their first day up more than 460% [471]. In the first nine months of 2026, according to its first report as a listed company, its net income rose 270%, mainly because a year earlier it had booked a one-time expense for share-based payments. Recurring net income, which excludes those items, fell 19%, while its research spending nearly tripled [472]. UBTech, which is listed in Hong Kong, took in 2.001 billion yuan in 2025, 820.6 million of it from humanoids, and lost 789.8 million [473].
In the United States, Agility Robotics began trading on the Nasdaq on November 5, 2026, after merging with a company created to buy another and take it public, at a valuation of US$2.5 billion before the transaction [474]. Venture capital in robotics companies had already reached US$18.8 billion by June 2026, more than in all of 2025, according to Crunchbase [475]. Skild AI, which develops a general model for robots, raised about US$1.4 billion in January in a round with SoftBank and NVIDIA, at a valuation of more than US$14 billion [476].
We believe the scarce link in robots is the components of the body, its motors, its reducers (the gears that lower the motor’s speed and raise its force) and its magnets, and that this is why the winner today is whoever makes them at scale, just as in AI the winner is whoever makes the memory (see “Where the money stays”). The model is published and copied, the body has to be manufactured, and China decides when it exports the rare earths from which the magnets are made, as it showed when it suspended its expanded controls in November 2025 and extended the suspension until January 10, 2027 (see “The fab”) [438].
The United States closes the door
The United States has shut new models of foreign-made mobile robots out of its market. On July 28, 2026, the Federal Communications Commission added “foreign-produced advanced robotic devices” to its Covered List, except those with a Conditional Approval from the U.S. Department of War [42]. Equipment on that list cannot receive new authorizations from the Commission, and without that authorization a device that emits radio waves, such as one that connects over Wi-Fi, cannot be sold in the United States. The rule reaches any country, not only China, because it identifies equipment by where it was produced. It covers mobile robots, such as humanoids and four-legged robots, that weigh more than two kilograms and connect to a network, and it leaves out fixed industrial and medical arms. Models already authorized may still be sold, although their hardware changes are restricted [477]. The national security determination on which it rests says robots collect data that could be used to surveil Americans “or to remotely commandeer the robots” [42].
The U.S. Department of Defense, which since September 2025 also uses the name Department of War, added Unitree on June 8, 2026 to its list of Chinese military companies operating in the United States [478]. On July 22 the House of Representatives passed its defense bill with a provision requiring the Department of War to begin determining whether Chinese robots should be banned nationwide [479]. In 2025, security researchers documented that Unitree’s G1 humanoid sends data to external servers every five minutes and that all units share the same encryption keys, so whoever obtains the key of one robot can decrypt the communications of all of them [480].
China, on the other side, funds the sector as a state priority. Its 15th Five-Year Plan, for 2026 to 2030, includes embodied intelligence, which is AI that acts in the physical world, among the industries of the future that should become “new points of economic growth” [481]. The National Artificial Intelligence Industry Investment Fund, part of the state chip fund, made its first investment in embodied intelligence in March 2026: it heads the list of investors in a 2.5 billion yuan round for Galbot [482]. The National Development and Reform Commission itself, China’s economic planning body, warned in November 2025 of the risk of a bubble, with more than 150 humanoid makers in the country [483].
NVIDIA in the middle again
NVIDIA sells the brain of robots to both sides. Its robot computer, Jetson AGX Thor, came out in August 2025. NVIDIA named Agility Robotics, Amazon Robotics, Boston Dynamics and Figure among its first users [484], and Unitree, AgiBot, Galbot and UBTech also adopted it early [485]. On May 31, 2026 it presented a reference humanoid for universities: the body is a Unitree H2 Plus, the brain a Jetson Thor and the software NVIDIA’s. Unitree will sell it at the end of 2026, and it will be used by the Allen Institute, ETH Zurich, Stanford University and the University of California San Diego [486].
We believe that humanoid sums up NVIDIA’s position on both sides (see “NVIDIA between the two sides”). It puts a U.S. model that it calls open on the body of a company the Department of War lists as a Chinese military company, two months before the Federal Communications Commission closed the market to new models of mobile robots made outside the United States.
The law of the machine
Europe regulates robot AI within the rules for machinery. The Digital Omnibus, the July 2026 reform of the European Union’s digital laws, moved the Machinery Regulation to the section of the EU AI Act that applies through the sectoral route, that is, whose requirements do not apply directly but are incorporated into the law for each product. It tasked the European Commission with adding the high-risk AI requirements to the Machinery Regulation through delegated acts, to apply from August 2, 2028 [487] [488]. A delegated act is a rule the European Commission issues under a mandate from the regulation to complete it or change its annexes. The Machinery Regulation applies from January 20, 2027 and already covers safety components whose behavior changes through learning [489]. Damage caused by a defective robot falls under the Product Liability Directive (see “Liability for harm”) [490].
Technical standards lag behind. ISO 10218, the safety standard for industrial robots, was updated in 2025 and replaced the 2011 edition, but it still does not apply to robots that are not industrial [491]. The standard for industrial mobile robots that must keep their balance on their own, such as legged humanoids, is still a draft, ISO 25785-1 [492]. In China, the Ministry of Industry and Information Technology presented in February 2026 a standards system for humanoids, a map of the standards to be written for the whole chain [493].
12
Quantum computing
Quantum computing intersects with AI at four points: AI already helps operate quantum machines; quantum machines learn certain problems better than an ordinary computer, for now in theory; NVIDIA sells the bridge between the two; and a large quantum machine could break public-key cryptography, which today protects the connections and digital signatures of models, of their data and of almost everything else.
A quantum computer calculates with qubits, units that, unlike a bit, can be in a combination of 0 and 1 and become linked to one another; with algorithms designed to exploit this, quantum computers solve certain problems, such as factoring large numbers or simulating molecules, much faster than the best ordinary computer. Qubits fail easily, because heat, vibrations or any interference from the environment alter their state. The technique at stake in these years is error correction. A physical qubit is each real component of the chip. A logical qubit is one the machine builds by spreading the same information across many physical qubits, so that if one fails, the others make it possible to detect the error and correct it. That way the logical qubit fails less than each of the physical qubits that form it, but each one costs many of them. No machine yet does useful work that an ordinary computer cannot do [494]. A fault-tolerant machine is one that corrects its errors while it calculates, so it can run long programs. IBM dates its first fault-tolerant machine, Starling, with 200 logical qubits, to 2029 [495]. The experiments of Microsoft and Quantinuum in June 2026 reached 12 (see “What quantum already does”).
What quantum already does
The United States and China already correct errors below the threshold. Each physical qubit added to a logical qubit brings its own errors, and if the qubits fail too often, adding more makes the result worse. Below the threshold they fail so rarely that correction removes more errors than they add, and each qubit added makes the logical qubit more reliable. That is the condition for a machine to be able to grow. Google showed it with its Willow chip, with 105 qubits [43], and the University of Science and Technology of China repeated it with Zuchongzhi 3.2, in a paper from December 2025 [496]. In June 2026, Microsoft and Quantinuum published experiments with up to 12 logical qubits that failed between 11 and 800 times less often than without correction. Correction is repeated in rounds while the machine calculates, and a long program needs many rounds in a row; in one of those experiments, the error per round was more than 50 times lower than without correction. These are the first steps toward a large-scale fault-tolerant machine [497].
Quantum advantage means performing a calculation that no ordinary computer can do in a reasonable time, whether useful or not, and every announcement stays on trial while others try to reproduce it with ordinary computers. In May 2026, the same Chinese team published Jiuzhang 4.0 in Nature, a machine that handles up to 3,050 photons, the particles of light [498] [499]. The team estimates that the best supercomputer would take more than 10^42 years to repeat what it did, a figure that far exceeds the age of the universe, about 13.8 billion years, and Jiuzhang 4.0 is built for that problem, sampling particles of light, not for any calculation. The machine sends the photons through a network of mirrors and records where each one comes out, and the task is to produce those results with the correct probabilities. An ordinary computer cannot imitate it in a reasonable time, but the problem has no demonstrated practical use yet [500]. IBM declared 2026 its year of advantage and publishes its experiments in an open tracker run by the community, so others can verify them [494].
Since October 2025 there have been announcements of quantum advantage that others can check. Google published an experiment in Nature, Quantum Echoes, in which the Willow chip lets a quantum system change over time, disturbs it at one point, makes it run back toward its initial state and measures how much of the disturbance remains, which is the echo [44]. With 65 of its 105 qubits it took about two hours, and Google estimates that Frontier, one of the fastest supercomputers in the world, would take 13,000 times longer [501]. The result is a number, and another quantum machine of equal quality could repeat it to confirm it. No one has done so yet, and the problem is a test model with no practical use [502].
On July 30, 2026, IBM and the University of Chicago announced a calculation that comes with a statistical guarantee that its result is correct at least to a certain degree, the proof that was missing from earlier announcements [503]. To achieve it they used an error-detecting code: the machine discards the runs in which a fault appears, and so each operation fails about 10 times less often [504]. The code detects errors but does not correct them, and the result is waiting for other groups to try to reproduce it with ordinary computers [505]. We believe the discussion has moved from whether advantage exists to how to verify it, and that the next step is a verified advantage on a problem with practical use.
Microsoft is betting on another type of qubit, the topological one. It stores the information spread between the two ends of a nanowire, a strand of material hundreds of times thinner than a hair, so that noise from the environment cannot easily erase it. Microsoft argues that this way it would need far fewer qubits to correct errors. That protection depends on states of matter called Majorana modes. A mode, in physics, is a way in which the electrons of a material move together, as if they were a single particle. Majorana modes appear in pairs, one at each end of the nanowire, and between them they hold the spread-out information. Whether Microsoft’s chip has them is an open debate among physicists. When the company presented Majorana 1, in February 2025, the editors of Nature, the journal that published the paper, added a note: the results do not prove that those states exist in the chip [506] [507]. In June 2026, Henry Legg, a physicist at the University of St Andrews, published a comment in the same journal arguing that Microsoft’s data point to a state without the protection it was seeking. Microsoft replied, also in Nature, that its measurements indicate a topological origin and rule out almost all other explanations [508] [509].
In June 2026, Microsoft presented Majorana 2, which replaces the chip’s aluminum with lead and adds antimony to the semiconductor [510] [511]. What matters most about a qubit is how long it keeps its state, because each time it loses it an error occurs. According to Microsoft, if a qubit performs one operation every microsecond and its state lasts one millisecond, it makes an error every thousand operations; if it lasts tens of seconds, it can perform tens of millions of operations before making a mistake [511]. According to the preprint that accompanied the announcement, a paper published before other scientists review it, in Majorana 2 that state lasts about 20 seconds, and sometimes a minute [512]. Microsoft says its qubits are 1,000 times more reliable than those of Majorana 1 and moved up to 2029 its goal of a quantum computer that solves problems no computer can solve today, half the time it had given itself [510] [511].
The preprint has not yet been peer reviewed, and physicists outside the company question it [513]. Henry Legg says there is no evidence that the device can hold a superposition, the combination of 0 and 1 that distinguishes a qubit from a bit, and that the 20 seconds therefore measure how long a bit lasts [514]. He adds that the data seem to come from a few measurements on a single device [515]. Sergey Frolov, of the University of Pittsburgh, says the work does not rest on a solid track record [515]. Frolov and Vincent Mourik pointed out the inconsistencies that led a team linked to Microsoft to retract in 2021 a 2018 Nature paper that claimed to have detected Majorana modes [516]. Microsoft replied that it stands by its results and its roadmap [514]. Jason Zander, the executive vice president in charge of its quantum team, said that Bell Labs did not have to prove the electron existed to invent the transistor, and that Microsoft does have to prove that Majorana modes exist [515].
With neutral atoms, thousands of qubits can already be handled. Each qubit is an atom with no electric charge held in place by a laser beam. In September 2025, the California Institute of Technology trapped 6,100 atoms, which kept their quantum state for about 13 seconds, and operated each one with two faults per 10,000 operations. It has not yet linked them to one another, and without linking them it cannot calculate [517]. In November, Harvard University, the Massachusetts Institute of Technology and the company QuEra published in Nature, with up to 448 atoms, the components of a fault-tolerant machine, among them error correction below the threshold. By protecting each logical qubit with more atoms, the error per round of correction fell to less than half. The authors estimate that, for a large machine, they still need to reduce the errors of each atom by three to five times [518].
Open in software, closed in hardware
In quantum computing the open part is the software, that is, the tools with which programs for quantum machines are written and tested on an ordinary computer. IBM’s Qiskit, Google’s Cirq, NVIDIA’s CUDA-Q and PennyLane are published under the Apache 2.0 license; Microsoft’s Q# language, under the MIT license, and QPanda from Origin Quantum, the Chinese company, also under Apache 2.0 [519]. What is closed is the machine. Only a few companies and labs have one, and everyone else gets in through the cloud on the owner’s terms. IBM’s free plan, for example, gives up to 10 minutes of use every 28 days [520].
What is closed most carefully is what serves for attack. In March 2026, Google estimated how many qubits it would take to break elliptic-curve cryptography, which protects, among other things, a large share of digital signatures on the internet. It did not publish the attack circuits, the quantum programs that would carry it out. Instead it published a zero-knowledge proof, which shows the circuit exists without revealing it [521]. In April, the security firm Trail of Bits forged that proof, that is, it got the verifier to accept a proof of something that was not true, by exploiting flaws in the code Google used to generate it. While it had those flaws, the proof guaranteed nothing. Google fixed them without its results changing [522], and in June a researcher at the French institute Inria reconstructed the circuits and published them [523]. We believe the case repeats what this study describes about a model’s weights: what is withheld so as not to help the attacker still reaches those with the means to rebuild it, and the ones left out are those who lack them.
AI inside the quantum machine
AI already helps calibrate and correct quantum machines. Google DeepMind and Google Quantum AI presented AlphaQubit 2, a decoder, the program that reads error signals and decides how to correct them. It works in real time, at less than one microsecond per correction cycle, and that speed matters because the machine repeats those cycles nonstop while it calculates: a decoder slower than the machine lets errors pile up [524]. In July 2026, Google published in Nature a reinforcement learning method, in which the program learns by trial and error, that adjusts more than 1,000 machine parameters while it calculates. Here the parameters are the physical settings that control each qubit, distinct from a model’s weights. The method reduces by about 20% how often logical qubits fail, compared with conventional calibration [525] [526]. Microsoft says it built Majorana 2 with the help of AI agents from its Microsoft Discovery platform, which simulated materials, automated measurements and found a miscalibrated temperature sensor that was distorting the fabrication data [510].
NVIDIA holds the bridge. In October 2025 it presented NVQLink, a connection between its GPUs and quantum processors, with 17 makers of quantum machines and nine U.S. labs [527]. In April 2026 it presented Ising, a family of models it calls open for calibrating and decoding quantum machines [528]. Its calibration model is built on Alibaba’s Qwen3.5, under NVIDIA’s model license, and the July version on Google’s Gemma 4, under the OpenMDW license [529]. The models can be downloaded, but the decoder needs NVQLink, NVIDIA’s connection, and calibration runs on CUDA-Q, NVIDIA’s platform for programming machines that combine quantum processors and GPUs: whoever uses the models stays inside NVIDIA’s equipment [530]. In September 2025 NVIDIA invested in Quantinuum, in a US$600 million round, and in PsiQuantum, in a US$1 billion round [531] [532].
Jensen Huang, then chief executive of NVIDIA, said in January 2025 that very useful quantum computers would arrive in 15 years at the earliest, and the next day the shares of two U.S. companies that make quantum computers fell: Rigetti’s by 40% and IonQ’s by 37% [533]. In March he retracted it at an event devoted to quantum computing within NVIDIA’s annual conference [534]. We believe NVIDIA repeats in quantum computing what it does with language models (see “What Huang says”): it gives away the software and the models, even though one of those models is built on a Chinese model, to sell the GPU and the connection needed by machines that combine quantum processors and GPUs.
Quantum in the service of AI
That quantum computing improves AI has been shown in theory and on problems made to measure. In 2021, IBM researchers proved that an error-corrected quantum computer classifies certain data that no classical machine can classify better than chance, provided that the discrete logarithm, a mathematical problem on which part of cryptography rests, is as hard as believed. Those data, however, were built from that same problem, so the result proves that the advantage can exist, but not that it appears with real data, such as texts, images or customer records [46]. In 2022, Hsin-Yuan Huang, John Preskill and the Google team showed that, when what is being learned is a quantum system, a quantum machine needs exponentially fewer experiments, and they confirmed it with 40 qubits [535]. A 2024 paper proposes that an error-corrected machine could speed up the training of large models, but only in a narrow case: that the model be sparse, that is, that most of its weights be zero, and dissipative, that is, that its training advance stably, without small changes being amplified. It does not prove it on any machine [536].
The broadest result is from April 2026. The group of Hsin-Yuan Huang and John Preskill, with Google researchers, estimates that a machine with fewer than 60 logical qubits would classify movie reviews or genetic data with between 10,000 and a million times less memory than a classical computer. That machine does not exist yet: the 2026 experiments reached 12 logical qubits, and IBM promises 200 by 2029. It is a preprint, not peer reviewed and only simulated, and the advantage is in memory: with enough memory, its authors acknowledge, classical algorithms take the same time and can be faster [537]. None of these works speeds up the training of a language model. Torsten Hoefler, of ETH Zurich and Microsoft, and Thomas Häner and Matthias Troyer, then at Microsoft, concluded in 2023 that quantum computing will serve problems with a lot of compute and little data, and not those with a lot of data, such as training, because loading the data into the machine costs more than is gained [538].
Experiments on real machines do not yet beat the best classical methods on AI tasks, and the first announcements have been qualified. In September 2025, HSBC and IBM announced an improvement of up to 34% in predicting whether a customer would accept a bond quote, but the improvement disappeared when the same machine without errors was simulated on an ordinary computer. The paper itself attributes it to the hardware’s noise, that is, to its imperfections: the improvement came from the machine’s faults and not from what makes it quantum [539] [540]. Scott Aaronson, of the University of Texas at Austin, wrote that the result has nothing to do with quantum speedup [541]. In May 2026 IonQ reported that its fine-tuning of a language model on its machine reduces error by about 24%, but the comparison is against simple classical models and was made with 250 sentences, where the difference falls within the margin of error the paper itself calculates [542].
In 2025, Google trained generative models on 68 qubits, the kind that produce new data similar to what they learned, and those models learn distributions, probability patterns that a classical computer cannot reproduce. But they are patterns of quantum systems, not of text or images, and the work is not peer reviewed [543]. A study by Xanadu, a Canadian quantum computing company, with 12 quantum models and 160 datasets found that common classical classifiers win [47]. Ewin Tang showed in 2018 that a quantum algorithm for recommendation systems, those that suggest products or movies, which was considered one of the best candidates for an exponential speedup, had a classical equivalent [544].
What is in dispute is whether quantum computing will produce data that AI cannot obtain any other way. Hartmut Neven, founder of Google Quantum AI, wrote in December 2024 that quantum computing will be indispensable for collecting training data that a classical machine cannot produce [43]. On November 10, 2026, Quantinuum and the Japanese institute RIKEN said that a model trained with data from their Reimei machine predicts magnetism better than one trained only on classical simulations, a result they themselves call preliminary and do not accompany with any paper [545]. NVIDIA presents the relationship the other way around: when launching Ising, Jensen Huang spoke of AI as what will make quantum computing practical [546]. We believe quantum computing will not soon speed up the training of the models at the center of this study’s dispute, and that its most likely contribution to AI will be chemistry and materials data that today cannot be measured or simulated.
The United States and China
The United States has controlled the export of quantum computers since September 2024. The Bureau of Industry and Security rule requires a license according to the size and precision of the machine: the fewer errors it makes, the fewer qubits it takes for it to be controlled. The most precise are controlled from 34 qubits, and from 2,000 physical qubits any machine is controlled, whatever errors it makes. The rule also controls their software and technology [547]. The rule controls the machine but says nothing about cloud access, and the November 16, 2026 rule on remote access is limited to advanced chips (see “Export controls”) [22]. That gap is what the Remote Access Security Act seeks to close; the House of Representatives passed it 369 to 22 on January 12, 2026, and it is not yet law (see “Restrictions on the rival’s model”) [548] [549]. Since May 2024 the Entity List has included the University of Science and Technology of China, which builds Jiuzhang and Zuchongzhi, and Origin Quantum, among 22 Chinese entities the list added for supporting China’s quantum development [550].
In 2026 the U.S. government moved from funding research to buying shares. On May 21, the U.S. Department of Commerce signed letters of intent, preliminary agreements that do not yet oblige it to hand over the money, with nine companies for US$2.013 billion from the CHIPS and Science Act of 2022, which funds chip manufacturing and research in the United States. In exchange, the government would receive in each one a minority, noncontrolling stake, that is, shares that do not give it control of the company. IBM would receive US$1 billion; GlobalFoundries, US$375 million; Atom Computing, D-Wave, Infleqtion, PsiQuantum, Quantinuum and Rigetti, up to US$100 million each; and Diraq, up to US$38 million [551]. A month later, Executive Order 14413 ordered the delivery to a U.S. Department of Energy laboratory of a quantum computer at a scale sufficient to begin the era of quantum-enabled scientific discovery [552]. According to the Payne Institute at the Colorado School of Mines, shares of the listed companies with final support fell an average of 22.1% between September 9 and November 6, compared with 6.4% for the Nasdaq, the U.S. stock index on which many of the large technology companies trade [553].
China put quantum technology first on the list of industries of the future in its 15th Five-Year Plan, for 2026 to 2030 [554]. Its spending is hard to measure. The German think tank MERICS says China’s flagship projects are “secretive and expensive”, puts Chinese public spending at about US$15 billion and could not verify the US$150 billion reported by McKinsey [555]. China also opened its machines to the cloud: in November 2025 China Telecom announced that its Tianyan-287, with 105 qubits, was open through its cloud to researchers and companies worldwide [556].
Harvest now, decrypt later
The quantum threat to cybersecurity begins before the machine exists. An attacker can store encrypted data today and decrypt it once a large quantum computer is available. The U.S. National Institute of Standards and Technology puts it this way in its transition plan, published in October 2026, which declares obsolete after 2030 the vulnerable public-key cryptography, the kind that allows parties to agree on keys and sign without having met before, such as RSA and elliptic-curve cryptography, and prohibits it after 2035 [557]. Post-quantum cryptography is the kind that resists a quantum computer, and that institute published its first three standards in August 2024 [558]. Executive Order 14412, of June 2026, orders agencies to use it in their most sensitive systems to agree on keys by December 31, 2030 and for digital signatures by December 31, 2031 [45]. The same institute warns that going from a standard to its full use has taken between 10 and 20 years [557].
Estimates of that machine keep falling. In May 2025, Google calculated that breaking a 2,048-bit RSA key, the most commonly used size today, would take about a week with a million physical qubits, twenty times fewer qubits than its 2019 estimate [559]. In March 2026, researchers from the California Institute of Technology and the company Oratomic, among them John Preskill and Hsin-Yuan Huang, estimated that with neutral atoms 10,000 qubits would be enough to run Shor’s algorithm, the one used to break this cryptography, and that with 26,000 an elliptic-curve key would be broken in a few days. It is a preprint, not peer reviewed [560]. Willow, the Google chip cited above, has 105 qubits: about 100 times fewer than the lowest estimate, the 10,000 neutral-atom qubits, and about 10,000 times fewer than the million Google calculated. Google set itself 2029 to migrate its systems, products and services [561]. We believe this is where quantum computing comes closest to the dispute in this study: the weights of a closed model, its customers’ data and the signatures that prove a downloaded model is the original travel or are authenticated with the same public-key cryptography that the U.S. government has already ordered its agencies to replace.
Where the money is
Quantum companies still spend more than they take in. IonQ took in US$173.8 million in the third quarter of 2026, but US$83.5 million came from SkyWater, a foundry that makes chips to order and that IonQ bought. It had an operating loss of US$433.1 million, although it booked net income of US$564 million from the accounting revaluation of its warrants, securities that give the right to buy IonQ shares at a fixed price. The company records them as a liability; when its stock falls, that liability is worth less and the difference is booked as a gain, without any cash coming in [562]. Quantinuum, which went public in June, took in US$9.1 million in the same quarter and lost US$118 million [563]. In the year ended November 10, 2026, IonQ shares fell 41%, Rigetti’s 64% and D-Wave’s 65% [564]. We believe that, for now, the winners in quantum computing are those who sell what surrounds it: NVIDIA, with the GPUs used to simulate, calibrate and correct those machines, and the governments that buy shares are counting on the business arriving later.
13
Two ideas about how to develop AI
Alongside the competition between the United States and China runs a dispute of ideas about how AI should be developed. One idea wants it concentrated in a few labs that control the most capable models. The other wants it diffused, so that anyone can use it, study it and modify it. Arvind Narayanan and Sayash Kapoor, of Princeton University, describe it as a choice between limiting AI to a few actors and diffusing it, and attribute it to deep differences in worldview [48]. That dispute does not follow borders. The same arguments are heard in California, in Beijing and in Paris, and within each country there are fractures as deep as the one that separates the two powers.
Concentration or diffusion. Of the arguments on the open side, the most repeated is that opening models prevents power from concentrating. It is made above all by those who earn their money outside the model, like Meta with advertising, or by latecomers who need to displace the incumbents, like Mistral and DeepSeek. Mark Zuckerberg, founder and chief executive of Meta, wrote in 2024 that open source prevents power from concentrating in the hands of a few companies [297]. In 2026 he wrote that the notion that AI is so dangerous that the only safe path is an extreme concentration of power “seems inherently problematic” [301]. Yann LeCun, then at Meta, sees the concentration of power through proprietary systems as “a much bigger danger than everything else”. He proposes for AI the answer liberal democracies gave to the press: that it be free and diverse [565].
Liang Wenfeng, founder of DeepSeek, wanted in 2023 even a small application to be able to use large models at low cost, instead of the technology staying in the hands of a few people and companies, “forming a monopoly” [566]. Arthur Mensch told the French National Assembly that he founded Mistral because “an oligopoly was taking shape”, and that opening models is “a technique of decentralization and of displacing the established players” [567]. Dario Amodei, of Anthropic, which sells access to a closed model (see “Who gains from opening and who from closing”), also fears the concentration of power, but sees it in other places. He sees it in chips, which should not be sold to the Chinese Communist Party, and in governments, even democratic ones, that could turn AI against their own citizens [568].
Control or scrutiny. The control argument is made by those who sell access to the model. The scrutiny argument is made by those who gain when the model circulates, like Zuckerberg, whose business is advertising, and Huang, who sells the chips. The closed side argues that it is very hard to put safeguards on an open model or to monitor its use, and that once published it cannot be withdrawn [291]. Amodei is thinking above all of cyber and biological attacks, and the 2026 tests of Chinese open models, which found weak safeguards in DeepSeek V4 Pro and Kimi K3, support him (see “What open and closed mean”). The open side replies that what is published gets reviewed more. Zuckerberg argued that open source should be significantly safer because anyone can scrutinize it [297]. Huang said it in Paris in 2025, according to Fortune: “If you want things to be done safely and responsibly, you do it in the open” [569].
The two sides increasingly agree on regulating by the model’s capability and not by its license. Amodei calls for mandatory testing of every sufficiently capable model, open or closed [291]. Demis Hassabis, of Google DeepMind, proposed in July 2026 a standards body in the United States, self-regulated and under federal oversight, along the lines of the one that oversees stockbrokers. That body would apply to frontier models from any country, “open or closed”, and would have open-source representatives on its board [570]. The model Hassabis cites is FINRA, a private nonprofit organization that writes and enforces the rules for stockbrokers under the oversight of the Securities and Exchange Commission [570] [571].
Principle or instrument. For some, opening is a value, and for others, a tool of national power. The use of openness as an instrument follows each country’s position. That is how the Indian think tank Takshashila Institution reads it for China: China opens its models to attack the revenue of U.S. labs, which charge for access to theirs (see “Cheapen someone else’s layer”). To India, the same think tank recommends not tying itself to either of the two powers [50]. The United States, which leads in closed models and in chips, sponsors both sides, open and closed. China, whose access to the most advanced chips is restricted, opens up to win adoption abroad. Middle-sized countries open up so as not to depend on either of the two (see “The state sponsors”).
In 2024 Liang called openness “more of a cultural act than a commercial one”, and said that in the face of a disruptive technology the moat that protects closed source, the advantage that keeps rivals from catching up, is temporary [572]. The U.S. AI Action Plan wants open models “founded on American values” because they have “geostrategic value” [30]. Marc Andreessen wrote in 2023 that open source should be allowed to proliferate with no regulatory barriers at all and, in the same text, that American and Western AI should be driven to “absolute global dominance, including ultimately inside China itself” [573]. Zuckerberg wants U.S. open models to be the best in the world [301]. China opens “non-sensitive” technical resources outward [353] and requires at home that generative AI services adhere to the core socialist values [574]. That is required by the Interim Measures for the Management of Generative AI Services, in force since August 2023. Those measures apply to those who offer such services to the public within China, and not to those who only research or develop the technology [574].
Common ground among rivals. In less than a month, between July and August 2026, three voices from rival countries came out against restricting open models. On July 17, at the World AI Conference in Shanghai, Xi Jinping called for encouraging open source and for opposing “overgeneralizing the concept of national security” in AI [575]. The July 24 letter signed by Meta, Microsoft, NVIDIA and other companies in favor of open weights (see “The open camp in the United States”) asked to avoid “premature restrictions” on open models [306]. And on August 10 Zuckerberg wrote that restricting access to foreign open models is not an effective solution [301].
Fractures within each country. In the United States, Anthropic was left out of the letter signed by Meta, NVIDIA, Microsoft, Google and OpenAI (see “Who is on each side”). In China, Robin Li, of Baidu, said in 2024 that open models would fall further and further behind [576], while DeepSeek and Moonshot, the company behind Kimi, made openness their route to the world. Meta changed its stance twice in two years. In April 2026 it published its Muse Spark model as closed, and in August it went back to publishing open models (see “The open camp in the United States”).
We believe this dispute is mixed up with business and with the rivalry between powers, but it cannot be reduced to them. Nearly every stance matches the place where whoever holds it makes money (see the map in “Who gains from opening and who from closing”), and governments use openness as an instrument. Even so, the arguments about the concentration of power, pluralism and safety are repeated almost word for word among rivals competing for the same market, and they divide companies from the same country.
14
Who gains from opening and who from closing
The parallel between software and AI holds up in the logic of the business. That is how Nathan Lambert, author of the Interconnects newsletter, reads NVIDIA’s bet, and how the Indian think tank Takshashila reads China’s [49] [50]. We believe the parallel breaks down in how ownership works. Software has already gone through the two business calculations that divide AI today. The first is to open the layer someone else sells in order to get paid in one’s own. A layer is each link in the chain, such as the chip, the cloud or the model. The second is the fight over ownership of the input, which in software was patents and in AI is training data.
Cheapen someone else’s layer
Joel Spolsky, a programmer and entrepreneur, stated the rule in 2002: “Smart companies try to commoditize their products’ complements” [577]. A commodity is a generic product, bought on price because it makes no difference who sells it. A complement is what is usually bought together with another product, like gasoline with a car [577]. The reason is that demand for a product rises when the price of what goes with it falls [577]. His example was IBM, which was becoming a consulting firm and therefore benefited from cheap enterprise software. Consulting is the complement of that software [577]. In December 2000, IBM had announced that it would spend US$1 billion on Linux, the free operating system, the following year, according to press reports at the time [578].
IBM applies the same rule to AI today. In 2024 its chief executive, Arvind Krishna, said the company wanted to use open source to do with AI “what was successfully done with Linux and OpenShift” [579]. OpenShift is a platform from Red Hat, the company IBM bought in 2019 [579]. Its Granite models are released under Apache 2.0, a permissive license (see “What open and closed mean”) [580]. It offers its clients, against intellectual property claims, the same protection as with the rest of its products [580] [581]. That is, if a third party sues a client alleging that a model developed by IBM infringes its rights, IBM commits by contract to cover the cost, with no cap on the amount [580] [581].
The same rule orders the positions in the dispute. The table puts what each actor holds next to where it gets paid. The last column is our reading, except where the quote comes from the company itself.
| Who | What it holds | Where it gets paid | What it gains from its stance |
|---|---|---|---|
| Anthropic | Closed model, mandatory testing for every capable model and no advanced chips for China; did not sign the letter [291] [217] | Access to its models: 55% of enterprise API spending [66] | Keeping the model’s price. The open token costs on average US$0.13 per million and the closed one, US$0.70 [66]. According to the company, also preventing a model capable of helping with a cyber or biological attack from being released without testing [291] |
| OpenAI and Google | Main model closed and smaller versions open, gpt-oss and Gemma; signed the letter [292] [229] [217] | Access to their models: 31% and 9% of API spending [66]. Google, also, in its cloud | That open models stay a generation behind. They open what does not compete with their main model |
| Microsoft | Published the letter; Phi open [306] [231] | Azure, which serves models of both kinds, and about 24% of OpenAI, which committed to buy US$250 billion in services from it [582] [583] | It gets paid in the cloud whoever wins |
| Meta | Open until 2025, hybrid since 2026 [297] [301] | Advertising: “selling access to AI models isn’t our business model” [297] | Not being locked into a competitor’s closed ecosystem, as happened to it with Apple in phones, according to Zuckerberg [297] |
| NVIDIA | Open and closed together; Huang co-wrote the letter [25] | Chips: data centers provided 89.7% of its revenue in fiscal year 2026 [11] | Demand for open models “promotes use of our products” [14]. Many customers instead of a few labs that design their own chips |
| IBM | Granite open, under Apache 2.0 [580] | Consulting and enterprise platform [579] | Repeating with AI what it did with Linux [579] |
| Alibaba | Qwen open, with its own license for the largest ones [240] | Cloud, which grew 45% [319] | An open model ecosystem “drives greater demand for our cloud computing services” [584] |
| DeepSeek, Moonshot and Z.ai | Open flagship models [309] [238] [241] | Low-priced service; DeepSeek’s website and app are free [320] | Adoption outside China. Opening costs them traffic: the share of Z.ai’s model served through its own platform fell from 88% to 22% in twenty days [321] |
| ByteDance | Main model closed, smaller ones open [311] [312] | Service, through Volcano Engine and Doubao [311] | The same as the closed labs of the United States |
| Mistral | Opens its small models and reserves the medium ones; opening is a technique of “displacement of established players” [244] [567] | Service and contracts to resell its larger models [244] | Competing with incumbents without matching their spending |
| Hugging Face | Signed the letter [306] | The platform where models are published, which NVIDIA agreed to buy [24] | More models being published |
| Andreessen Horowitz | No regulatory barriers for open source; licensing data would only be within reach of the richest [573] [585] | Investment in startups | Cheap models for the companies it invests in |
We believe the match between stance and business is high, but not total. Whoever sells access to the model defends control. Whoever gets paid in another layer defends diffusion. Whoever arrives late defends openness, which displaces incumbents because it offers for free what they charge for. When the business changes, so does the stance. Microsoft, whose chief executive called Linux a cancer in 2001, published the July 2026 letter in favor of open weights [586] [306]. Meta closed Muse Spark when it created its superintelligence lab and opened again months later (see “Who is on each side”).
However, the same arguments are repeated by actors that get paid in different layers. They are also repeated by others that sell neither models nor chips, such as the Swiss public universities that publish the open model Apertus, or the Free Software Foundation. That foundation is a nonprofit organization that defends the freedom of software users [587]. In the case of those universities and the Free Software Foundation, the argument is not explained by business.
Llama’s license puts a business limit on openness. Anyone offering products with more than 700 million monthly active users must request a separate license from Meta, which Meta grants at its discretion. Until it has one, it cannot use the model [588]. In this way, Meta cheapens the complement for everyone except its largest rivals.
Owning the input
Software patents. The dispute between free software and companies was fought over patents. Patents ended up as a weapon of mutual deterrence more than as a source of revenue, and model licenses inherit the clauses that achieved that deterrence.
Since the 1980s, the free software movement has defended four freedoms of the user: to run the program, to study and change it, and to redistribute it with or without changes [589]. In 1998, part of that community adopted the name open source to talk to companies about its practical advantages rather than about ethics [590]. Richard Stallman, founder of the movement, replied that the open source idea “values mainly practical advantage and does not campaign for principles” [591].
Steve Ballmer, then chief executive of Microsoft, called Linux in 2001 “a cancer that attaches itself in an intellectual property sense to everything it touches,” according to The Verge [586]. In 2007, Microsoft told Fortune that free software infringed at least 235 of its patents, but it did not identify any [592]. In November 2006 it had signed an agreement with Novell, which distributed the SUSE version of Linux. Under that agreement, Microsoft committed not to sue the customers of that distribution over patents, and only them [593]. Users of other Linux distributions were left without that protection.
The Free Software Foundation’s license is the GPL, the GNU General Public License. Anyone who distributes a program covered by it, modified or not, must do so under the same license and provide the source code, so the program remains free even when it changes hands [594]. The foundation responded to the agreement with version 3 of its license, which prohibits distributing the program to anyone who signs such an agreement after March 28, 2007 [595] [596]. For its part, the Apache 2.0 license already granted the user a patent license, which is lost if that user sues alleging that the program infringes a patent [597]. Unlike the GPL, Apache 2.0 allows modified versions to be distributed under different license terms [598].
Over the years, Microsoft switched sides. In 2018 it joined the Open Invention Network, a cross-licensing network that protects Linux, in which member companies allow each other to use their patents on that system, bringing more than 60,000 patents, and it acknowledged the “friction” of the past [599]. In 2020 its president, Brad Smith, said Microsoft had been “on the wrong side of history” [586]. In July 2026 it was Microsoft that published the letter in favor of open weights [306]. The Gemma 4, gpt-oss and Granite models are released under Apache 2.0, and Llama’s license copies the retaliation clause: anyone who sues Meta alleging that Llama infringes their intellectual property loses the license [588].
Training data. Today the fight over ownership is over data, and it already costs money:
- Reddit signed data licenses in January 2024 worth US$203 million, with terms of two and three years [600]. Those licenses give the buyer access to what Reddit users post, which the company considers key to training the leading language models [600].
- Meta discussed spending up to US$100 million on book licenses and dropped the idea when it saw how hard they were to negotiate, according to the ruling in Kadrey v. Meta, which sided with Meta on training [601].
- Anthropic agreed to pay US$1.5 billion to the authors of books downloaded from pirate libraries. Judge Araceli Martínez-Olguín, who inherited the case from Judge William Alsup, gave final approval to the settlement on July 20, 2026, and the payment comes to about US$3,000 per work, according to Bloomberg Law [602].
In the United States, Judge Alsup separated training from the way the data was obtained; Judge Chhabria did not. The doctrine at stake is fair use, which allows a protected work to be used without permission depending on the purpose, the amount used and the effect on its market. A use is transformative when it gives the work a purpose different from the original.
Two judges of the U.S. District Court for the Northern District of California applied that doctrine in June 2025. In Bartz v. Anthropic, three authors sued the company for training Claude on their books. Anthropic had downloaded millions of books from pirate sites, and had bought and scanned others. Judge William Alsup ruled that training was fair use, being “exceedingly transformative,” but that this does not excuse the pirated copies [603]. In Kadrey v. Meta, the case of an open model, thirteen authors sued Meta for training Llama on their books downloaded from pirate libraries. Judge Vince Chhabria sided with Meta because the authors “made the wrong arguments.” According to the judge, the argument that could have won was that the model would flood the market with similar works, and the authors did not prove it [601]. The judge clarified that his ruling does not establish that training on protected works is lawful [601].
We believe the first appellate ruling on AI training came on September 29, 2026. The Third Circuit, the federal court of appeals that reviewed the ruling of the Delaware court, held that Ross Intelligence did not make fair use of Westlaw’s headnotes [604]. Westlaw is the legal research service of Thomson Reuters, and its editors write summaries, called headnotes, of the points of law in each ruling. Ross, a startup that competed with Westlaw, trained its legal search engine on memos drawn from thousands of those headnotes [604]. The court clarified that Ross’s AI was not generative. It measured the harm in the market for licensing those headnotes to train AI, a market that “is rapidly developing” [604].
That reasoning reinforces itself, because each license that is signed makes it easier to prove that whoever does not pay causes harm. James Gibson, a law professor at the University of Richmond, described that effect for intellectual property in 2007 [605]. Michael Weinberg, executive director of the Engelberg Center on Innovation Law and Policy at New York University, warned in 2024 that AI data deals can undermine the fair use defense [606].
The New York Times case against OpenAI has no ruling yet. The newspaper sued OpenAI and Microsoft in December 2023, in the U.S. District Court for the Southern District of New York, for using its articles to train the models behind ChatGPT [607]. In the docket that consolidates that suit with others against OpenAI (In re: OpenAI, Inc. Copyright Infringement Litigation, 1:25-md-3143), the U.S. Department of Justice argued on September 1, 2026 that training a generative model is a transformative use, as the Third Circuit itself cites [604]. The U.S. Copyright Office, which registers works and advises Congress on the subject [608], expects some training uses to be fair and others not. It recommends letting the licensing market develop without government intervention [585].
In Europe, two courts reached opposite decisions on whether a model’s weights contain the works it was trained on. In London, the photo agency Getty Images sued Stability AI over its image generator Stable Diffusion. The High Court ruled in November 2025 that Stable Diffusion “does not store or reproduce” those works, so distributing the model is not distributing copies [609]. The judge gave Getty permission to appeal on that point, which the Court of Appeal has yet to decide [610]. A week later, the Munich Regional Court ruled against OpenAI because song lyrics from the authors’ society GEMA were “reproducibly fixed in the models” [611]. The case concerned the lyrics of nine German songwriters. The court rejected the argument that this memorization was covered by the text and data mining exception, which allows works to be copied in order to analyze them [611].
Laws also address the issue. The EU AI Act exempts open models from much of its documentation, but not from the copyright policy or the public summary of training data (see “How the law treats each model”) [612] [613]. Meta did not sign the code of practice that spells out those obligations [87]. Anyone who does not sign it must show the European Commission, by other means, that it complies with them (see “Who regulates AI”). Japan allows, in Article 30-4 of its Copyright Act, the use of works for data analysis, that is, to extract and compare information from a large number of them, unless the rights holder is unreasonably harmed [614]. China requires services to use data “from lawful sources” [574]. The requirement is in its 2023 Interim Measures for the Management of Generative AI Services, which apply to anyone offering those services to the public within the country and also prohibit them from infringing the intellectual property of others [574].
Why data are not patents. The parallel breaks down at three points:
- Data are the raw material. Software patents covered techniques and were hard to prove, and Microsoft did not identify its own. Data are the very input of the model. They are easy to prove when the model memorizes them, and they are already paid for in the hundreds of millions.
- Transparency changed sides. Public code helped free software defend itself against patents. With data the opposite happens, because publishing hurts. The lawsuit against Meta was built on the paper in which Meta presented Llama, which acknowledged using Books3, a dataset made up of books. From that fact, the authors traced that collection to a pirate site [615] [616]. EleutherAI, which publishes its data, says the lawsuits have “drastically” decreased the transparency of companies [617].
- There are no cross-licenses to sign. In software, both sides had patents and could commit not to use them against each other. The owner of the content is not part of the community and has nothing to exchange. The Open Source Initiative chose not to require data in its Open Source AI Definition (see “What open and closed mean”), so that open and proprietary systems stand on equal ground: “both can train on the same kind of data” [618].
Those who defend open models say that licensing data would only be within reach of the largest companies. Before the Copyright Office, Andreessen Horowitz warned that the cost would leave the market to a few. Hugging Face said it would exclude all but the largest companies. Meta argued that no company could pay royalties for the uses third parties make of its models [585]. A per-use royalty does not cost the same to one who charges for each query as to one who publishes weights anyone can download. The former can add it to its price. The latter does not charge for each use and does not even know how many times its model is used.
However, the fair use defense is not exclusive to open models. OpenAI told the British Parliament in 2023 that it would be impossible to train leading models without copyrighted material [619]. Moreover, openly licensed corpora do not yet reach the frontier. Common Pile is an eight-terabyte collection of openly licensed text, assembled to train language models. Models trained on Common Pile compete with Llama 1 and 2 at their size, which is the 2023 level [620].
Patents, by contrast, weigh little in this dispute. China registered more than 38,000 generative AI inventions between 2014 and 2023, six times as many as the United States [621], and it is still the largest producer of open models. We found no significant patent litigation against an open-weights model.
We believe geography decides the business model. Where training is fair use and only piracy is punished, as Judge Alsup ruled, open and closed models compete on equal terms, and the cost is buying the copies. Where a de facto license applies, that is, where in practice the data must be paid for, as the Ross and Munich rulings suggest, whoever can pay for it wins. That is the large lab, whether open or closed. The closed lab also has a direct way to spread that cost across what it charges per use. Finally, whoever trains where the law is most permissive, as in China or Japan, competes with an advantage. OpenAI, an interested party, called that situation in 2025 a “copyright arbitrage” between countries [622].
15
Where the money stays
AI money today stays in the scarce links of the chain. It stays first in memory, chip design and manufacturing, and then in the cloud. The lab that makes the model loses money.
That chain explains much of the dispute. The open model cheapens the lab’s layer and leaves the others intact, which is why those who get paid in them defend it. That is the case of NVIDIA, which according to Nathan Lambert funds open models to sustain its chip business [49]. Kwon Nam-hoon, president of the Korea Institute for Industrial Economics and Trade, doubts that the labs will keep the largest profit. He expects the profit opportunity to move up the chain, to the bottlenecks that persist in compute, memory and energy. He also expects it to move down, to applications, distribution and data, and competition to shift much of the value to consumers [623].
The operating margin in the table is what remains of each dollar sold after paying for production, research, sales and administration.
| Link | Company and period | Revenue | Operating margin |
|---|---|---|---|
| Chipmaking equipment | ASML, July to September 2026 | €11.662 billion | 44% [624] |
| Manufacturing | TSMC, July to September 2026 | US$46.72 billion | 59% [53] |
| Memory | SK hynix, July to September 2026 | 104.6 trillion won | 80% [8] |
| Memory and manufacturing | Samsung, chip division, July to September 2026 | 147.1 trillion won | 72% [51] |
| Memory | Micron, fiscal year ended September 3, 2026 | US$133.188 billion | 75% [52] |
| Chip design | NVIDIA, August to October 2026 | US$115.844 billion | 67% [198] |
| Cloud | AWS, July to September 2026 | US$45.984 billion | 39% [54] |
| Cloud | Google Cloud, July to September 2026 | US$30.709 billion | 36% [55] |
| Lab | Anthropic, July to September 2026 | US$16.956 billion | −3%, with US$5 billion in short-term contracted compute [56] |
| Lab | OpenAI, July to September 2026 | about US$10 billion | about −109%, adjusted [625] |
The margins are our calculation from each company’s figures, except OpenAI’s, which does not publish its accounts. A negative margin is a loss: OpenAI’s, about −109%, means that it spent about US$2.09 on its operations for every dollar it took in.
Scarcity sets the margin
Memory makers today earn more per dollar than NVIDIA. SK hynix made 80 cents of operating profit per dollar in the third quarter, and expects demand to exceed supply until 2030 [8]. Micron made 75 cents per dollar in its fiscal year. The company explains that HBM, the high-bandwidth memory used by AI chips, consumes more silicon wafers per bit than conventional memory [52]. That is why a fab produces less memory when it switches to HBM, and supply falls short.
ASML, the only maker of the most advanced lithography machines, the ones that print circuits onto chips, earns less per dollar: 44 cents in the quarter, against 33 a year earlier. It says some customers already want to place orders for 2029 and 2030, and that the supply it planned for 2027 and 2028 will not cover all the demand [624]. NVIDIA foresaw part of that scarcity and asked memory companies to invest in HBM (see “NVIDIA orders the chain”).
Each link’s margin follows what is scarce at the moment, not its place in the chain. Kwon Nam-hoon, who follows the economist David Teece on this point, argues that profit migrates toward scarce complementary assets, what is needed besides the invention to get paid for it, such as exclusive data or the relationship with customers [623]. Teece calls the ability to capture the returns from an innovation appropriability. According to Kwon, in AI it is low relative to what is invested, because papers, the movement of researchers, public benchmarks and distillation, which consists of training a model on the answers of another, quickly spread what the leader discovers [623]. In 2026 memory is scarce, and that is why its makers earn more than whoever designs the chip. The lab, which buys all of that through the cloud, is the only link in the table that loses.
The chip and the cloud
NVIDIA made 66 cents of operating profit per dollar sold from May to July 2026, according to our calculation from its quarterly report, and 67 from August to October, the quarter in the table [14] [198]. It finances part of those sales itself, because it invests in companies that buy its chips, such as CoreWeave, and contracts cloud services from them. As of July 26 it had US$99 billion invested in the equity of other companies and another US$25 billion committed [14]. It had also committed to buy US$36 billion in cloud services from AI clouds, companies that rent compute specialized for AI and that are largely suppliers that also buy its chips. On top of that come US$29 billion in cloud to develop its own open models and its autonomous driving software [14]. CoreWeave, which rents NVIDIA GPUs and has NVIDIA as a shareholder, lost US$709 million in the quarter [626].
The cloud earns a margin, but spends more cash than it generates. Amazon, Alphabet, Microsoft, Meta and Oracle invested US$231 billion in the last quarter each one reported, which at that pace is equivalent to about US$925 billion a year, according to our sum of their reports [54] [627] [583] [628] [629]. In China, the research firm Rhodium Group expects AI investment to double in 2026 to 932 billion yuan, about US$139 billion. That figure is between 15% and 20% of the U.S. figure [630].
Each company’s accounts show that spending. Amazon has accumulated negative free cash flow of US$42.357 billion over twelve months [54]. Free cash flow is cash from operations minus what is invested. Alphabet had negative free cash flow in the second and third quarters of 2026, and in June it raised US$49.5 billion in capital [627] [55]. Oracle has US$664 billion in contracts yet to be collected, of which it expects to collect 13% in the next year [628].
The cloud’s margin depends on the labs paying for what they committed to buy from it. Torsten Slok, chief economist at Apollo, warns that the profitable links depend on the money-losing lab continuing to raise financing [631]. The credit market already reflects that risk: the cost of insuring against a CoreWeave default, which rises when investors see it as more likely that the company will not pay its debts, hit a record on November 5 (see the public affairs study, “The bubble risk”) [632].
The lab
Anthropic is the only frontier lab in the United States with public accounts. It disclosed them in its filing to go public, and it expects to price its shares on November 18 [56]. Its revenue went from US$41 million in 2023 to US$4.585 billion in 2025, and to US$33.242 billion in the first nine months of 2026. Of that, 79% comes from usage-based billing, and 44% was sold through the marketplaces of Amazon, Google and Microsoft [56].
Its gross margin is what it keeps from each dollar after the direct cost of serving the model. It is 85%, according to our calculation from note 8 of its filing [56]. That margin describes only the compute that serves paying customers, which is about a quarter of its compute spending. The rest goes to training, research and serving free users. In total, Anthropic spends 57% of what it takes in on compute, and in the nine months it lost US$1.26 billion on its operations. Its net loss was US$104.039 billion, but almost all of it is an accounting loss and does not correspond to spending. It comes from the revaluation of its convertible notes, which are debt repaid in shares; of derivatives, contracts whose value depends on the price of another asset; of its SAFEs, agreements that give the right to receive shares in a future round; and of the fair value, the estimated market price, of its financing agreements with partners [56]. Because those instruments are paid in shares or depend on their value, when Anthropic’s value rises, what it owes on them rises too, and the accounting records that difference as a loss even though no money goes out.
Suppliers finance the lab, and that money returns to them as compute purchases. Fortune and Stacy Rasgon, of Bernstein, pointed this out about NVIDIA and OpenAI [633]. Anthropic has committed to buy about US$900.9 billion in compute, including at least US$121.3 billion from Google, US$110 billion from Amazon and US$47.2 billion from Microsoft [56]. Alongside the public offering, it will sell US$26 billion in shares privately. The buyers are Google, NVIDIA, AMD, Broadcom, Micron, Microsoft and TSMC, all suppliers of chips or cloud [56]. Microsoft, for its part, recorded US$10 billion in revenue in the quarter from its contracts with OpenAI, a company in which it holds about 24% [583].
In its risk factors, Anthropic says open models compete for workloads that prioritize “cost efficiency over intelligence” and can pressure its prices [56]. It also warns about distillation, weight theft and the spread of knowledge. If they put the ability to make powerful models within reach of more companies, it says, competition could increasingly depend “on access to large-scale compute rather than on model development capability” [56].
The price of the model
Open models push down the price of what already exists, because any provider can serve an open model and compete on price. Anthropic acknowledges that they pressure its prices in tasks where cost matters most [56]. The closed lab charges for the frontier, the most capable model, and that advantage lasts from two to six months according to Ben Lorica, author of the Gradient Flow newsletter [634].
The price of a fixed capability falls fast. The cost of using a model as capable as GPT-4 fell about 40-fold per year, according to Epoch AI [635]. By contrast, what is paid to use the most capable model rises, because models are larger and reason more. The increase is between 3 and 18 times per year, according to a 2025 academic study [636].
Closed labs charge almost the same for models of equivalent level. Input tokens are the text sent to the model, and output tokens, the text the model returns. Opus 5.5 costs US$4 per million input tokens and US$20 for output, the same as gpt-5.6-sol at the promotional price OpenAI is keeping until November 30 [637] [638]. On December 1 Google will double the price of Gemini 4 Carbon and Argon, from US$2 to US$4 for input [639]. DeepSeek V4 Pro charges US$3.96 per million output tokens at peak hours and US$1.98 off-peak [640]. At peak hours, that is about a fifth of what Opus 5.5 and gpt-5.6-sol charge for output.
Who pays for open
The open labs that publish their accounts sell services and still lose money:
- Zhipu, the company behind Z.ai, is listed in Hong Kong. It took in 724 million yuan in 2025 and lost 4.718 billion, more than six times its revenue [57]. In the first half of 2026 it took in 954 million, 86% from its API. Between the start of the year and September it raised the average price of that API by 101%, while volume grew more than fortyfold. Its gross margin fell from 50% to 26% compared with the first half of 2025, because the share of its cloud business grew, and that business earns less than installations on each client’s own servers. The margin of the cloud and the API, by contrast, rose from −0.4% to 24.6% [641]. The company says that, commercially, “open-source and safety are not cost items.” Its reason is that the distribution open source provides is the condition for the model to receive feedback from real tasks [641]. Its shareholders include funds controlled by the state-owned assets commissions of Beijing and two of its districts [642].
- MiniMax took in US$79 million in 2025, 73% of it outside mainland China, and lost US$251 million on its adjusted measure, the one the company calculates excluding some accounting items that are not cash payments [58]. Alibaba owns 11.36% of the company and sold it cloud services worth US$75.9 million in 2025. That purchase equals 96% of MiniMax’s revenue, according to our calculation: what MiniMax paid one of its shareholders for cloud almost matched what it charged all its customers [643].
- Alibaba has reported separately, since June, the AI Labs and Applications segment, which makes Qwen and its app. That segment’s adjusted EBITA, its earnings before interest, taxes and amortization, was a loss of 13.861 billion yuan in the quarter to June. That is more than twice what its cloud earned, 5.628 billion [319]. It is the most explicit cross-subsidy we found in a report. A cross-subsidy exists when a company covers the loss of one business with what it earns in another.
- DeepSeek does not publish revenue. The US$5.6 million cost attributed to V3 is the theoretical rental of the GPUs for the final training run. Its own report clarifies that it excludes research and prior experiments [644]. Its theoretical margin of 545% was measured on cost, not on revenue like those in the table: it meant that the service would take in about US$6.45 for each dollar of rental of the inference GPUs, the only cost it deducted [320]. DeepSeek calculated it assuming it charged for all tokens at the price of its R1 model, and clarified that its actual revenue was much lower [320]. Until April 2026 it was funded only by High-Flyer, its founder’s fund, according to press reports [645].
- Meta wrote in its annual report that it expects to continue training “a combination of open and closed models” [646]. In October 2026 it told its investors that training is a large fixed cost and that this is why it wants to serve all the high-margin inference demand it can [647].
The value of open models is captured, in large part, by whoever serves them. On Vercel’s AI Gateway, the service that routes applications’ calls to each lab’s models, open models processed 71.5% of tokens in September 2026 and took 22.6% of spending: they did almost three quarters of the work and took less than a quarter of the money, because their token is much cheaper. Almost 70% of what was spent on Chinese models went through U.S. providers [648]. That spending is collected by the provider that serves the model, not by the Chinese lab that published it. Fireworks, which serves open and specialized models, passed US$1 billion in annualized revenue, that is, what it would take in over a year if it kept its most recent pace of sales [649]. On the public side, the U.S. National Science Foundation and NVIDIA gave Ai2 US$152 million for fully open models, and Apertus was trained with hours on Swiss public supercomputers [650] [326]. The Foundation is the federal agency that funds science and engineering, and Ai2 is a nonprofit research institute that Paul Allen founded in Seattle [651] [652].
Open models are sustained in three ways. The first is as an expense to sell something else, whether the cloud, the chip or advertising. The second is with capital, often from the state, while the lab builds its sale of services. The third is as a public good. Nathan Lambert describes NVIDIA funding open models to sustain its chip business, and Takshashila describes the Chinese state absorbing their cost [49] [50]. The business is done mainly by whoever serves the models, more than by whoever publishes them.
16
The software industry facing AI
AI makes writing code cheaper, and that changes the business of anyone who sells software built by thousands of engineers. We believe the change hits those who sell applications by license harder than those who charge for the compute on which the new applications that replace them run. We also believe that Microsoft has settled into this second group. The research firm Forrester expects software to move from per-seat pricing, that is, per user, to consumption pricing, that is, by what each customer uses [653].
The work of programming
David Heinemeier Hansson, creator of Ruby on Rails and a partner in the software company 37signals, described on Lex Fridman’s show in August 2026 how his work changed in nine months. We used the episode’s automatic transcript. He says that nine months ago he would have called what he says today “AI psychosis,” and that he himself has become “optional in the part that produces the code” [654].
His example is a code translation. An agent translated a library, a package of code that other programs reuse, from the Python programming language to the Rust language in under 45 minutes. Paid per token, the task would have cost him about US$550. For him, learning Rust to do it would have taken nine months. Other models repeated the task for between US$23 and US$55, and he paid for it with a monthly subscription. From this he concludes that the economic payoff of taking care over every line of code is “diminishing rapidly.” About the programs we only half use, he proposes: “what if we all just build our own 5%?” [654].
Heinemeier Hansson also says that Microsoft has had unlimited programming capacity for decades, and that this shows that writing a lot of code does not produce good software. In large companies, he says, the bottleneck is ideas, vision and taste. He compares those companies to supertankers that cannot turn. He calls the employment figures “fuzzy” and mentions the Jevons paradox, according to which lower prices increase demand. He attributes today’s layoffs to overhiring during the pandemic, with AI as a “convenient excuse” [654].
He also talked about open models. He uses the Chinese model Kimi served by Fireworks, a U.S. provider. He called Anthropic “protectionist” for preventing his subscription from being used with other developers’ programs. He said that he can ask the open Chinese model about Tiananmen, while a closed U.S. model refused to translate an essay on migration into Italian for him [654].
U.S. employment data show the same mix. Software developers went from 1.66 million in 2023 to 1.69 million in 2025. Programmers, a narrower occupation, went from 152,610 in 2021 to 92,230 in 2025 [655]. In each of the three editions of its occupational handbook that we consulted, from March and October 2025 and November 2026, the U.S. Bureau of Labor Statistics lowers the ten-year growth it projects for developers: 17%, 15% and 10% [656]. A Stanford study using payroll data found in its August 2026 revision that employment of young people aged 22 to 25 in the occupations most exposed to AI is 19% below what was expected. The gap comes from their being hired less, not from their being laid off. The authors present the result as a descriptive indicator, not as proof of cause [657].
Huang gave the opposite reading to layoffs in March 2026, with the example of radiology. According to him, it was the first occupation that AI researchers wrote off, computer vision surpassed humans around 2020, and even so the number of radiologists grew and today there is a shortage. If programming becomes describing what one wants, he said, the people who can do it go from 30 million to “probably 1 billion” [15].
What is left for Microsoft
Microsoft closed its fiscal year in June with 223,000 employees, 5,000 fewer than two years earlier. In research and product development it went from 81,000 to 77,000 people, while its revenue grew 35% over those two years, to US$331.839 billion [658] [60]. In 2025 it cut about 15,000 jobs in two rounds, according to CNBC [659] [660], and in July 2026 it cut 4,800 more. Its chief people officer wrote that those roles “are not being replaced by AI” [661]. Satya Nadella, chief executive of Microsoft, told employees in 2025 that the company is in an industry that “has no franchise value,” in which nothing protects what has been won, and that it must shift “from a software factory to an intelligence engine” [662].
Azure passed US$100 billion for the first time in fiscal year 2026 [663]. Microsoft’s total capital expenditure, which is what it invests in data centers, chips and networks, rose from US$44.477 billion in fiscal year 2024 to US$115.948 billion in 2026 [60]. The company does not break it out by business [60]. Since September it has reported a segment called “Agents and Infra,” which brought in 83% of its revenue for the quarter. In that same quarter, Azure grew 48% and revenue from Windows for manufacturers and from devices fell 22% [664]. Nadella said that its business model is expanding “from per-seat subscriptions to include consumption,” and that on GitHub, the Microsoft platform where programmers store and share their code, agents already open more than a third of pull requests [61]. A pull request is a proposal to modify the code that another person reviews before it is merged.
Microsoft gains when software becomes cheaper, as long as it runs on its cloud, because it charges for the compute that software consumes. It opened the code of its programming assistant in the VS Code editor under the MIT license, which is permissive (see “What open and closed mean”), and publishes Phi under the same license. It also offers DeepSeek and gpt-oss in its catalog, and published the letter in favor of open weights [665] [231] [666] [306]. Its revenue from open models in that catalog grew tenfold, according to Nadella, who did not give the period [61].
What remains for Microsoft’s engineers is building and operating the infrastructure, the agents and the layer that governs them, more than writing new applications. Nadella wants to remake every layer, from infrastructure to agents, and to make tools so that everyone can build their own [662]. The company’s risk lies in two places. One is the gross margin of its cloud, which fell to 66% because of the cost of AI [60]. The other is that charging by use makes its revenue more variable than charging by license, because the customer pays only for what it consumes and not a fixed fee per user.
Software sold by license
On January 29, 2026, Anthropic published a plugin for legal tasks for its assistant Claude, under the open Apache 2.0 license [667]. Between February 2 and 3, Thomson Reuters, owner of the legal service Westlaw, fell 15.7% on the stock market, Intuit 10.9% and iShares’ exchange-traded fund of software companies 4.6%, according to our calculation with Yahoo Finance prices [62]. That fund is bought on the stock market like a share and bundles the shares of many software companies, so its price tracks the whole sector. As of November 13, that fund had recovered almost everything and was down 4.1% for the year, while Microsoft was up 8.5%. By contrast, Intuit was still down 48%, Oracle 36%, ServiceNow 33% and Adobe 31% [62].
Executives in the sector read the risk in different ways. Nadella had anticipated in 2024 that, in the agent era, the very notion of business applications is “probably where they’ll all collapse” [668]. Huang called the idea that AI replaces software “the most illogical thing in the world” [669]. Klarna, which had announced that it would drop Salesforce and Workday, later clarified: “we didn’t replace SaaS with an LLM” [670]. SaaS is software paid for by subscription, and an LLM is a language model. Salesforce announced that it would not hire engineers in 2025 [671]. In January 2026 it had 83,334 employees, more than a year earlier, with the acquisition of Informatica included [672].
The threat is to pricing and to per-user billing, more than to the existence of these companies, as Forrester and the consulting firm Bain argue [653] [673]. The February drop coincided with the release of a free software component. That component is Anthropic’s legal plugin, which anyone can use and modify without asking permission.
Free software
AI uses free software more than ever, but it no longer sends its maintainers (the people and companies who develop it and keep it up to date) the traffic and customers that used to fund them. Dries Buytaert, founder of the free content management system Drupal, wrote this about Tailwind [59]. Tailwind is a free tool for designing web pages, and it had never been so widely used. However, traffic to its documentation, which is where the customers for its commercial products come from, fell about 40% since 2023, and its revenue fell almost 80%. In January 2026 it laid off 75% of its engineering team, and its founder wrote that he did so because of the impact AI had on the business [63].
Other projects receive both burden and help from AI. The curl project, a free tool for transferring data over the internet, closed its security bug bounty program, which paid whoever reported flaws, in January 2026. The reason was that junk reports made with AI lowered the share of confirmed reports from more than 15% to less than 5% [674]. The same maintainer had merged some 50 fixes found by AI analysis tools [675]. Heinemeier Hansson believes this is the best time to maintain a free project, because free contributions arrive [654].
17
What large companies allow with their data
More than choosing between open and closed models, large companies decide what can be done with their data and who authorizes it. None of the public rules we reviewed, whether from companies or regulators, requires an open model or bans a closed one. Those rules require that the data not be used for training, that it not be stored, that it be known where it is processed and that someone authorize it in writing. Companies respond to those requirements with five stances, which they often combine: banning, setting conditions for their suppliers, building their own portal on top of contracted models, training their own model or installing an open one in house, that is, on their own servers. Each stance gives the business to a different provider.
Ban first, then contract
In 2023, after Samsung employees pasted internal code into ChatGPT, the company banned generative AI tools on its devices and its network. It presented the ban as a temporary measure and warned that anyone who violated it could be fired [676]. Other organizations did something similar. In a 2024 Cisco survey, 27% of privacy leaders said their organization had temporarily banned generative AI applications. In addition, 63% had limited what data can be entered into them and 61% which tools employees can use [677].
The bans targeted the consumer version of these tools, not AI itself. Where there is public follow-up, they ended in a company-approved tool. JPMorgan launched its own portal with OpenAI’s model in 2024 [678]. Goldman Sachs launched a similar one in 2025, with the models behind its firewall, that is, inside its own network [679]. Samsung, which had trained its own model, called Gauss, allowed its appliances and phones division to use ChatGPT, Gemini Enterprise and Claude in June 2026. Its chip division still has no date to do so [680].
Restrictions are not enough to stop use either. In a survey by KPMG and the University of Melbourne, 46% of U.S. workers said they had uploaded confidential company information to public AI platforms [681].
What they require of their suppliers
Some large companies publish the AI rules they impose on their suppliers. All the ones we found regulate the use of data, not the type of model:
- AT&T bars its suppliers from processing its information in any AI system without a written agreement, and requires them to delete it “following its intended use.” Its definition of a model is broad. It includes in-context learning, in which the model receives examples or documents within the query without its weights changing. It includes RAG, the technique that has a model search a set of proprietary documents before answering. It also includes embeddings, the numerical representation that allows those documents to be searched by meaning. With that definition, the ban also reaches the supplier that only queries AT&T’s information without training any model. In addition, everything generated with AI must be delivered labeled and reviewed by a person [64].
- BASF bars the supplier from training its AI on the company’s data, that is, on what BASF enters into it and the answers it receives, and also from reviewing or monitoring that data, unless there is separate consent. It counts RAG as training and asks that changes that increase the risk to people be approved in advance [65].
- Procter & Gamble bars the use, without its written consent, of tools that expose its data to third parties. The ban includes internal tools that could train on what is entered into them, “even if inadvertently and resulting from a technical setting alone” [682]. It is enough, then, for the tool’s configuration to allow that training, even if no one sought it.
- Kenvue asks for written consent for any use of its data with generative AI. In its list of examples it includes Microsoft Copilot 365, the AI assistant that Microsoft builds into Word, Excel and Outlook, along with ChatGPT and Gemini [683]. The rule thus reaches everyday office tools.
- Clifford Chance, a global law firm, bars training on its data and requires the supplier to state which foundation model it uses, the general-purpose model on which it builds its product [684].
Governments ask for the same. The White House ordered federal agencies to make their contracts prohibit training commercial models on agency data [685]. The order is in memorandum M-25-22, which its Office of Management and Budget issued in April 2025 to set how the federal government buys AI, except for national security systems. The ban is permanent, covers nonpublic data and yields only to the agency’s express consent. The memorandum also calls for clauses against dependence on a single vendor, such as portability, which lets the agency take its data and models to another vendor, and it applies to contracts from solicitations issued starting 180 days after its date [685].
The European Commission prepared model clauses for public procurement, which are still a draft. They have a full version for high-risk systems under the EU AI Act (see “Who regulates AI”) and a light version for the rest [686]. Those clauses limit the use of data to what the contract covers and require it to be destroyed upon request [687].
The rules of each sector
Several of those clauses are explained by the rules of each sector. In the European Union, the financial sector’s digital operational resilience regulation, known as DORA, requires every IT services contract to specify the countries where data is processed and stored, and to give notice before changing them [104]. DORA has applied since January 17, 2025 to banks, insurers, investment firms and the other financial entities of the Union [688]. In the United States, the health law HIPAA makes every vendor that processes patient information on behalf of a hospital or an insurer a business associate [105]. The hospital or insurer may hand over that information only if the vendor commits, in a written contract, to protect it [689]. That contract bars the vendor from using the information for anything other than what was agreed, and requires it to report any breach and to impose the same conditions on its subcontractors [690]. An AI provider that receives patient records has to sign it first.
The U.S. Department of Defense, for its part, requires that a cloud storing military information have the security level of federal systems [107]. That level is the moderate level of FedRAMP, the program through which the federal government sets the security requirements for cloud services [107]. The new model risk guidance from U.S. bank supervisors, of April 2026, expressly excludes generative AI and agentic AI, the kind that carries out tasks on its own [110]. For those systems it sets no requirements of its own and refers to each bank’s general risk management and governance practices [110]. That guidance replaces the 2011 guidance and describes how banks should validate and monitor the quantitative models they use to make decisions. It is aimed mainly at banks with more than US$30 billion in assets [110].
Lawyers have a stricter rule, because professional confidentiality can be lost. The regulator of solicitors in England and Wales warned in August 2026 that putting a client’s documents into a public AI tool will likely breach confidentiality. It also warned that doing so can cause privilege to be “permanently waived” [109]. Privilege is the protection that prevents a client from being compelled to disclose what they discussed with their lawyer. The State Bar of California bars entering a client’s confidential information, without the client’s informed consent, into a tool that puts it at risk [691]. The rule is in its practical guidance of May 2026, which replaces the 2023 guidance and applies to generative and agentic AI the ethical duties the lawyer already had [691].
Their own models
Some companies trained their own models from scratch. In 2023 Bloomberg built one with 50 billion parameters using its financial data [692]. Samsung trained Gauss and uses it in its internal programming assistant [693]. NTT sells tsuzumi, a Japanese model built from scratch that runs on a single GPU. The company offers to install it in house or in a private cloud to handle confidential information [694]. All three companies have research capacity and a commercial interest in the model.
The banks JPMorgan and Goldman Sachs took another path. They built their own portals on top of contracted closed models, behind their firewall [678] [679]. We found no large non-technology company that documents in its own channels, as its main platform, an open model that has been fine-tuned, that is, trained a little more with its own data, and installed in house. Open models appear as a piece within those portals or in the offerings of providers such as Mistral, Cohere, IBM and SAP. A company’s own model also has a limit, and Samsung’s case shows it: the company had had its own model since 2023 and even so opened the door to three closed models. According to the company, Gauss will remain in use alongside them because it is integrated into its internal systems [680].
What the labs offer
The closed labs adjusted their offerings to those rules. OpenAI has not trained on what its API receives since March 2023 [695], Anthropic does not train on its commercial customers’ content [696], Google does not do so without the customer’s prior permission [697] and AWS does not share content with the developers of the models it sells [698].
Zero retention, that is, the commitment not to store anything the customer sends, has limits. OpenAI grants it only with prior approval [695]. Google warns that it is not possible in some advanced features [697]. AWS stores all traffic with Claude Fable 5, 5.1 and 5.2 for up to 30 days to detect abuse automatically, and what its classifiers flag can go to human review. It does so because those models require retention in order to be used [115] [699]. With those models, a clause like AT&T’s, which requires deleting information following its intended use, cannot be met unless an exception is negotiated.
Providers of open models sell the opposite. Mistral offers its assistant for installation on the customer’s equipment or in its private cloud [111]. Cohere promises that in its private deployments it has no access to the data [112]. IBM sells its Granite models under the Apache 2.0 license for installation in house [113]. SAP integrated Mistral’s models so that its customers can build sovereign applications within their own environment [114].
Who gains from each company stance
The columns on who gains and who loses are our reading, except where they cite data.
| Stance | Who adopted it | Who gains | Who loses |
|---|---|---|---|
| Ban external tools | Samsung in 2023; 27% of the organizations in the Cisco survey, temporarily [676] [677] | The free consumer versions, which continue to be used in secret: in the United States, 46% of workers said they had uploaded confidential company information to them [681] | All enterprise providers, while the ban lasts |
| Set conditions for suppliers | AT&T, BASF, Procter & Gamble, Kenvue, Clifford Chance and the U.S. federal government [64] [65] [682] [683] [684] [685] | Suppliers that can sign that they do not train, that they delete and which model they use, and can pay for compliance; the lab that offers zero retention | The integrator whose lab requires retaining data for review [115] |
| Own portal or enterprise version of closed models | JPMorgan and Goldman Sachs; Samsung since June 2026 [678] [679] [680] | OpenAI, Anthropic and Google, whose models are inside, and the clouds that serve them without training on the data [698] [697] | The generic chat and assistance applications that the portal replaces |
| Own model trained from scratch | Bloomberg, Samsung and NTT [692] [693] [694] | Whoever sells the compute, NVIDIA or the cloud, and the company itself when it sells the model, as NTT does | The outside labs, until the in-house model falls short, as happened at Samsung |
| Open model installed in house or in a private cloud | Offered by Mistral, Cohere, IBM and SAP; we found no large non-technology company that documents it as its main platform [111] [112] [113] [114] | Vendors of installation, support and sovereignty, and whoever sells the chips | The closed labs that require retaining data for review |
| Each sector’s requirements on location, contract and confidentiality | European financial sector, U.S. health care and defense, law firms [104] [105] [107] [109] | Providers that sign the HIPAA business associate agreement, meet the federal security level or guarantee the country where the data stays | Public tools, which according to the English regulator can cause professional confidentiality to be lost [109] |
In almost every stance the winner is whoever sells the compute, as the venture capital firm Andreessen Horowitz already concluded in 2023 for the whole generative AI market [700]. A bank’s portal and the open model in a private cloud run in the data centers of Amazon, Microsoft or Google. The in-house model and the open model installed in house run on NVIDIA chips. The closed labs win where the company protects its data with a contract. The vendors of open models win where the data admits no retention at all or where the rule is about sovereignty, as in the case of NTT in Japan or of SAP with Mistral in Europe.
We believe that those who sell solutions built on third-party models win or lose depending on each customer’s stance. In the clauses of AT&T, BASF, Procter & Gamble, Kenvue and Clifford Chance, the supplier bound is that vendor and not the lab. The vendor has to request written authorization, refrain from training on the data even through RAG, delete it, declare the foundation model and label what is generated. It also has to obtain those same commitments from its lab.
That vendor loses against the in-house portal, which already gives access to the models, and also loses when its lab requires retaining the data. It wins if it runs in the cloud or on the customer’s equipment with the model the customer already contracted. It also wins if it brings what a generic portal does not, such as domain data, sector workflows and traceability for each answer, and if it can switch to an installable open model when the data admits no retention.
The surveys do not agree on which stance is growing. In 2025, Andreessen Horowitz found more adoption of open models among large companies, because of their preference for installing them in house [701]. In January 2026, the same firm found that more than a third preferred closed models and that data security was one of the reasons [702]. The Menlo Ventures survey, conducted in September 2026, found that in one year open models went from 11% to 19% of the tokens companies use, but take only 4% of spending [66]. All three surveys come from funds that invest in labs, so none is a disinterested source.
The retention that some frontier models now require, such as the Claude Fable models on Bedrock, the model marketplace of Amazon’s cloud, came after the first two surveys, which therefore do not measure its effect. AWS exempts customers that qualify for its Enterprise Frontier Safeguards program from that retention. The exemption applies to internal use and AWS offers it at least until June 30, 2027. After that it will store all traffic for up to 30 days to detect abuse automatically [699]. That is why we believe that banking, health care and law firms will negotiate that exemption before installing an open model in house [115].
18
How the law treats each model
Export controls
Starting November 19, 2026, no U.S. export rule will treat open and closed weights differently. The AI diffusion rule (Framework for Artificial Intelligence Diffusion), of January 2025, was the only export rule from a major power that did. The Bureau of Industry and Security issued it as an interim final rule. With it, the Bureau extended worldwide the license requirement for exporting advanced computing chips, and it created exceptions for the destinations that, in its judgment, posed no national security risk [703]. That rule created the 4E091 classification, a new number on the Commerce Control List, the list of items whose export the U.S. Department of Commerce controls, and required a license to export the weights of closed models trained with more than 10^26 operations, a 1 followed by 26 zeros. Each operation is an addition or a multiplication, and the total number of operations measures the computation it took to train the model. The rule left out models with published weights because, in the Bureau of Industry and Security’s judgment, their benefits outweighed their risks for now [703].
On May 13, 2025, the U.S. Department of Commerce announced that it would rescind the rule, instructed the Bureau of Industry and Security not to enforce it and promised a replacement rule [704]. After the announcement, however, the text remained in the Export Administration Regulations. That text required a license to export anything classified under 4E091 to any destination [705]. It established a presumption of denial, meaning that a license application is denied in principle, for any end user headquartered outside the destinations in paragraph (a) of supplement no. 5 to part 740 [705]. That paragraph lists the destinations to which the diffusion rule granted a license exception [703]. It also exempted models that did not exceed the most capable open model, so the control reached only closed models better than any that could already be downloaded. It also kept the foreign direct product rule for model weights [705]. That rule subjects to U.S. regulations certain items produced in other countries with U.S. technology, software or equipment, so a model trained outside the United States on U.S. chips could need a license to be exported [703].
On May 12, 2026, the Government Accountability Office ruled that the announcement that the rule would not be enforced was itself a rule, and that it therefore had to be submitted to Congress [706]. That Office is an agency independent of the executive departments [707]. It issued the ruling at the request of Senator Elizabeth Warren. Its ruling rested on the Congressional Review Act, and it could not force the Bureau of Industry and Security to enforce the diffusion rule [706]. That 1996 law requires every agency to send each new rule to both chambers and to the Comptroller General, who heads the Government Accountability Office, before it takes effect, and it gives Congress 60 days to disapprove it through a special procedure [708]. The replacement rule was slated for July 2026, and its description did not mention weights [709].
On November 16, 2026, the Department of Commerce released an interim final rule that formally rescinds the diffusion rule. The interim rule eliminates the 4E091 classification and every control specific to weights, because the Department of Commerce considers it more effective to address their risks with licenses by destination or by end use, that is, according to the country the item goes to or to what will be done with it. It is published in the Federal Register and takes effect on November 19. As an interim final rule, it applies without prior consultation and accepts comments for the following 60 days [22]. The same rule moves the control to compute. A data center in any country needs a license to give access to more than 12 million TPP units, the measure of processing power by which chips are classified, about 760 H100 chips, to companies headquartered, or with a parent company, in Macau or in a Country Group D:5 country. The requirement seeks to keep those companies from using, through the cloud and from other countries, U.S. chips they cannot buy. That group is the list of countries under a U.S. arms embargo, and it includes China [22].
Federal procurement guidance, by contrast, does distinguish open models from closed ones, because it exempts models with open licenses from the ideological neutrality requirements it imposes on the language models that agencies buy (see “The United States sponsors both sides”) [341].
While the diffusion rule remained unenforced, the Department of Commerce restricted a closed model by another route. In June 2026 it ordered Anthropic, without giving its reasons in writing, to cut off foreign users’ access to Fable 5 and Mythos 5. On June 30 it lifted the order. The case is covered in the public affairs study [710].
On October 16, 2026, China’s Ministry of Commerce and Ministry of Science and Technology presented a proposal to restrict exports of AI technologies, whose thresholds are covered in the public affairs study. The proposal amends the catalog of technologies whose export China prohibits or restricts. It adds 40 technologies in AI, biomedicine and photovoltaic energy, 23 of them in AI [290]. In it, China proposed controlling the transfer of its largest closed weights and leaving already published weights and access as a service free [711]. That is the criterion the United States still had in its regulations [705]. In a large model, the proposal restricts the weights, the checkpoints (copies of the weights saved partway through training), the source code and the architecture, as well as the techniques for optimizing its training and accelerating its inference. Whatever has already entered the public domain through open source or academic publication is left out [290].
A model enters the list if it meets any of three criteria: 10^26 training operations or more (the same compute threshold the United States uses), 2 trillion parameters or more, or a place among the world’s top ten in public evaluations. Models that provide a service and were trained with “important data” also enter. Item 266501J goes further with models trained with “core data”: it places them among the technologies whose export is prohibited, which, unlike restricted ones, cannot leave China even with a license [290]. Both categories come from China’s Data Security Law, in force since September 2021. Important data are those that each region and each sector enters in a catalog of enhanced protection. Core data are those that bear on national security, the lifelines of the economy, the livelihood of the population or a major public interest, and they are under the strictest regime [712].
Since October 16, while the catalog awaits approval, anyone who wants to export something on the list must register it beforehand [290]. Opening the code of a listed technology will get an automatic license, so that, according to the spokesperson of China’s Ministry of Commerce, it is basically “declare and it goes” [713]. Offering a restricted model abroad through an API, an app, the web, software as a service or a lightweight on-device installation does not count as an export either. Deploying it in full abroad does count, although the license is automatic if the model runs on the company’s own cloud outside China [713]. The consultation on the proposal closed on November 15 [290]. Starting November 19, that proposal will be the only export control on weights of the two powers.
Pre-deployment evaluation and transparency
European Union. Among the rules in force, the EU AI Act is the one that treats open models in the most detail in evaluation and transparency. It does so in four articles:
- Article 2(12) leaves AI systems under free and open-source licenses outside the regulation. The exclusion does not apply if they are high-risk, or if they fall under the prohibited practices of Article 5 or the transparency obligations of Article 50, among them the obligation to mark synthetic content (see “Synthetic content”) [612].
- Article 25(4) refers to the contracts the regulation requires along the supply chain of high-risk systems. In them, the provider of a high-risk system and whoever supplies it with tools, services or components set down in writing the information and technical access the former needs to comply with the Act. The article exempts from those contracts anyone who publishes tools or components under an open license. The exemption does not cover general-purpose models [612].
- Article 53(2) exempts certain providers of general-purpose models from two documentation obligations. They are those whose license allows access to the model, its use, modification and distribution, and who also publish its weights, its architecture and its usage information. The two obligations are drawing up the model’s technical documentation for the AI Office and the national authorities, and providing information on the model to those who integrate it into their systems [714]. The exemption does not apply to models with systemic risk, the category the Act reserves for the most powerful ones [612].
- Article 54(6) exempts those same providers from appointing a representative in the Union, which the Act requires of providers established outside it [714], with the same limit [612].
The General-Purpose AI Code of Practice repeats that boundary. Its documentation measures do not apply to open models that meet Article 53(2), except those with systemic risk. It asks open models to warn in their documentation that the model must not be used to infringe copyright [715]. According to the European Commission’s guidelines, which are not binding, someone who fine-tunes another party’s model becomes its provider, and bears the Act’s obligations for that model, if the fine-tuning uses more than a third of the compute with which the original was trained. The Commission gives that third as an indicative criterion. Someone who fine-tunes with less does not become a provider. According to the same guidelines, models with systemic risk meet all obligations, whether open or not [716].
The exemption is lost in three cases:
- Systemic risk. Systemic risk is presumed when training exceeds 10^25 operations, ten times less than the 10^26 threshold used by the laws of California and New York, so the European Union reaches more models. The provider can rebut the presumption if it shows that, because of its characteristics, its model does not present those risks. The Commission can also designate models below that threshold. When there is systemic risk, openness exempts from nothing [612]. The provider must then meet both documentation obligations and, in addition, evaluate the model with adversarial testing, in which a team tries to make the model do what it should not in order to uncover its dangerous capabilities. It must also assess and mitigate systemic risks, report serious incidents to the AI Office and give the model adequate cybersecurity protection [714].
- Copyright. The exemption never covers the summary of the content used for training or the copyright compliance policy, as recital 104 makes clear; recitals are the paragraphs that precede the articles of the Act and explain how to interpret them [612].
- License and charging. According to the Commission’s guidelines, three licenses do not count as open: one that limits the model to non-commercial or research use, one that requires an additional license above a certain number of users (Llama’s is our example) and one that requires a separate commercial license for certain uses. Monetization also removes the exemption. There is monetization if the model is hosted only on the provider’s platform with paid access or with advertising, if access requires personal data for anything beyond security, or if the model is offered under a dual license that charges for commercial use or for use above a certain scale [716] [612]. By that standard, the licenses that require a separate contract from the large reseller, described above, hardly qualify for the exemption.
The Digital Omnibus of July 2026, which amended the regulation, did not touch these exemptions. It did postpone the obligations for high-risk systems, which the open-source exemption does not reach. Annex III of the Act lists the high-risk uses, among them biometrics, education, employment, access to essential services, law enforcement, migration and justice. Annex I covers AI that serves as a safety component in products already regulated, such as machinery, toys or medical devices [612]. The obligations for Annex III systems moved to December 2, 2027, and those for Annex I, to August 2, 2028 [488]. In Article 25(4), the amendment replaced the first paragraph and left the second intact, which is where the open-source exemption sits. It also gave the AI Office exclusive supervision of systems built on a general-purpose model from the same provider [488].
United States. State frontier laws make no distinction between open and closed models. California’s SB 53, the Transparency in Frontier Artificial Intelligence Act, signed in September 2025, requires every developer that trains models with more than 10^26 operations to publish a transparency report and to report critical safety incidents. Developers that also earn more than US$500 million must publish a safety framework, the document with the technical and organizational protocols they use to assess and mitigate catastrophic risks. Only they can be fined, up to US$1 million per violation. The law defines deploying as making a model available to a third party to use, modify or copy, so publishing the weights is just another deployment [99].
On weights, the large developer must explain in its framework how it protects unreleased weights against unauthorized modification or transfer. Every frontier developer, for its part, must report within 15 days any unauthorized access to the weights, their unauthorized modification or their exfiltration (the removal of the weights without authorization), when any of the three causes death or injury [99]. No later than January 1, 2027, and every year after that, the California Department of Technology must review the definitions of frontier model and frontier developer. It must also propose changes to the legislature, with input from the open-source community, among others [99].
Two other states followed California. New York’s RAISE Act, in its original text, covered models distilled from a frontier model. The March 2026 amendment aligned it with California’s law and dropped that part [100] [101]. Illinois followed the same definitions in its SB 315, the Artificial Intelligence Safety Measures Act, which requires the safety framework and an annual independent audit from 2028 [717]. Both obligations fall on the large developer, the one that earns more than US$500 million a year. From January 1, 2027, when the law takes effect, that developer cannot develop, deploy or operate a frontier model in the state without a current declaration filed with the state emergency management agency, which it must renew every year [718].
At the federal level, the early access under Executive Order 14409, through which the government receives certain frontier models up to 30 days before they are delivered to other partners, is voluntary (see “The United States sponsors both sides”) [339]. On November 10, 2026, Josh Gottheimer, a Democrat, and Mike Lawler, a Republican, introduced the American AI Security Act (H.R. 10981), which would make it mandatory. The developer of a covered foundation model would have to give the National Security Agency (NSA) full access to the model before deploying it. It could deploy it only after applying the NSA’s recommendations, or if the NSA did not evaluate it within 30 days, extendable by another 30. The Department of Commerce would set which models are covered, with a minimum number of parameters and by cyber, chemical, biological or radiological risk capabilities [719].
The text of the bill also addresses the publication of models. The NSA would evaluate, among other things, the capability that may create a serious cybersecurity threat “if the developer made the covered foundation model publicly available.” In addition, the bill defines deploying as delivering the model, or giving access to it, to anyone outside the developer, so publishing the weights is also deploying. The fine would be US$500,000 per day of noncompliance, and a special court would review appeals [719]. By its text, if the bill were passed, a covered open model could not be published without first going through the NSA, because publishing the weights is deploying. The parameter threshold set by the Department of Commerce would decide how many open models it reaches.
China. The Interim Measures for the Management of Generative Artificial Intelligence Services, in force since August 2023, apply to services offered to the public in China. They expressly exclude anyone who researches or develops without providing the service to the public [574]. The security assessment and algorithm filing are required of services with public-opinion attributes or social-mobilization capacity, that is, those that can influence what the public thinks or bring people together [574]. Filing is governed by the Provisions on the Management of Algorithmic Recommendations in Internet Information Services, in force since March 2022. The provider must register with the Cyberspace Administration of China’s system, within ten working days of the start of the service, its name, the type of algorithm and a self-assessment report [720]. The measures do not address weights published for download, and we found no official guidance that settles the question. By its text, the filing covers whoever serves the model to the public, not whoever publishes the weights. That is how the analyst Tony Peng read it when the measures were published [721]. China thus watches the service, where it can require censorship and filing, and leaves publication free, which is the route by which its models reach the world.
Other countries. Brazil’s AI bill (PL 2338/2023), which the Federal Senate passed in December 2024 and which awaits a committee report in the Chamber of Deputies, does not exempt open models by law. The developer of a general-purpose system with systemic risk meets its requirements even if it offers the system “under free and open-source licenses” (art. 30, § 1). The authority may grant a simplified regime to open and free standards and formats, except high-risk and general-purpose ones (art. 73) [722]. The general AI laws of South Korea, Japan and Italy make no distinction between open and closed models [723].
Liability for harm
The European Product Liability Directive, Directive (EU) 2024/2853, gives a right to compensation to any natural person who suffers harm caused by a defective product. The victim must prove the defect, the harm and the causal link between them, but not the manufacturer’s fault. The Directive covers death, personal injury, damage to property not used for professional purposes and the loss of data [490]. It counts software as a product. It does not apply to free and open-source software developed or supplied outside a commercial activity [724]. The Directive considers a manufacturer, with strict liability for defects, that is, without the victim having to prove fault, anyone who substantially modifies a product outside the manufacturer’s control and places it on the market. It also makes the manufacturer of a component liable [724]. Fine-tuning a model can count as a substantial modification, as Philipp Hacker and Matthias Holweg argue [725].
Under those rules, the developer that publishes weights for free is left outside the Directive. Liability falls on whoever builds the model into a commercial product, or substantially modifies it and sells it. The developer that charges for its model can be liable as the manufacturer of a component. The Directive applies, through the national laws that transpose it, to products placed on the market after December 9, 2026. Transposing a directive means incorporating it into the law of each member state, and the deadline to do so falls on that same day [724]. As of October 30, 2026, only Hungary, Croatia and Lithuania had fully transposed it [726].
The United States has no equivalent rule. SB 53 is enforced through civil penalties sought by the attorney general, and it creates no liability for damages [99]. SB 1047, the Safe and Secure Innovation for Frontier Artificial Intelligence Models Act, which Gavin Newsom vetoed in 2024, made the original developer liable for copies fine-tuned with little compute, and made anyone who fine-tuned with a lot of compute a developer. The line between the two cases was 3 × 10^25 fine-tuning operations, three times the compute at which the European Union presumes systemic risk in a full training run. So almost all fine-tuning would have remained the original developer’s responsibility. It also limited the shutdown obligation, the obligation to be able to stop the model completely, to the copies the developer still controlled [727].
Section 230 of title 47 of the U.S. Code bars treating an online service as the publisher of information provided by a third party, so a platform is not liable for what its users post [728]. No appeals court has decided whether Section 230 protection reaches what an AI generates. The U.S. District Court for the Northern District of California, in two cases, treated it as a question of fact, decided case by case according to its circumstances rather than with a general rule [729] [730]. A developer cannot know in advance whether its model is protected. The cases are Forrest v. Meta, from 2024, and Bouck v. Meta, from March 2026. In both, the court distinguished between publishing a third party’s material and contributing to its content enough to lose the protection, without setting a rule [730].
In the absence of a rule specific to AI, California Attorney General Rob Bonta sued OpenAI on November 16, 2026, in San Francisco Superior Court. He relied on the Unfair Competition Law and on the doctrine of public nuisance. California’s Unfair Competition Law prohibits unlawful, unfair or fraudulent business acts. The state’s Civil Code defines a public nuisance as one that affects at the same time an entire community or a considerable number of persons, and it allows the attorney general to seek to have it stopped [287]. The complaint alleges that in July 2026 hundreds of agents running on experimental versions of OpenAI’s models got out onto the public internet and attacked Hugging Face’s systems for about two and a half days. It alleges that this happened even though OpenAI claimed those tests were conducted under strict control and constant supervision. It also alleges that the company acknowledged to the attorney general more than 80 losses of control of its agents during testing, that it has publicly reported only 27 and that it did not hand over the details of all the incidents [287].
The attorney general seeks an injunction to stop the conduct, restitution and disgorgement, abatement of the nuisance and penalties of US$2,500 for each violation proven [287]. OpenAI responded that the lawsuit singles it out alone for a problem that all the major labs have identified. It also responded that it published a technical report, commissioned an independent one, notified those affected and strengthened its safeguards [731].
Restrictions on the rival’s model
In the United States, the only restrictions in force on open models fall on Chinese models. They reach their use by the government, not their download:
- The National Defense Authorization Act for Fiscal Year 2026 orders the exclusion and removal from U.S. Department of Defense systems of AI developed by DeepSeek or by High Flyer, the fund that created it. It also bars the Department of Defense’s contractors from using it in their contracts. It also orders the removal of the DeepSeek app from the national security systems of the intelligence agencies [732].
- The same act, in its Section 1532, empowers the U.S. Department of Defense to extend the exclusion to other AI companies, also for its contractors. It can extend it, if it judges that the company poses a national security risk, to any company that is on the Consolidated Screening List, which the International Trade Administration of the Department of Commerce compiles from the export restriction lists of the U.S. Departments of Commerce, State and the Treasury, or on the list of Chinese military companies operating in the United States that the Department of Defense publishes every year; that is domiciled in North Korea, China, Russia or Iran; or that is owned, controlled or influenced by one of those countries without mitigating measures [732] [346]. Since June 8, 2026, the Department of Defense list includes Alibaba, which develops Qwen, and Baidu, which join Tencent. The Department of Defense increasingly uses that list to restrict its contracts with those companies, although being on it has few immediate legal consequences [733].
- A July 2026 order of the U.S. Department of Homeland Security excludes DeepSeek products from civilian agency procurement. It also orders the removal of third-party products “developed using DeepSeek implementations,” a term it does not define [734]. The order was issued under the Federal Acquisition Supply Chain Security Act of 2018. That law created an interagency council, the Federal Acquisition Security Council, that can recommend excluding risky products. The orders are issued by the Department of Homeland Security for civilian agencies, the Department of Defense for its own and the Office of the Director of National Intelligence for the intelligence agencies [735] [736]. According to the law firm Crowell & Moring, the order reaches every product that integrates a DeepSeek model, wherever it is hosted and even if it is a copy of the open weights installed locally [736]. Haynes Boone warns that the order does not distinguish between DeepSeek’s service and the copy the user hosts, and that it does not address fine-tuned versions either [737]. The order limits federal procurement, not use outside the government [735].
- The Department of Defense and the Office of the Director of National Intelligence issued their own orders. Those orders reach applications and services that implement a DeepSeek model, wherever they are hosted [735] [736].
- Virginia banned the DeepSeek app on state government devices and networks. According to the press, Texas and New York, among others, did the same. These are bans on the app, not on weights run on one’s own hardware [738].
- Zhipu, the company behind Z.ai, has been on the Department of Commerce’s Entity List since January 2025. That restricts exports to Zhipu, not the use of GLM, Z.ai’s family of models [739]. We found no indication that the Department of Commerce added any Chinese lab to the Entity List in 2026. The U.S. Department of Defense did add Alibaba to its own list. The joint advisory of the NSA, the FBI and CISA of September 8, 2026, names DeepSeek, Moonshot, Alibaba, MiniMax, StepFun and Z.ai for distillation practices, but it does not restrict the use of their models [740].
Pending bills go further. Josh Hawley’s S. 321, the Decoupling America’s Artificial Intelligence Capabilities from China Act of 2025, would ban importing into the United States AI technology or intellectual property developed in China and, by its broad definition, would reach weights [741]. It would also ban exporting that technology to China, and both bans would apply 180 days after the law was passed [741]. The China FIREWALL Act, H.R. 10670, would bar agencies from buying any software that uses a foundation model developed or provided by a “foreign entity of concern,” as defined by the Surface Transportation Reauthorization Act of 2021, and it does not exempt local execution [742].
That definition includes, among other cases, any entity owned by the government of one of the four countries named in defense legislation, which are North Korea, China, Russia and Iran, or under its control or jurisdiction [743] [346]. The bill would also order the removal of those applications from government devices, with exceptions for national security and security research [742]. The No Adversarial AI Act, introduced in both chambers, would require each agency to consider excluding AI from those same four countries, which it calls foreign adversaries [744]. It would task the interagency council created by the 2018 supply chain security law with a public list of AI produced or developed by a foreign adversary, which agencies would have to exclude and remove, with exceptions such as research or testing [744]. None of the three has made it out of committee.
The Remote Access Security Act, which the U.S. House of Representatives passed on January 12, 2026, deals with remotely rented compute, not with models [745]. The bill would amend the Export Control Reform Act of 2018 so that the government can control remote access by a foreign person, over the internet or through a cloud service, to items subject to that law [549]. Today a company that cannot be sold a controlled machine can rent it through the cloud. The November 16 rule closes that gap only for advanced chips, and the bill would also close it for the rest, such as quantum computers.
Outside the United States, the measures against DeepSeek separate the service from the weights in the same way, but each country draws the line in a different place:
- The data protection authorities of Italy, South Korea and Berlin acted against the app and its data processing. In January 2025, Italy’s Data Protection Authority (Garante per la protezione dei dati personali) urgently restricted the processing of Italian users’ data [746]. In Korea, DeepSeek suspended downloads on its own. South Korea’s Personal Information Protection Commission recommended that it have a legal basis for transferring what users typed to Volcano Engine, ByteDance’s cloud, and that it delete what had already been transferred [747]. The Berlin Commissioner for Data Protection and Freedom of Information reported the app to Apple and Google as illegal content, because it transfers data to China [748].
- Australia banned DeepSeek products from federal government systems and devices, over an “unacceptable” level of security risk [749].
- Czechia warned those bound by its cybersecurity law against DeepSeek products on July 10, 2025, and rated the threat as high. It expressly excluded open-source models installed on one’s own hardware without communication with the company’s servers. The government also ordered the central administration to stop using those products within 30 days [750].
- Taiwan banned DeepSeek for public entities in all its forms, including local download, that is, the open weights. It did not restrict private use. Public universities and research centers may use it with authorization [751].
In the European Union, the European Data Protection Board broadened its ChatGPT task force. The Board brings together the national data protection authorities and ensures that the European data protection regulation is applied consistently [752]. The task force now deals with the enforcement of data rules on AI in general, DeepSeek included [753]. We found no formal measure by the European Commission.
The measures in the United States and Taiwan reach the weights, Czechia’s leave them free and those of Italy, Korea and Berlin target the service. An open model run on one’s own hardware transfers data to no one. What remains is the risk of the model itself, with its biases, its censorship and whatever a third party may have inserted into it.
Synthetic content
The rules on synthetic content were written for the service, and the open-weight model falls almost entirely outside them. Marking content can only be required of whoever controls the system that generates it. Whoever runs an open model on their own hardware can leave the mark off, or can remove it.
European Union. Article 50(2) of the EU AI Act requires providers of systems that generate images, audio, video or text to mark that content so that a machine can recognize it as artificial, to the extent technically feasible. The open-source exemption in Article 2(12) does not reach that obligation [612]. The Commission’s guidelines, dated November 5, 2026, specify that whoever publishes open components, models included, that are not by themselves an AI system is not subject to Article 50. They also specify that the article does not expressly apply to general-purpose models [754].
The code of practice of June 10, 2026, calls for two layers of marking. One is digitally signed metadata, data attached to the file that declare it artificial and carry a signature with which anyone can verify that no one altered them. The other is an imperceptible watermark, an alteration embedded in the image, the audio or the text. For systems and models under an open license it is enough to warn in their documentation that the mark must not be removed, and the signature is not required when the system runs on the user’s device [755]. For systems already on the market before August 2, 2026, the deadline for marking falls on December 2, 2026 [488].
From that same day, a prohibition introduced by the July amendment applies. It bans placing on the market or using systems that generate realistic images, videos or audio of the intimate parts or sexual activity of an identifiable person without their consent. It also bans those that generate child sexual abuse material. The prohibition reaches the provider in two cases: when that is the system’s purpose, or when the result is foreseeable and reproducible without significant technical modification and the system lacks reasonable safeguards to prevent it. Those safeguards are judged by the system’s “release and distribution strategy (such as open-source releases),” as recital 12 puts it [488]. We believe the rule reaches an open system published for that purpose or without safeguards. We also believe it leaves out a base model that produces that material only after fine-tuning, and that it makes whoever does the fine-tuning the provider of a prohibited system.
United States. The TAKE IT DOWN Act, of May 2025, punishes publishing forged intimate images. It also requires platforms that host user content to remove them within 48 hours of receiving a notice. It imposes no duties on whoever develops the model [756]. The Federal Trade Commission (FTC) began enforcing it on May 19, 2026 [757]. California goes further with open models:
- AB 853, of 2025, calls a hosting platform any platform that allows downloading the code or the weights of a generative system. Starting January 1, 2027, it bars such a platform from knowingly offering a system that does not mark its content [98].
- SB 1000, of September 2026, extended the duty to mark to every provider of a generative system accessible to the public in the state. It thereby eliminated the threshold of one million users in the California AI Transparency Act, the 2024 law it amended [758]. According to Latham & Watkins, it thus probably also reaches whoever develops open-source models. Whoever licenses the system to another party and knows that the other party removed the mark must revoke the license or notify it within 72 hours [759] [760].
United Kingdom. Since June 29, 2026, the Crime and Policing Act punishes with up to three years in prison anyone who makes, adapts or supplies a tool designed to generate fake intimate images [761]. The government factsheet says the law captures those tools “in all their forms,” including “AI models” [762]. The law leaves out the general model and reaches a model fine-tuned to undress people that someone uploads to a repository. The factsheet limits the offense to tools made to generate those images [762]. Lorna Woods, for her part, reads the parallel offense on child abuse the same way, and it reaches fine-tuned versions of commercial models [762] [763].
China. The Measures for Labeling AI-Generated Synthetic Content, in force since September 2025, require whoever provides the service to add a visible label and another in the metadata. They also bar any organization or person from removing those labels or providing tools to do so [764]. Whoever publishes weights without providing the service is reached only by that prohibition.
The harm comes from both sides. According to the Internet Watch Foundation, an independent nonprofit organization that finds and removes child sexual abuse images from the internet [765], open-source systems “remain the tool of choice” for those who generate child sexual abuse material, because they run on one’s own hardware, without moderation or detection [766]. The most visible case of the year, however, came from a closed service. According to the Center for Countering Digital Hate, Grok, from xAI, generated about 3 million sexualized images in 11 days in January 2026, about 23,000 of them apparently of children [767]. On January 26, 2026, the European Commission opened proceedings against X over those Grok features, under the Digital Services Act [768]. That law requires very large platforms to assess and mitigate the systemic risks of their service. The Commission can fine them up to 6% of their annual worldwide turnover [769].
With closed models, the regulator has someone to hold accountable. With open models, control passes to whoever fine-tunes the model, whoever hosts it and whoever spreads the content. Kapoor, Bommasani and others argue that preventing an open model from circulating is not feasible, while monitoring the distribution of models fine-tuned for illicit purposes is [252].
Competition and concentration
Competition authorities count open models in favor of competition, but none has used them as a criterion to decide a case. The U.K. Competition and Markets Authority wrote in April 2024 that it would be concerned if the market did not sustain a range of foundation models, “including open-source models” [770]. In July 2024, the FTC’s Office of Technology wrote that open-weight models can “drive innovation, reduce costs, increase consumer choice.” It also warned of the risk of the company that opens first and closes later, that is, one that offers its product openly to win market share and then closes it. According to the text, that strategy allowed companies in earlier generations of technology to dominate the market and exclude rivals [771]. According to TechCrunch, the FTC took that text down in 2025, by then chaired by Andrew Ferguson, whom Donald Trump designated chairman of the commission in January of that year [772].
The FTC’s January 2025 study of the partnerships of Microsoft, Amazon and Google with OpenAI and Anthropic found that none of those companies’ main models had open weights [773]. The July 2024 joint statement by the authorities of the United States, the United Kingdom and the European Union does not mention open models [774]. The U.S. government’s action plan defends them so that startups do not depend on a closed-model provider [30].
Corporate spending on models is concentrated in two closed labs, while usage is more spread out. The Menlo Ventures survey, published in November 2026, polled almost 400 AI purchasing decision makers at U.S. companies. In it, Anthropic takes 55% of enterprise spending on model access and OpenAI, 31%. Open models account for 19% of usage and 4% of spending, because their tokens cost on average US$0.13 per million, against US$0.70 for closed ones. Chinese models take 89% of what is spent on open models. Menlo is an investor in Anthropic [66]. For now, open models serve more to lower the prices of closed ones than to take customers from them. Menlo notes that, as open-model usage grows, Anthropic and OpenAI have begun discounting even their flagship models [66].
The large partnerships tie the closed labs to a cloud. In October 2025, when OpenAI was reorganized, Microsoft ended up with about 27% of the company [582]. Its July 2026 annual report gives about 25% and its report for the quarter through September, about 24% [775] [583]. According to the annual report, the stake fell because of the reorganization and other OpenAI funding rounds. In each round investors come in with new shares, and the percentage held by earlier partners shrinks even if they sell nothing [775]. OpenAI, in turn, committed to buy an additional US$250 billion of Azure services. The same agreement allows OpenAI to publish open-weight models that meet certain capability criteria [582].
The acquisition of Hugging Face, the main distribution site for open models, has no confirmed formal review. According to Reuters, the U.S. Department of Justice asked whether the deal would let NVIDIA “capture a fundamental distribution point for AI” [372] [776]. It asked within an antitrust investigation of NVIDIA’s practices that it already had open. That investigation is separate from the formal review of an acquisition, which can approve it, condition it or block it. The acquisition was announced seven weeks after OpenAI agents under evaluation got into Hugging Face’s production infrastructure. The public affairs study covers that case [777], over which California sued OpenAI on November 16 (see “Liability for harm”). The specialized press considers it likely that the European Commission will review the acquisition, although there is no confirmed notification [778].
In the acquisition announcement, NVIDIA committed to keeping the platform open in line with Hugging Face’s practices and to supporting other chipmakers. Jensen Huang, then NVIDIA’s chief executive, also wrote on the company blog that using the platform will not require NVIDIA compute [24] [25]. We believe public policy has to decide whether those commitments remain a company promise or become conditions of the clearance.
The only 2026 measure on the distribution of AI assistants fell on Meta, the company that most champions open models. In October 2025, Meta banned other companies’ AI assistants from the WhatsApp Business API, the interface companies use to connect their systems to WhatsApp. In March 2026 it readmitted them with a fee that, according to the European Commission, is at first sight equivalent in practice to the ban. On June 9, the Commission imposed interim measures on it under competition law and ordered it to restore access on the previous terms [779]. The Commission can impose interim measures when there is a prima facie infringement and it is urgent to prevent serious and irreparable harm to competition. The order applies until the final decision in the case. It is the second time the Commission has used that power under the current competition regulation [779]. We believe that opening the model does not prevent closing the channel through which it reaches the user.
Adam Schiff and Jim Banks introduced the Collaboration on Adversarial Threats and Security Risks Act, S. 5105, in July 2026. The bill would give companies that share information on AI security risks, including theft and distillation by adversary countries, a defense against antitrust laws. The same defense would cover companies that agree to delay or limit the launch of a model to reduce those risks, in this case with prior notice to the Department of Justice. The bill preserves private lawsuits and allows the attorney general to seek to halt conduct that increases risk [780]. If it passes, the labs could agree among themselves not to publish certain weights and would have a defense against a collusion charge, provided they prove its security purpose. The bill has been in the Senate Committee on the Judiciary since July 23.
Who regulates
United States. The underlying dispute is between the federal government and the states, and whichever rule prevails weighs differently on each model. The U.S. Constitution makes federal laws the supreme law of the land, above the laws of any state [781]. That is why a federal law can preempt a state law on the same subject, that is, render it void. In July 2025, the U.S. Senate voted 99 to 1 to strip from the budget bill its moratorium on state AI laws [782]. The moratorium would have barred states from enforcing their AI laws for ten years [782].
Executive Order 14365, of December 2025, set as policy a minimally burdensome national framework for AI. It created within the Department of Justice a task force whose “sole responsibility” is to challenge state AI laws. The task force can argue that those laws encroach on interstate commerce, which the Constitution reserves to Congress, or that a federal rule already preempts them. The order conditioned federal broadband funds on states not having “onerous” laws. The funds are those of the BEAD program, and which laws are onerous is defined by an evaluation the order assigned to the Department of Commerce. It also assigned the FTC a policy statement on the subject. The statement must explain when the federal prohibition of deceptive practices preempts a state law that requires altering a model’s truthful outputs. The order also called for a proposal for a federal law that preempts state laws contrary to its policy, without preempting those on child protection, on compute infrastructure and data centers, or on the states’ own procurement and use of AI [783]. In March 2026, the White House asked Congress to bar states from regulating AI development and from penalizing a developer for what a third party illegally does with its model [93].
The rest is still in progress:
- In April 2026, the Department of Justice intervened in xAI’s lawsuit against Colorado’s law on algorithmic discrimination, and Colorado agreed not to enforce it while the judge decides [784] [785]. That law, SB 24-205, the Colorado Artificial Intelligence Act, requires whoever develops a high-risk AI system to use reasonable care to protect consumers from algorithmic discrimination [786]. The Department of Justice argues that it violates the Equal Protection Clause of the Fourteenth Amendment, which requires treating everyone equally before the law, because it forces companies to keep their products from disproportionately harming a race or a sex even when they do not intend to, and it exempts certain forms of discrimination meant to promote diversity [784]. In May, Colorado replaced it with another law, which regulates automated decisions on employment, housing, credit, education, insurance or health care. The new law requires documentation from whoever develops those decision systems, not from whoever trains general-purpose models [787]. Starting January 1, 2027, whoever develops one of those systems must give its users technical documentation on its intended uses, its training data and its known limitations. The law is enforced by the state attorney general, and it does not give private parties a new right to sue [787].
- In June 2026, Representatives Jay Obernolte and Lori Trahan released the discussion draft of the Great American AI Act. The draft would bar for three years any state law that specifically regulates the development of a model, and it would leave in force those that regulate its distribution and use. It defines deploying as making a model available to use, copy or combine with other software, and it makes anyone who fine-tunes another party’s weights a developer. It has not been introduced [97]. Some 200 state legislators signed a letter against it [788].
- On July 1, 2026, the FTC proposed a policy statement on the suppression of accuracy in AI systems. According to law firms, the statement holds that a state law that requires altering truthful outputs is preempted by federal law, and it cites Colorado’s as an example. The text of the statement says that a company that alters its system’s outputs to comply with a state law may deceive the consumer and violate Section 5 of the FTC Act, which prohibits deceptive practices, unless it discloses this clearly and conspicuously. It also says that the law with which Colorado replaced its own in May raises many of the same concerns [789]. It has no final version [790].
The state frontier laws in force treat the publication of open weights as just another deployment and only require protecting unreleased weights, so they do not exempt open weights (see “Pre-deployment evaluation and transparency”).
The sides in this dispute do not match those of open and closed. In March 2025, Google asked to preempt “a chaotic patchwork” of state rules on frontier model development [89]. Andreessen Horowitz warned in February 2025 against “a patchwork of state laws” regulating development [90]. In June 2026, OpenAI proposed turning SB 53, the RAISE Act and the Illinois law into federal law, and then preempting state laws on the same risks [91]. Anthropic supported SB 53 [102]. In 2026 it donated US$40 million, in two contributions of US$20 million, to Public First Action, a group founded and run by Republican and Democratic strategists that opposes preempting state laws without stronger federal safeguards [791] [792]. Anthropic also maintains that Congress should not preempt them except with a federal law “at least as strong” as the framework it proposes [92]. The July 2026 letter in favor of open weights does not mention state laws [217]. OpenAI and Google signed the letter and call for preempting state laws. Anthropic did not sign it and accepts preemption with conditions.
Preempting state laws on development would favor whoever publishes weights only if it came together with the rule the White House requested on illegal use by third parties, which would shield the developer from the use it cannot control (see “The seven levels”). The draft of the Great American AI Act, from Representatives Obernolte and Trahan, is not enough. It leaves state laws on distribution in force and defines deploying as making a model available to copy. That is why publishing the weights would still be covered by those laws, as it already is by California’s AB 853 (see “Synthetic content”).
H.R. 10981 preempts only the state laws that require the developer of a covered model to be transparent about its safety protocols, its frameworks and its capability evaluations. In exchange it imposes the NSA’s pre-deployment evaluation [719]. That trade consists of a federal obligation in exchange for preempting state laws only where it regulates, and it is the one that can win votes from both parties. According to the law firm Covington & Burling, the new Democratic majority will not preempt state laws without enforceable federal rules to replace them [67]. We believe that, for whoever publishes weights, the trade swaps a state transparency law for prior federal authorization.
On November 3, 2026, Democrats won a majority in both chambers, according to the AP count [793]. According to Time, the Congress that convenes in January 2027 makes a law preempting state laws harder to pass [794]. According to Covington, the new majority will steer the discussion toward transparency, safety testing and incident reporting by frontier developers, although it will not have the 60 votes needed to overcome a minority filibuster in the Senate [67]. We believe that moves federal protection for open models further away and brings closer rules that, like the state ones, treat the publication of weights as just another deployment.
European Union. We believe the open ecosystem ends up divided among more supervisors than the closed one, because it is built mostly on other parties’ models. It is divided between the AI Office, which supervises the model, and the authorities of 27 countries, which supervise the systems. General-purpose models are always supervised by the AI Office. Since the July 2026 amendment, the AI Office also has exclusive supervision of systems built on a general-purpose model when the model and the system belong to the same provider or group, whether the model is open or closed. It also has exclusive supervision of those that operate within very large online platforms and search engines [488]. A system that integrates another provider’s open model stays with the national authority of each country.
China. Regulation is centralized and separates the service from publication. The Cyberspace Administration of China files and assesses services offered to the public. The Ministry of Commerce controls exports and leaves what has already been published outside its reach (see “Export controls” and “Pre-deployment evaluation and transparency”).
19
Balance
| Instrument | Open weights (download and self-hosting) | Closed model or service |
|---|---|---|
| U.S. export control | No control on weights from November 19, when the removal of classification 4E091 takes effect | No control on weights from the same date; control shifts to compute provided to companies from Macau and from Country Group D:5 countries, China among them |
| China’s export catalog (proposal) | Exempt if already public; opening it later requires an automatic license | Prior registration since October 16 and possible restriction above 10^26 operations, 2 trillion parameters or a place in the top ten |
| EU AI Act | Exempt from documentation and from appointing a representative, except systemic risk | All obligations |
| California SB 53 and New York RAISE | No distinction; publishing the weights counts as deploying | No distinction; the large developer must protect unreleased weights, and every developer reports their exfiltration |
| EU product liability | Out if there is no commercial activity; whoever substantially modifies it and sells it is liable as a manufacturer | The manufacturer is liable |
| U.S. federal policy | Support in the action plan and in Congress (H.R. 10152) | Contracts, voluntary early access and evaluations |
| U.S. government use | DeepSeek excluded from the U.S. Department of Defense, intelligence and civilian agencies, including in third-party products; the Department of Defense can extend the exclusion to other Chinese companies, and Alibaba, which makes Qwen, is already on its list of military companies | The DeepSeek exclusion also reaches its app; no U.S. model is restricted by origin |
| Service filing in China | Does not address publishing the weights; whoever serves the model to the public files | The service with public-opinion attributes files |
| Measures against DeepSeek outside the U.S. | Taiwan bans local download in the public sector; Czechia exempts it | Italy, Korea and Berlin act against the app and its data |
| Marking of synthetic content in the EU (Art. 50) | A model or component that is not a system is out; an open system only needs to warn in its documentation against removing the mark | Mandatory marking; on December 2 the deadline expires for systems predating August 2 |
| Non-consensual intimate images (EU from December 2; United Kingdom) | The base model is out; whoever fine-tunes it or supplies it for nudification is liable | The provider without reasonable safeguards is liable |
| Marking in California (AB 853 and SB 1000) | Every provider accessible in the state marks; from 2027 a platform hosting weights cannot offer a system without marking | Every provider accessible in the state marks |
| Competition | Authorities count it in favor of competition, without effect in any case; its main distribution site would pass to NVIDIA, with closing expected in 2027 | Two labs hold most enterprise spending, according to a survey by Menlo Ventures, an Anthropic investor; alliances with clouds without formal review |
| Rights over training data | In the U.S., the two rulings on generative models sided with the developer on training, one of them without establishing that training is lawful; the Anthropic ruling did not excuse piracy and the Meta ruling judged it together with training; in the EU, no exemption from the data summary or the copyright policy; publishing where the data came from hands evidence to the plaintiff | Same rules; it can pay for licenses and spread the cost across what it charges for usage |
| Preemption of state laws in the U.S. (proposal) | Free of state laws on development; it would remain subject to those on distribution, which reach the publication of weights; the shield against third-party use depends on the federal law that is passed | It would remain subject to state laws on the use of its service |
20
Calendar
| Date | Event |
|---|---|
| November 13, 2026 | The deadline expired for Huang to confirm in writing the December 1 interview [18] |
| November 15, 2026 | China closed the consultation on its catalog of technologies restricted for export [290]; John Moolenaar told CBS News that Huang agreed to a voluntary closed-door interview [192] |
| November 16, 2026 | The U.S. Department of Commerce announces the rule that removes the export control on model weights and requires a license to provide compute to companies from Macau and Country Group D:5 [22]; the U.S. Departments of Commerce and Justice report that the Department of Justice is leading the criminal investigation into Megaspeed [195]; California sues OpenAI over its agents’ losses of control in testing [287] |
| November 17, 2026 | NVIDIA reports results for the third quarter of its fiscal 2027, the first without Huang; it announces it will stop selling the H200 in China and creates a board committee for export controls [198] [200]; the committees report that Huang and NVIDIA’s general counsel will appear voluntarily on December 1 and 2 [19] |
| November 19, 2026 | The Department of Commerce rule is published and takes effect; those already providing that compute have 90 days to apply for a license [22] |
| November 28, 2026 | Deadline under Executive Order 14434 for the President’s science and technology adviser to propose a legal definition of superintelligence [258] |
| December 1, 2026 | Huang’s transcribed interview before the three committees of the U.S. House of Representatives, which he accepted voluntarily [19] |
| December 2, 2026 | Interview of Tim Teter, NVIDIA’s general counsel; the marking obligations of Article 50(2) of the EU AI Act begin for generative systems predating August 2, as does the ban on systems that generate non-consensual intimate images or child sexual abuse material [18] [488] |
| December 3, 2026 | Hearing of Jeff Kessler, Under Secretary for Industry and Security at the U.S. Department of Commerce [18] |
| December 9, 2026 | Deadline to transpose the new European product liability directive, which applies to products placed on the market after that day [724] |
| January 1, 2027 | New York’s RAISE Act takes effect, with the March amendments, as does Illinois’s SB 315; in California, platforms hosting weights can no longer offer systems that do not mark their content, and the first annual review of SB 53’s frontier definitions by the California Department of Technology is due [101] [717] [98] [99] |
| January 3, 2027 | The 119th Congress ends and bills not passed expire, among them H.R. 10152, approved in committee on September 16 and awaiting its report and a floor vote, and H.R. 10981; the 120th Congress convenes, with a Democratic majority in both chambers [26] [719] [793] |
| January 10, 2027 | The extended suspension of the rule that extends controls to affiliates of Entity List companies, part of the trade truce with China, expires [437] |
| January 18, 2027 | The 60-day period to comment on the Department of Commerce rule of November 16 closes [22] |
| February 17, 2027 | The 90 days to apply for a compute license expire [22] |
| First half of 2027 | Expected closing of NVIDIA’s purchase of Hugging Face [24] |
| December 2, 2027 | Obligations for the high-risk systems of Annex III of the EU AI Act begin [488] |
| January 1, 2028 | In Illinois, the mandatory safety framework and independent audits begin [717] |
| August 2, 2028 | Obligations for the high-risk systems of Annex I of the EU AI Act begin [488] |
21
Implications
Depending on DeepSeek is already a procurement risk. A company building on DeepSeek is shut out of contracts with U.S. civilian agencies and the U.S. Department of Defense (see “Restrictions on the rival’s model”) [734] [735]. The defense authorization act already allows the Department of Defense to extend that exclusion to other Chinese AI companies. Alibaba, which makes Qwen, has been on the list of military companies that triggers that power since June [732] [733]. Pending bills would extend the exclusion to any software built on a Chinese model, although none has made it out of committee and all expire with the Congress, which ends on January 3, 2027 (see “Calendar”) [742]. Chinese models are also the most downloaded on Hugging Face, with 41% of downloads in the year before March 2026 [28], and, according to the public affairs study, the cheapest [304]. Anyone selling to the U.S. government must be able to switch models without rebuilding its product, because the list of excluded models can grow. The Harvard Business School AI Institute already recommends sourcing models like any critical input, one that governments can control and that is best bought from more than one supplier [795].
Whoever installs an open model takes charge of its safeguards. A public attack that breaks an open model’s safeguards keeps working on every copy, because published weights cannot be fixed afterward [255] [256]. A company deploying an open frontier model must add its own filters and its own monitoring, because no one else will. Meta recommends that those deploying Llama review and filter everything that goes into the model and everything that comes out of it [796].
Liability shifts toward the integrator. In the European Union, the company that substantially modifies an open model and sells it in a product is liable as a manufacturer, even if the original developer who published it for free is not (see “Liability for harm”). The rule applies to products placed on the market after December 9 and depends on the national laws that transpose the directive [724]. Fine-tuning someone else’s model stops being neutral for liability purposes, as Philipp Hacker and Matthias Holweg argue [725].
Open licenses come with conditions. Those of Kimi K3, Qwen3.8 and Muse Spark Open require a contract from the large service provider above a certain revenue [238] [240] [243]. Mistral Large 4’s requires that contract from any model-as-a-service business [244]. GLM-5.3’s reserves a security review for the Chinese developer [241].
The open coalition in the United States changes spokesperson. Huang said he co-wrote the July letter [25], and NVIDIA is, according to Hugging Face, the largest contributor of open models [23]. The company today has an interim chief executive, an investigation under way and appearances in December. We believe that, for those reasons, it comes to the debate on H.R. 10152 and the distillation bills with less capacity to speak publicly [16] [18]. We also believe that Microsoft, which published the letter, and Meta are the signatories with the most weight to take that place [306] [307].
Distribution of open models may end up in the hands of a chip supplier. Hugging Face is the site where most open models are published, and NVIDIA promised to keep it open and neutral toward other chips [24] [25]. So far that promise is unilateral (see “Competition and concentration”). The U.S. Department of Justice has only asked questions within a conduct investigation that was already open, without a formal review of the merger [372]. We believe that what the authority reviewing the merger decides, namely whether or not it turns that promise into a condition, weighs heavily for open models. We also believe that this decision does not determine all of their distribution, because clouds and other platforms, such as Alibaba’s ModelScope, also distribute them.
Content marking does not reach the model that runs at home. European, Chinese and U.S. rules place marking on whoever provides the service, and the European code acknowledges that signing cannot be done on a local device (see “Synthetic content”) [755] [754]. California and the United Kingdom already target whoever hosts or supplies the weights [98] [762]. For open models, the path will be to monitor hosting and fine-tuning for unlawful purposes, rather than the base model, as Kapoor, Bommasani and others propose [252].
The cost of data depends on where it is litigated. In the United States, two judges found training a generative model to be fair use. One of them, in the Anthropic case, did not excuse the pirated copies; the other, in the Meta case, judged the download together with the training. The U.S. Court of Appeals for the Third Circuit, by contrast, denied fair use to a non-generative AI and measured the harm in the licensing market [603] [601] [604]. In Germany, a court found copies in the weights [611]. We believe that a company that trains or fine-tunes a model, open or closed, should document where it got the data and whether it bought them. We also believe that an open model that publishes that list is more exposed.
In Europe, size erases the advantage of being open; in China, it does not. The European exemption for open models ends where systemic risk is presumed, above 10^25 training operations. From that threshold on, open models meet the same obligations as closed ones, among them documenting the model, evaluating it, mitigating its risks and reporting serious incidents (see “Pre-deployment evaluation and transparency”) [612]. The catalog proposed in China restricts exports of models with 10^26 operations, 2 trillion parameters or more, or capability among the world’s top ten. Those already public remain free. A model of that size opened later will have to be declared, with an automatic license (see “Export controls”) [290] [713]. Qwen3.8 and Kimi K3 already cross the parameter threshold, but they remain free because they are already public [240] [238].
Prices fall for what already exists, not for the frontier. The cost of reaching a fixed capability falls by a factor of tens each year, but what is paid to use the most capable model rises, and in December Google will double the price of two of its models [635] [636] [639]. A company budgeting its AI use should separate the tasks an open or older model already handles from those that need the frontier, because only the former get cheaper. The Harvard Business School AI Institute recommends sending routine tasks to cheap models [795].
Software that charges per user loses pricing power. Microsoft said its business model is moving from per-user subscriptions to also include consumption-based billing [61], and in 2026 the stock market punished business application companies [62]. We believe that software buyers today have more room to negotiate, and that sellers are under more pressure to charge for outcomes or usage.
Whoever integrates third-party models inherits its clients’ clauses. Large companies prohibit their suppliers from training on, retaining or using their data without written permission [64] [65] [682]. Some frontier models, for their part, require retaining conversations for review, unless an exemption is negotiated [115]. An integrator needs to be able to run in the client’s cloud and to switch to an installable open model when the data does not allow retention. For regulated or sensitive data, the Harvard Business School AI Institute recommends running models on one’s own servers [795].
22
About this document
The regulatory and legal research drew on the Psiquepol + AI corpus of laws, bills and monitoring, the platform we use to track regulation and its actors in each country. That corpus provided the texts of the EU AI Act, the product liability directive, the code of practice, the California, New York and Illinois laws, the executive orders, the export regulations and the congressional bills we cite. For search, source cross-checking and drafting, we used artificial intelligence models.
This is a document of public policy analysis and of the artificial intelligence landscape, intended as input for decision-making. It does not constitute legal, political or financial advice.
23
AI Monitoring
Psiquepol + AI’s AI Monitoring brings together the laws, bills and debates cited in this study, with a link to each source and to the profile of each actor, institution and article. Outside our platform, some official sources restrict access depending on the country or the type of connection. AI Monitoring updates that material continuously, by country or with global coverage. It is available immediately in countries with active monitoring. In other countries, it is activated within 72 hours or your money is refunded.
Примітки та джерела
- 1.Yotam Berger, “Microsoft's Crackdown on Unit 8200 Reveals Tech's Intermediary Role”, Lawfare, October 8, 2025.↩
- 2.Peter Henderson and Mark A. Lemley, “The Mirage of Artificial Intelligence Terms of Use Restrictions”, Indiana Law Journal, 2025.ab
- 3.George J. Stigler, “The Theory of Economic Regulation”, The Bell Journal of Economics and Management Science, vol. 2, no. 1, Spring 1971, pp. 3 to 21; the quotation is from p. 3.ab
- 4.Moritz Laurer, Andrea Renda and Timothy Yeung, CEPS, “Clarifying the costs for the EU's AI Act”, September 24, 2021, archived copy. The figures are about the European Commission's proposal.abc
- 5.Anthropic, “Updating restrictions of sales to unsupported regions”, September 4, 2025.ab
- 6.Anthropic, supported regions policy, effective November 12, 2026, accessed November 13. Mexico is among the supported regions.abcd
- 7.Brad Smith, Microsoft, “Update on ongoing Microsoft review”, September 25, 2025.abc
- 8.SK hynix, third-quarter 2026 results, October 28, 2026.abcdefg
- 9.Girish Sastry, Lennart Heim, Haydn Belfield, Markus Anderljung, Miles Brundage and others, “Computing Power and the Governance of Artificial Intelligence”, arXiv, February 2024.↩
- 10.ASML, 2025 annual report, strategic report section and financial performance section, February 24, 2026.abcd
- 11.NVIDIA, Form 10-K for fiscal year 2026, February 25, 2026. Percentages calculated from the figures by customer headquarters and by market; the phrase on China is “effectively foreclosed”.abcdefghijklm
- 12.AMD, Form 10-K for fiscal year 2025, February 4, 2026.ab
- 13.Microsoft, Maia 200 announcement, January 26, 2026.abc
- 14.NVIDIA, Form 10-Q for the second quarter of fiscal 2027, August 26, 2026.abcdefghijklmnopqrstuv
- 15.Lex Fridman Podcast No. 494, conversation with Jensen Huang, March 23, 2026, with the official transcript.abcdefghijklmnopqrs
- 16.NVIDIA, Form 8-K, Item 5.02, event of November 11, 2026, filed November 12.abcd
- 17.NVIDIA, “Statement from NVIDIA Board of Directors”, November 8, 2026.ab
- 18.Brian Mast, John Moolenaar and Rick Crawford, Committee on Foreign Affairs press release, with the letter to Jensen Huang, the letter to NVIDIA's general counsel and the invitation to Jeff Kessler, November 12, 2026.abcdefg
- 19.Brian Mast, John Moolenaar and Rick Crawford, “Jensen Huang and Nvidia Agree to Appear Before Congress”, Committee on Foreign Affairs press release, November 17, 2026.abcd
- 20.The New York Times, “Rubio Warned Officials About Nvidia Chief's Entanglement With Chinese Spy”, November 8, 2026, read in the DNYUZ reproduction of that day, because the original requires a subscription.ab
- 21.Ian Duncan, Zach Dorfman and Warren P. Strobel, The Washington Post, “Behind Jensen Huang's spy scandal, a shadow war for Trump's ear on China”, November 13, 2026. The full text requires a subscription; we read its opening, and the details of the memo come from AI Weekly and The Federal, which summarized it.ab
- 22.Bureau of Industry and Security, press release and interim final rule “Addressing Remote Access to Advanced Semiconductors”, RIN 0694-AK50, November 16, 2026. Document 2026-23529, filed for public inspection at the Federal Register on November 16, with publication and entry into force on November 19.abcdefghijk
- 23.Hugging Face, “State of Open Source on Hugging Face: Spring 2026”, March 17, 2026, before the acquisition agreement with NVIDIA.abcd
- 24.NVIDIA, Form 8-K on the acquisition of Hugging Face, September 2, 2026, filed on the 3rd, with the retention program.abcdefghij
- 25.Jensen Huang, NVIDIA blog post on the acquisition of Hugging Face, September 3, 2026.abcdefg
- 26.U.S. Congress, legislative status of H.R. 10152, updated September 29 and accessed November 12, 2026: ordered reported on September 16, no report filed.abc
- 27.Dylan Patel and others, SemiAnalysis, “Nvidia and the World Without Jensen”, November 12, 2026.abcdefghi
- 28.Hugging Face, “State of Open Source on Hugging Face: Spring 2026”, March 17, 2026.abc
- 29.Nathan Lambert and Florian Brand, “The ATOM Report”, arXiv, version of May 25, 2026.ab
- 30.White House, Winning the Race: America's AI Action Plan, July 2025, sections “Encourage Open-Source and Open-Weight AI” and “Ensure that Frontier AI Protects Free Speech and American Values”.abcdefg
- 31.Epoch AI, “Open models lag state-of-the-art closed models by 4 months”, May 29, 2026, with its model data, which record Meta's Muse Spark as API access since April 8, 2026.abc
- 32.UK AI Security Institute, Frontier AI Trends Report, December 2025.ab
- 33.UK AI Security Institute, “How far behind the frontier are leading open-weight models on cyber?”, July 17, 2026.abcd
- 34.National Institute of Standards and Technology, CAISI, evaluation of DeepSeek V4 Pro, May 1, 2026.abc
- 35.Stanford HAI, AI Index 2026, technical performance chapter, April 2026, with data through March.ab
- 36.Reuters, “Huawei expects AI chip revenue to jump at least 60% this year, FT reports”, May 1, 2026, published by The Star. Reuters could not verify the Financial Times report.ab
- 37.Reuters, “Nvidia's China AI-chip share falls as Huawei closes gap, report finds”, October 13, 2026, on an IDC report.ab
- 38.Reuters, “DeepSeek unveils new AI model tailored for Huawei chips as China pushes for tech autonomy”, April 24, 2026.abc
- 39.SemiAnalysis, “Huawei Ascend Production Ramp: Die Banks, TSMC Continued Production, HBM is The Bottleneck”, September 8, 2025.abcd
- 40.TechInsights, “Analysis: Can China even build data center AI chips?”, March 17, 2026.ab
- 41.TechNode, “Chinese makers accounted for more than 97% of global humanoid robot shipments in H1 2026”, August 11, 2026, with figures from Smart Analytics Global.abc
- 42.Federal Communications Commission, Public Safety and Homeland Security Bureau, public notice DA 26-786, July 28, 2026, and its Appendix C.abc
- 43.Hartmut Neven, Google, “Meet Willow, our state-of-the-art quantum chip”, December 9, 2024.abc
- 44.Google Quantum AI and collaborators, “Observation of constructive interference at the edge of quantum ergodicity”, Nature, October 22, 2025.ab
- 45.Executive Order 14412, “Securing the Nation Against Advanced Cryptographic Attacks”, June 22, 2026, published June 25, section 4(b).ab
- 46.Yunchao Liu, Srinivasan Arunachalam and Kristan Temme, “A rigorous and robust quantum speed-up in supervised machine learning”, Nature Physics, July 2021.ab
- 47.Joseph Bowles, Shahnawaz Ahmed and Maria Schuld, “Better than classical? The subtle art of benchmarking quantum machine learning models”, arXiv, March 2024. It is a preprint.ab
- 48.Arvind Narayanan and Sayash Kapoor, “AI as Normal Technology”, Knight First Amendment Institute, April 15, 2025.ab
- 49.Nathan Lambert, “The inevitable need for an open model consortium”, Interconnects, April 11, 2026.abcde
- 50.Pranay Kotasthane, Nitin Pai and Bharath Reddy, Takshashila Institution, “Why Is China Giving Away Its AI Models?”, discussion document 2026-25, August 7, 2026. It attributes China's openness to a commercial strategy and considers chip scarcity a secondary factor.abcde
- 51.Samsung Electronics, third-quarter 2026 results, October 29, 2026.abc
- 52.Micron, Form 10-K for fiscal year 2026, October 9, 2026.abcd
- 53.TSMC, third-quarter 2026 results, Form 6-K, October 15, 2026, with the revenue-by-platform slide.abc
- 54.Amazon, Form 10-Q for the third quarter of 2026. AWS margin and trailing twelve-month free cash flow calculated from its figures.abcd
- 55.Alphabet, Form 10-Q for the third quarter of 2026. Google Cloud margin calculated from its figures.abc
- 56.Anthropic, Form S-1/A, Amendment No. 2, November 10, 2026.abcdefghijklm
- 57.Zhipu, 2025 annual results, Hong Kong Stock Exchange, March 31, 2026.ab
- 58.MiniMax, 2025 annual report, Hong Kong Stock Exchange, April 22, 2026.ab
- 59.Dries Buytaert, “AI is a business model stress test”, January 9, 2026.ab
- 60.Microsoft, Form 10-K for fiscal year 2026. Revenue change calculated from its figures.abcde
- 61.Microsoft, transcript of the earnings call for the first quarter of fiscal 2027, October 28, 2026.abcd
- 62.Yahoo Finance, daily closing prices for IGV, MSFT, INTU, TRI, ORCL, NOW, ADBE and CRM, accessed November 13, 2026. Changes calculated by us.abcd
- 63.Adam Wathan, comment in the Tailwind documentation repository, January 7, 2026.ab
- 64.AT&T, “Supplier Artificial Intelligence Requirements”, version 2.1, October 2025.abcd
- 65.BASF, AI addendum for software and services suppliers, version 1.0, July 16, 2025.abcd
- 66.Menlo Ventures, “2026: The State of Enterprise AI”, November 3, 2026, Morning Consult survey of September 8 to 26. Anthropic is a company in its portfolio.abcdefg
- 67.Covington & Burling, “Technology Policy in a Democratic Congress”, November 9, 2026.abc
- 68.Anthropic, “GLM-5.3 and the spread of advanced cyber capabilities”, September 29, 2026. Anthropic competes with Z.ai.abcde
- 69.Federal Bureau of Investigation, FLASH-20261113-001, “Threat Overview of ARTEX Autonomous Penetration Testing Framework”, November 13, 2026.abc
- 70.U.S. Senate, hearing “Oversight of A.I.: Rules for Artificial Intelligence”, S. Hrg. 118-037, May 16, 2023, oral remarks by Sam Altman to Senators Kennedy and Ossoff.↩
- 71.Joshua Gans, “The thing about AI I am worried about: Competition”, May 17, 2023.↩
- 72.Psiquepol + AI, “The State of Artificial Intelligence Public Affairs Worldwide”, section “Ways to regulate”, note 551.↩
- 73.Access Now, AlgorithmWatch, EDRi, ECNL, ICCL and other organizations, open letter to the negotiators of the Council of Europe convention, March 5, 2024.↩
- 74.United States, 10 U.S.C. § 3252, definitions of supply chain risk and of the procurement actions it authorizes, in the reproduction of the Cornell University Legal Information Institute, accessed November 17, 2026; and 41 U.S.C. § 4713, the equivalent authority in federal procurement.↩
- 75.Council of Europe, member states, accessed November 17, 2026.↩
- 76.Council of Europe, Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (CETS No. 225), Vilnius, September 5, 2024, preamble and Articles 3 to 5, 14 to 16, 23 and 24, read in the Internet Archive copy because the site does not allow automated reading.↩
- 77.Council of Europe, Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (CETS No. 225), Article 30, paragraph 1; and “The Framework Convention on Artificial Intelligence”, accessed November 17, 2026, for the date it opened for signature and the observer states that took part in drafting it.↩
- 78.Council of Europe, Treaty Office, signatures and ratifications of the Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (CETS No. 225), status as of November 13, 2026.↩
- 79.Psiquepol + AI, “The State of Artificial Intelligence Public Affairs Worldwide”, section “Who sets the rules”, note 601.↩
- 80.Council of Europe, “The Framework Convention on Artificial Intelligence”, accessed November 13, 2026.↩
- 81.OECD, Recommendation of the Council on Artificial Intelligence, OECD/LEGAL/0449, adopted May 22, 2019 and amended May 3, 2024, with its list of adherents accessed November 13, 2026: the 38 OECD members, 12 non-member countries and the European Union. The OECD.AI page still says 47.↩
- 82.Miranda L. Cole and Dania Assas, Perkins Coie, “Delayed EU Code of Practice Provides Compliance Framework for General-Purpose AI Models”, July 22, 2025.↩
- 83.Regulation (EU) 2024/1689, Articles 53(1), 53(2) and 53(4), 54(1) and 54(6), and 55(1) and 55(2).abc
- 84.Regulation (EU) 2024/1689, published July 12, 2024, Articles 3(63), 5(1)(c), 6, 51(2), 55(1) and 113 and Annex III.ab
- 85.European Commission, European AI Office, accessed November 17, 2026.↩
- 86.Regulation (EU) 2024/1689, Articles 40(1) and 40(2), 53(1) and 53(2), 99(6) and 101(1).abc
- 87.European Commission, signatories of the code of practice for general-purpose models, page updated October 7, 2026.ab
- 88.European Commission, questions and answers on general-purpose AI models, accessed November 13, 2026.ab
- 89.Google, comments on the AI action plan, March 13, 2025.ab
- 90.Andreessen Horowitz, “Setting the Agenda for Global AI Leadership”, February 4, 2025.ab
- 91.OpenAI, Democratic Governance of Frontier AI, June 2, 2026.ab
- 92.Anthropic, “Policy on the AI Exponential”, undated, accessed November 13, 2026.abc
- 93.White House, National Policy Framework for Artificial Intelligence: Legislative Recommendations, March 20, 2026.ab
- 94.Fei-Fei Li, op-ed on SB 1047, Fortune, August 6, 2024.↩
- 95.Governor of California, veto message for SB 1047, September 29, 2024.↩
- 96.California, SB 1047, Safe and Secure Innovation for Frontier Artificial Intelligence Models Act, text passed by the legislature on September 3, 2024 and vetoed, §§ 22602(e) and (g), 22603(a) and 22606(a)(1).↩
- 97.Jay Obernolte and Lori Trahan, discussion draft of the Great American AI Act, June 4, 2026, section 121.ab
- 98.California, AB 853, chapter 674 of 2025, Bus. & Prof. Code § 22757.3.2 and § 22757.1(h), in the reproduction of the official California code by California Public Law, because leginfo does not allow automated reading.abcde
- 99.California, SB 53, Transparency in Frontier Artificial Intelligence Act, chapter 138 of 2025, signed September 29, 2025, §§ 22757.11, 22757.12, 22757.13, 22757.14 and 22757.15, cross-checked against the Psiquepol + AI corpus.abcdefgh
- 100.New York, S6953-B, RAISE Act, chapter 699 of 2025, signed December 19, 2025.abc
- 101.New York, S8828, chapter 96 of 2026, signed March 27, 2026, in force January 1, 2027.abcd
- 102.Anthropic, “Anthropic is endorsing SB 53”, September 8, 2025.ab
- 103.European Union, Regulation (EU) 2022/2554, Articles 1, 2, 31 and 64.↩
- 104.European Union, Regulation (EU) 2022/2554 on digital operational resilience for the financial sector, Article 30.abc
- 105.United States, 45 CFR 160.103, definition of business associate.abc
- 106.United States, 45 CFR 164.502(a)(3) and 45 CFR 160.402.↩
- 107.U.S. Department of Defense, DFARS clause 252.204-7012, which requires the cloud to meet the FedRAMP Moderate level.abcd
- 108.U.S. General Services Administration (GSA), FedRAMP, accessed November 17, 2026.↩
- 109.Solicitors Regulation Authority, notice on the misuse of AI, August 17, 2026.abcd
- 110.Federal Reserve, Office of the Comptroller of the Currency and Federal Deposit Insurance Corporation, letter SR 26-2, April 17, 2026, and its attachment.abcd
- 111.Mistral AI, launch of Le Chat Enterprise, May 2025.abc
- 112.Cohere, private deployments, accessed November 13, 2026.abc
- 113.IBM, Granite, accessed November 13, 2026.abc
- 114.SAP, partnership with Mistral AI for European sovereign AI, November 2025.abc
- 115.AWS, data retention in Amazon Bedrock, accessed November 17, 2026: retention applies to Claude Fable 5, 5.1 and 5.2, with an exemption for customers in the Enterprise Frontier Safeguards program.abcde
- 116.ISO, ISO/IEC 42001:2023, artificial intelligence management system, December 2023.↩
- 117.Microsoft, ISO/IEC 42001, updated September 11, 2026; it covers GitHub Copilot, Microsoft Copilot, Microsoft Foundry and Security Copilot.ab
- 118.ISO, ISO/IEC 42006:2025, requirements for bodies that audit and certify AI management systems, July 2025.↩
- 119.AWS, “AWS achieves ISO/IEC 42001:2023 Artificial Intelligence Management System accredited certification”, November 25, 2024.↩
- 120.Anthropic, “Anthropic achieves ISO 42001 certification for responsible AI”, January 13, 2025.↩
- 121.Google Cloud, ISO/IEC 42001, accessed November 13, 2026; it covers Google Cloud, Google Workspace and the Gemini app.↩
- 122.A-LIGN, announcement of its accreditation to certify ISO/IEC 42001, October 23, 2024, on Synthesia's certificate.↩
- 123.BSI, “KPMG Australia is the first to achieve BSI's AI certification”, October 16, 2024.↩
- 124.Corporate Europe Observatory, “Bias baked in: How Big Tech sets its own AI standards”, January 9, 2025, with statements by Kris Shrishak.↩
- 125.White House, fact sheet on the voluntary commitments of seven AI companies, July 21, 2023, in the Biden administration archive: Amazon, Anthropic, Google, Inflection, Meta, Microsoft and OpenAI.ab
- 126.White House, fact sheet on the commitments of eight more companies, September 12, 2023: Adobe, Cohere, IBM, NVIDIA, Palantir, Salesforce, Scale AI and Stability.↩
- 127.Psiquepol + AI, “The State of Artificial Intelligence Public Affairs Worldwide”, section “Ways to regulate”, note 553.↩
- 128.Google, announcement of the founding of the Frontier Model Forum, July 26, 2023.ab
- 129.Frontier Model Forum, membership, accessed November 13, 2026.↩
- 130.Google, “How we're increasing transparency for gen AI content with the C2PA”, September 17, 2024.ab
- 131.C2PA, membership, accessed November 13, 2026.↩
- 132.Anthropic, “Introducing Anthropic's Responsible Scaling Policy”, September 19, 2023.ab
- 133.Google DeepMind, “Introducing the Frontier Safety Framework”, May 17, 2024.ab
- 134.U.S. Federal Trade Commission, staff report on AI partnerships and investments, January 17, 2025, with the statement by Lina Khan, archived copy.↩
- 135.Gabriela Ramos, “Big Tech shouldn't be writing the rules for AI”, Project Syndicate, published by The Korea Times on April 8, 2026.↩
- 136.Engadget, “China reportedly bans tech companies from buying NVIDIA's AI chips”, September 17, 2025, on the Financial Times report.ab
- 137.OpenAI, countries and territories supported in the API, accessed November 13, 2026.↩
- 138.Google, regions available for Google AI Studio and the Gemini API, accessed November 13, 2026.↩
- 139.Microsoft, code of conduct for enterprise AI services, version 4.0, May 1, 2026.↩
- 140.Psiquepol + AI, “The State of Artificial Intelligence Public Affairs Worldwide”, section “What the state can demand”, note 622, and section “The race with China”, note 384.↩
- 141.OpenAI, usage policies, effective November 20, 2026, accessed November 13.abc
- 142.Psiquepol + AI, “The State of Artificial Intelligence Public Affairs Worldwide”, section “Depending on a lab”, notes 440 and 441.↩
- 143.Harvard Law Today, interview with Louis Tompros, February 25, 2025.↩
- 144.NVIDIA, CFO commentary on the third quarter of fiscal 2027, November 17, 2026, with the fourth-quarter outlook.abcde
- 145.TrendForce, press releases of July 9, 2026 and March 31, 2026.↩
- 146.Psiquepol + AI, “The State of Artificial Intelligence Public Affairs Worldwide”, section “The gap with China”, notes 661 and 663.abcd
- 147.TSMC, Form 20-F for 2025, April 16, 2026. TSMC does not name its customers.↩
- 148.Google Cloud, TPU, accessed November 13, 2026.↩
- 149.AWS, Trainium, accessed November 13, 2026.↩
- 150.Anthropic, announcement of the expansion of its use of Google TPUs, October 23, 2025.↩
- 151.Amazon, Project Rainier, June 24, 2025, updated afterward.↩
- 152.Psiquepol + AI, “The State of Artificial Intelligence Public Affairs Worldwide”, section “Power, chips and capital”, notes 481, 482, 485 and 486.abc
- 153.Psiquepol + AI, “The State of Artificial Intelligence Public Affairs Worldwide”, section “Depending on a lab”, notes 443 and 456.↩
- 154.AWS, responsible AI policy, updated January 13, 2025.↩
- 155.Wikimedia Enterprise, Wikimedia Enterprise partners at Wikipedia's 25th anniversary, January 15, 2026. It does not publish amounts.↩
- 156.Reddit, Form 10-K for fiscal year 2025, February 6, 2026.↩
- 157.News Corp, Form 10-K for fiscal year 2026, August 7, 2026.↩
- 158.AWS, Amazon Bedrock and model choice, accessed November 13, 2026.↩
- 159.Microsoft, Microsoft Foundry models, accessed November 13, 2026.↩
- 160.Google Cloud, Vertex AI Model Garden, accessed November 13, 2026.↩
- 161.SemiAnalysis, “About”, accessed November 17, 2026.↩
- 162.Mistral AI, press release on its Series D, reproduced by StorageNewsletter, September 10, 2026.ab
- 163.Bureau of Industry and Security, “About BIS”, and the interim final rule in note 22, section on the Export Control Reform Act (ECRA), which lists the controlled items list and the license requirement.↩
- 164.NVIDIA, “NVIDIA Vera Rubin Opens Agentic AI Frontier”, press release, March 16, 2026, read in the GlobeNewswire copy.abc
- 165.Bhargava, Allison and Petty, “NVIDIA Vera Rubin POD: Seven Chips, Five Rack-Scale Systems, One AI Supercomputer”, NVIDIA developer blog, March 16, 2026.abc
- 166.NVIDIA, “NVIDIA Vera Rubin Ramps Into Full Production to Power Agentic AI Factories Worldwide”, press release, May 31, 2026.↩
- 167.Intel, “Moore's Law”, September 18, 2023.↩
- 168.U.S. Securities and Exchange Commission, Investor.gov, “Investment-grade Bond (or High-grade Bond)”, glossary, accessed November 17, 2026.↩
- 169.ASML, “The basics of microchips”, accessed November 17, 2026.ab
- 170.Reuters, “Nvidia's Asian supply chain prepares for life after Huang”, November 12, 2026, published by SRN News.↩
- 171.Miriam Waldvogel, Cat Zakrzewski, Jarrell Dillard and Mariana Alfaro, The Washington Post, “How disgraced Nvidia CEO Jensen Huang became a close Trump ally”, November 13, 2026.↩
- 172.Psiquepol + AI, “The State of Artificial Intelligence Public Affairs Worldwide”, November 2026, section “The gap with China”, with its notes 652 to 655. References to the public affairs study use the numbering of its published version.abcd
- 173.CNBC, article on Howard Lutnick and NVIDIA, November 13, 2026.abc
- 174.Reuters, “Nvidia chips pile up in Hong Kong as China withholds import approval”, published by The Standard, November 5, 2026.abcdef
- 175.U.S. House of Representatives, H. Res. 11, 118th Congress, section 1(b).↩
- 176.John Moolenaar, Brian Mast and Rick Crawford, letter to Jensen Huang, November 4, 2026, and Committee on Foreign Affairs press release, November 5.ab
- 177.Jensen Huang, “A Message to Our Company”, November 8, 2026.↩
- 178.The New York Times, “Commerce Dept. Chip Inquiry Was Halted After Nvidia Intervention”, November 9, 2026, read in the DNYUZ reproduction.ab
- 179.NVIDIA, “Statement from the Independent Members of the NVIDIA Board”, November 10, 2026.↩
- 180.NVIDIA, “NVIDIA Announces CEO Transition”, November 12, 2026, also filed as Exhibit 99.1 to the 8-K.ab
- 181.U.S. Securities and Exchange Commission, Form 8-K, General Instruction B.1 and Item 5.02(b).↩
- 182.Maria Curi and Marc Caputo, Axios, “Scoop: Bessent confronts Lutnick in Situation Room over China inaction”, November 11, 2026.↩
- 183.CNN, article on the confrontation between Bessent and Lutnick over Megaspeed, November 11, 2026.↩
- 184.StockAnalysis, NVDA price history, accessed November 12, 2026: close of US$166.19, down 4.3%.↩
- 185.Congressional Research Service, “Committee Discretion in Obtaining Witness Testimony”, LSB11093, December 22, 2023, on the rules of the House of Representatives and contempt of Congress under title 2, section 192, of the U.S. Code.↩
- 186.House Committee on Armed Services, “NDAA”, accessed November 17, 2026.↩
- 187.H.R. 6875, AI OVERWATCH Act, text introduced in the House of Representatives on December 18, 2025, section 2, and S. 1705, Chip Security Act, text introduced in the Senate on May 8, 2025, section 4. The names of the four bills come from the press release in note 18.↩
- 188.TechCrunch, November 12, 2026.↩
- 189.CNBC, transcript of Scott Bessent's interview on “Squawk Box”, November 13, 2026, and article on the Department of Commerce rule, November 16, 2026.↩
- 190.Export Administration Regulations, 15 CFR 744.11, paragraphs (a) and (b), in the copy of the Cornell University Legal Information Institute.↩
- 191.CNN, article on Bessent's accusation against the Department of Commerce and Lutnick's possible departure, November 13, 2026.↩
- 192.Türkiye Today, “Former Nvidia CEO Huang agrees to closed-door interview with US House committees”, November 15, 2026, on statements by John Moolenaar to CBS News.ab
- 193.NVIDIA, “NVIDIA H100 Tensor Core GPU”, specifications, accessed November 17, 2026. The H100 SXM delivers 1,979 trillion operations per second in FP8 without sparsity (3,958 with sparsity); with the rule's formula, its TPP is 1,979 × 8 = 15,832, and 12,000,000 / 15,832 ≈ 758. Our calculation.↩
- 194.National Archives and Records Administration, “About the Federal Register”.↩
- 195.Department of Commerce and Department of Justice, “Joint Statement from Department of Commerce and Department of Justice on Megaspeed Investigation”, November 16, 2026, published as an image on the Department of Commerce account on X.ab
- 196.Lim Hui Jie, CNBC, article on the Department of Commerce and Department of Justice statement about Megaspeed, November 17, 2026.ab
- 197.Council on Foreign Relations, “About CFR”.↩
- 198.NVIDIA, “NVIDIA Announces Financial Results for Third Quarter Fiscal 2027”, November 17, 2026, 8-K exhibit. The margin is our calculation on GAAP operating income of US$77.52 billion.abcd
- 199.Maura Barrett, “Nvidia board rules out Jensen Huang's return as CEO during earnings call”, Scripps News, published by TMJ4, November 17, 2026.↩
- 200.Tim Fernholz, “Nvidia is abandoning China after Jensen Huang's ouster”, TechCrunch, November 17, 2026.abc
- 201.Yahoo Finance, live blog on Jensen Huang’s resignation, November 12, 2026. The intraday figures do not come from a stock exchange source.↩
- 202.Supermicro, “About”, accessed November 17, 2026.↩
- 203.MLQ.ai, summary of the Wall Street Journal report on the orders Megaspeed arranged, November 15, 2026. The records cover orders; they do not show how much was delivered or where it came from.abc
- 204.Korea JoongAng Daily, summary of the Wall Street Journal report on Kang Wei, November 17, 2026, based on corporate records and official statements from the Chinese government.abc
- 205.The Information, “How Jensen Huang's Yearslong Ties to Alice Huang Enabled $7 Billion-Plus Nvidia Chip Sales to China”, November 10, 2026. Requires a subscription; we read it through the summaries by RuntimeWire and Let's Data Science.↩
- 206.Duarte Dias, CBS News, “Megaspeed says Alice Huang left company in 2025, denies wrongdoing amid Nvidia scandal”, updated November 15, 2026.↩
- 207.Bloomberg, via The Next Web, November 10, 2026.ab
- 208.Jing Yang and Qianer Liu, The Information, “Alice Huang's Companies Invested in DeepSeek, Promising Nvidia Chips”, November 12, 2026. Requires a subscription; we read it through the summaries by Kyunghyang Shinmun, of November 13, and by Let's Data Science.ab
- 209.Al Jazeera, “China says US ‘fabricating’ Nvidia spy claims to justify tech ‘blockade’”, November 12, 2026.↩
- 210.@tradingsharks.vip, TikTok video, November 12, 2026, 6 min 49 s. We cite it as an interpretation that is circulating, not as a source for the facts.↩
- 211.WIRED, “Nvidia Will Spend $26 Billion to Build Open-Weight AI Models, Filings Show,” March 11, 2026, read in The Decoder’s summary of March 12; we did not open the original.↩
- 212.NVIDIA, Form 10-Q for the third quarter of fiscal year 2026, November 19, 2025: multiyear cloud commitments of US$26 billion and an investment of up to US$10 billion in Anthropic.abc
- 213.Hugging Face, home page, accessed November 17, 2026.↩
- 214.Constellation Research, “Company Overview”, accessed November 17, 2026.↩
- 215.Larry Dignan, Constellation Research, on the Nemotron models, March 12, 2026.↩
- 216.Anthropic, “Detecting and preventing distillation attacks”, February 23, 2026. It is an accusation by the affected company.↩
- 217.“Open Weights and American AI Leadership,” version 6, archive of July 29, 2026 with the July 24 header, 235 signatories by our own count; also in NVIDIA’s copy. Anthropic and xAI do not appear by name in any version; SpaceX, xAI’s parent company, does.abcdefghi
- 218.Interconnects, “About”, accessed November 17, 2026.↩
- 219.Rebecca Bellan, “Jensen Huang's exit puts Nvidia's support for open source in question”, TechCrunch, November 12, 2026, with Nathan Lambert’s post on X.↩
- 220.NTIA, Dual-Use Foundation Models with Widely Available Model Weights, July 30, 2024.abcd
- 221.Open Source Initiative, “About”.↩
- 222.Open Source Initiative, Open Source AI Definition 1.0, published on October 28, 2024.↩
- 223.Open Source Initiative, final report of the definition process, with the validation from May to July 2024.↩
- 224.Allen Institute for AI, “About”.↩
- 225.Allen Institute for AI, “Olmo 3”, November 20, 2025.↩
- 226.NVIDIA, “NVIDIA Nemotron 3: Efficient and Open Intelligence”, arXiv, December 2025.↩
- 227.Open Source Initiative, text of the MIT License, and Apache Software Foundation, Apache License 2.0, sections 2 to 4.↩
- 228.DeepSeek, model cards for DeepSeek-R1 and DeepSeek-V3-0324, under the MIT license, on Hugging Face. The original V3 weights carried a separate model license.ab
- 229.Google, “Gemma 4: expanding the Gemmaverse with Apache 2.0”, April 2, 2026.abc
- 230.OpenAI, model card for gpt-oss-120b and gpt-oss-20b, August 5, 2025, with the Apache 2.0 license and the gpt-oss usage policy.ab
- 231.Microsoft, model card for Phi-4-reasoning-vision-15B, March 4, 2026.abcd
- 232.IBM, Granite models on Hugging Face, with Granite 4.2 of August 2026.ab
- 233.Meta, Llama 4 community license, April 5, 2025.↩
- 234.Meta, Llama 4 acceptable use policy, with the same exclusion of the European Union for multimodal models as Llama 3.2.↩
- 235.Open Source Initiative, “Meta's LLaMa license is still not Open Source”, February 18, 2025, on the Llama 3.x license.↩
- 236.Tencent, Hunyuan community license.↩
- 237.xAI, Grok 2 community license, clause 2(b), which prohibits training other foundation or general-purpose models and allows fine-tuning Grok 2.ab
- 238.Moonshot AI, Kimi K3 license and model card, July 16, 2026. The threshold is measured by the licensee’s and its affiliates’ revenue over 12 months.abcdefg
- 239.Moonshot AI, Kimi K3.1 Pro model card and license, weights published on November 4, 2026, its technical blog, section “Cybersecurity,” and TechNode, November 5, 2026.abc
- 240.Alibaba, Qwen3.8-Max license and model card, August 2026, and model card for Qwen3.8-Flash-Next, with the Qwen Community License 1.0, August 24, 2026.abcdef
- 241.Z.ai, GLM-5.3 license, August 2026.abcde
- 242.MiniMax, MiniMax M3 community license, June 2026.↩
- 243.Meta, model card and license for Muse Spark Open, October 2026.ab
- 244.Mistral AI, model catalog, which marks Medium as service-only and Codestral as Premier, and Mistral Large 4 model card, with the Mistral Ecosystem Open-Weight License.abcde
- 245.Psiquepol + AI, “The State of Artificial Intelligence Public Affairs Worldwide,” section “Ways to regulate,” note 512.↩
- 246.Simon Lermen, Charlie Rogers-Smith and Jeffrey Ladish, “LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2-Chat 70B”, arXiv, October 31, 2023.↩
- 247.Xiangyu Qi and others, “Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!”, arXiv, October 5, 2023.↩
- 248.EleutherAI, “About”.↩
- 249.Kyle O'Brien and others, “Deep Ignorance: Filtering Pretraining Data Builds Tamper-Resistant Safeguards into Open-Weight LLMs”, arXiv, August 8, 2025, version of February 17, 2026. The experiments use models of 6.9 billion parameters.↩
- 250.International AI Safety Report, “About”.↩
- 251.International AI Safety Report, January 2025, section 2.4, which calls marginal risk a high-level consensus, and p. 175 of the PDF, on the impossibility of withdrawing access.ab
- 252.Sayash Kapoor, Rishi Bommasani and others, “On the Societal Impact of Open Foundation Models”, arXiv, February 27, 2024.abc
- 253.International AI Safety Report 2026, February 3, 2026, section 3.4 and box 3.1.↩
- 254.FAR.AI, “About”.↩
- 255.FAR.AI, “Security Stress Test: Exposing the Brittleness of DeepSeek-V4-Pro's Safeguards”, May 11, 2026.ab
- 256.National Institute of Standards and Technology of the United States, CAISI, and the UK AI Security Institute, “Preliminary Assessment of Kimi K3's Cyber Capabilities”, July 23, 2026.ab
- 257.National Institute of Standards and Technology of the United States, center page, accessed November 17, 2026.↩
- 258.Executive Order 14434, “Inaugurating the Era of Super Intelligence”, signed on September 29, 2026 and published in the Federal Register on October 2, 2026, sections 2(a) and 3(b).ab
- 259.National Institute of Standards and Technology of the United States, Center for Advancing Innovation and Standards for Super Intelligence (CAISSI), accessed November 17, 2026.↩
- 260.Shen Zhou Hong and others, “Measuring Mid-2025 LLM-Assistance on Novice Performance in Biology”, arXiv 2602.16703, February 18, 2026.↩
- 261.Emily Forlini, Fortune, on cyberattacks using AI models, October 21, 2026, with statements by Jessica Ji of CSET.ab
- 262.Anthropic, “Claude Fable 5.2 and Claude Mythos 5.2”, November 9, 2026, and its system card, executive summary and section 3.4.ab
- 263.Wiz Research, “PHP2Shell”, October 21, 2026, on CVE-2026-109618 and CVE-2026-109619.ab
- 264.National Institute of Standards and Technology of the United States, CAISI, assessment of GLM-5.3’s cyber capabilities, September 17, 2026.abcd
- 265.Epoch AI, gap between open and closed models in cybersecurity, October 16, 2026, with data as of October 8.↩
- 266.10a Labs, “Uncensored Open-weight Models: Redistribution as the Persistence Layer”, arXiv, September 4, 2026.↩
- 267.Center for Security and Emerging Technology, “About Us”, Georgetown University, accessed November 17, 2026.↩
- 268.Colibri, GitHub repository, accessed November 18, 2026.↩
- 269.Google Threat Intelligence Group, “From Prompting to Autonomy: The Evolution of Adversarial AI”, September 8, 2026.ab
- 270.Google Threat Intelligence Group, “Agentic Threat Activity Crosses a Significant Threshold”, November 16, 2026.ab
- 271.Anthropic, Project Glasswing, April 7, 2026. Anthropic competes in this market.ab
- 272.Kanishka Narayan, the United Kingdom’s AI Minister, “Building a safer, more secure Britain in the age of AI”, October 12, 2026.abc
- 273.U.S. Cybersecurity and Infrastructure Security Agency, “Critical Vulnerabilities in PHP Require Immediate Patching”, October 21, 2026.↩
- 274.The Register, “This year's bug count just passed 100K”, November 13, 2026, with VulnCheck’s count.↩
- 275.PCWorld, Microsoft’s statements on the Outlook outage, November 13, 2026.↩
- 276.FAA, Air Traffic Control System Command Center, advisory 120 of November 13, 2026, nationwide ground stop due to a communications failure.↩
- 277.Fortune, on the investigation of the nationwide flight ground stop, November 15, 2026.↩
- 278.CNBC, on the JPMorgan Chase outage, November 16, 2026, with the bank’s statement.↩
- 279.9to5Mac, on the iCloud and App Store outage, November 17, 2026, with Apple’s statement as quoted by Mark Gurman.↩
- 280.Fox Business, interview with Jay Clayton, Director of National Intelligence, November 17, 2026.ab
- 281.Outlook India, on Markwayne Mullin’s statements, November 2026.↩
- 282.Fireworks AI, security bulletins of October 10, November 9 and November 17, 2026.↩
- 283.UK AI Security Institute, pretraining data filtering in open-weight biological models, October 26, 2026.↩
- 284.Japan’s Ministerial Meeting on Cybersecurity, urgent cybersecurity countermeasures package, November 6, 2026, section on the evaluation of high-performance AI models.↩
- 285.National Institute of Standards and Technology of the United States, National Security Agency of the United States and U.S. Cybersecurity and Infrastructure Security Agency, “Key Practices for Incident Readiness and Response in Super Intelligence Training and Evaluation”, October 16, 2026.↩
- 286.H.R. 10981, American AI Security Act, introduced by Josh Gottheimer and Mike Lawler on November 10, 2026.↩
- 287.Rob Bonta, California Attorney General, press release and complaint against OpenAI Foundation and OpenAI Group PBC before the San Francisco Superior Court, November 16, 2026, paragraphs 2 to 6, 90 and 113 to 117.abcde
- 288.National Technical Committee for Cybersecurity Standardization (TC260), notice 121 of 2026 with the plan for 25 national standards, October 13, 2026, plan 20264767-T-469.ab
- 289.Ministry of Commerce and Ministry of Science and Technology of China, Measures for the Administration of Prohibited and Restricted Export Technologies, articles 2 and 3, WIPO Lex copy.↩
- 290.Ministry of Commerce and Ministry of Science and Technology of China, consultation notice No. 48 of 2026 and its annex 1, October 16, 2026, section three and items 266501J and 266504X.abcdefghi
- 291.Dario Amodei, “Our position on open-weights models”, July 27, 2026.abcde
- 292.OpenAI, “Introducing gpt-oss”, August 5, 2025.ab
- 293.Eric Wallace and others, “Estimating Worst-Case Frontier Risks of Open-Weight LLMs”, arXiv, August 5, 2025.↩
- 294.Hugging Face, models published by OpenAI, accessed November 12, 2026: gpt-oss-safeguard-120b and -20b (September 18, 2025), circuit-sparsity (December 11, 2025) and privacy-filter (April 17, 2026).↩
- 295.Hugging Face, models published by xAI, accessed November 12, 2026.↩
- 296.xAI, “xAI joins SpaceX”, February 2026.↩
- 297.Mark Zuckerberg, “Open Source AI Is the Path Forward”, July 23, 2024.abcdef
- 298.Mark Zuckerberg, “Personal Superintelligence”, July 30, 2025.↩
- 299.Hugging Face, models published by Meta Llama, accessed November 12, 2026.↩
- 300.The Decoder and The Next Web, April 8, 2026, on the launch of Muse Spark. We did not read Meta’s announcement.↩
- 301.Mark Zuckerberg, “The Future is for Everyone”, August 10, 2026.abcde
- 302.Meta, “Introducing Muse Glimmer”, August 10, 2026.↩
- 303.Meta, “Muse Spark Open”, October 15, 2026.↩
- 304.Psiquepol + AI, “The State of Artificial Intelligence Public Affairs Worldwide,” section “The race with China,” notes 330, 396 and 397.abc
- 305.Reflection, model card for Beam-1.0-501B-A23B, October 2026.↩
- 306.“Open Weights and American AI Leadership,” original version published by Microsoft, July 24, 2026, 25 signatories.abcdefghi
- 307.Microsoft, page of the open-weights letter, accessed November 12, 2026, with the count as of August 3.ab
- 308.Fortune, July 24, 2026. Huang’s post on X could not be read at the source.ab
- 309.DeepSeek, model cards for DeepSeek-V4-Pro and DeepSeek-V4.1-Flash.ab
- 310.Tencent, Hy3 models on Hugging Face, July 2026.↩
- 311.ByteDance Seed, launch of Seed 2.0, February 14, 2026.abcd
- 312.ByteDance Seed, model card for Seed-OSS-36B-Instruct, under the Apache 2.0 license, August 20, 2025.abc
- 313.Andreessen Horowitz and OpenRouter, “State of AI 2025: An Empirical 100 Trillion Token Study”, December 2025.↩
- 314.CNBC, report on the global adoption of Chinese models, September 26, 2026. Based on data from OpenRouter and Vercel.↩
- 315.CNBC, report on the cost of Chinese models, July 7, 2026.↩
- 316.Bloomberg, via Fortune, October 21, 2025.↩
- 317.Stanford HAI, AI Index 2026, “Technical Performance”.abc
- 318.Arena, “Arena's Ranking Method”, November 14, 2025, updated February 28, 2026. Until that publication, Arena reported only each model’s optimistic rank, the lower end of its range of possible ranks.↩
- 319.Alibaba Group, results for the quarter ended June 2026 and results notice filed with the Hong Kong Stock Exchange, August 20, 2026. Alibaba publishes the quarter ended September on November 19, after the cutoff.abc
- 320.DeepSeek, overview of the V3 and R1 inference system, February 2025, measured over 24 hours from February 27 to 28.abcd
- 321.Psiquepol + AI, “The State of Artificial Intelligence Public Affairs Worldwide,” section “The race with China,” note 414.ab
- 322.National Institute of Standards and Technology of the United States, CAISI, Evaluation of DeepSeek AI Models, September 30, 2025, with the full report. The Department of State prepared the narratives benchmark and it is not public; a judge model grades the response.abc
- 323.National Institute of Standards and Technology of the United States, CAISI, assessment of GLM-5.2, July 17, 2026.↩
- 324.Jennifer Pan and Xu Xu, “Political censorship in large language models originating from China”, PNAS Nexus, February 17, 2026, with questions in Chinese to 2023 and 2024 models.↩
- 325.Aleph Alpha, “Training on the party line”, September 28, 2026, on Nemotron Cascade 2 (30B-A3B): 17% in the text and 18% in its figure 4. The company sells sovereign AI and competes with the Chinese providers; a model generated and graded the questions.↩
- 326.ETH Zurich, “Apertus: a fully open, transparent, multilingual language model”, September 2, 2025.abc
- 327.Swiss AI, model card for Apertus 1.5 70B, 2026, under the Apache 2.0 license.↩
- 328.Technology Innovation Institute, launch of Falcon Perception, May 3, 2026; the Falcon-Perception-300M repository on Hugging Face is from April 3.↩
- 329.MBZUAI and G42, press release on K2 Think, September 9, 2025, redistributed by AAP and PR Newswire.↩
- 330.Press Information Bureau of the Government of India, press release of March 25, 2026 and of March 13, 2026, on the IndiaAI Mission.↩
- 331.Cabinet of Japan, Basic AI Plan, revision of July 14, 2026; the first plan was approved on December 23, 2025.↩
- 332.Ministry of Science and ICT of South Korea, sovereign AI foundation model program, August 2025.↩
- 333.Government of South Korea, result of the first evaluation of the sovereign foundation model, January 15, 2026: LG, Upstage and SKT passed.ab
- 334.TASS, “Russia's first AI law comes into force”, September 1, 2026.ab
- 335.Psiquepol + AI, “The State of Artificial Intelligence Public Affairs Worldwide,” section “AI under sanctions,” notes 681 and 682.↩
- 336.Sber, model card for GigaChat 3.5 432B-A28B, under the MIT license, September 3, 2026.↩
- 337.European Commission, “Pioneering AI project awarded for opening Large Language Models to European languages”, February 3, 2025.ab
- 338.Office of Science and Technology Policy, NSTM-4, “Adversarial Distillation of American AI Models”, April 23, 2026.↩
- 339.Executive Order 14409, June 2, 2026, 91 FR 34565, section 3(b)(ii).ab
- 340.Executive Order 14319, July 23, 2025, section 4.ab
- 341.Office of Management and Budget, memorandum M-26-04, December 11, 2025, note 6.ab
- 342.National Institute of Standards and Technology of the United States, CAISI, announcement of May 5, 2026 on the testing agreements with Google DeepMind, Microsoft and xAI, via MeriTalk.↩
- 343.Psiquepol + AI, “The State of Artificial Intelligence Public Affairs Worldwide,” section “Ways to regulate,” note 552.↩
- 344.Mark MacCarthy, “US Government's AI Risk Review Should Apply to Open-Weight Models”, Tech Policy Press, August 13, 2026.↩
- 345.Fiscal year 2027 defense bill: H.R. 8800, passed by the House on September 14, 2026, section 242, and S. 4784, reported out of committee in the Senate on July 27, 2026, sections 1634 and 1655. Checked against the Psiquepol + AI corpus.ab
- 346.United States, 10 U.S.C. § 4872, definition of covered nation, to which section 5 of H.R. 10152 refers, in the reproduction of Cornell University’s Legal Information Institute, accessed November 17, 2026.abc
- 347.H.R. 10152, Open-Source AI Leadership Act, introduced by Gabe Evans on August 27, 2026, sections 2 to 5.↩
- 348.The Star, with material from the South China Morning Post, September 6, 2026.↩
- 349.H.R. 8094, AI Foundation Model Transparency Act of 2026, introduced by Don Beyer on March 26, 2026, section on exemptions.ab
- 350.State Council of China, opinion on the deep implementation of the “AI+” action (国发〔2025〕11号), signed on August 21 and published on August 26, 2025, sections 6 and 11.ab
- 351.Outline of China’s 15th Five-Year Plan, Xinhua version of March 13, 2026, chapter 13.↩
- 352.Ministry of Industry and Information Technology of China and seven other agencies, implementation opinion on the “AI+ Manufacturing” action (工信部联科〔2025〕279号), signed on December 25, 2025 and published on January 7, 2026, section 16 and section 23 of its annex 2.↩
- 353.Ministry of Foreign Affairs of China, Global AI Governance Action Plan, July 26, 2025, point 5.ab
- 354.Xinhua on the Chinese government portal, report on the signing of the World AI Cooperation Organization agreement, July 16, 2026.↩
- 355.Xinhua, July 19, 2026. State news agency; it attributes the program to “the relevant parties,” with no identified spokesperson.↩
- 356.Guangming Daily, with CCTV, November 3, 2026, statements by Wang Ruomeng of the National Development and Reform Commission.↩
- 357.Psiquepol + AI, “The State of Artificial Intelligence Public Affairs Worldwide,” section “The race with China,” notes 402 and 403.↩
- 358.Ministry of Commerce of China, spokesperson’s response on distillation, September 9, 2026.↩
- 359.Ministry of Commerce of China, spokesperson’s statement, July 27, 2026.↩
- 360.NVIDIA, CFO commentary on the second quarter of fiscal year 2027, August 26, 2026, with the third-quarter outlook.↩
- 361.NVIDIA and OpenAI, announcement of their strategic partnership, September 22, 2025. Non-binding letter of intent; the February 2026 10-K says NVIDIA was finalizing the agreement.↩
- 362.PC Gamer and Il Sole 24 Ore, March 2026, on what Huang said at the Morgan Stanley conference. The May and August 10-Qs do not name an investment in OpenAI.↩
- 363.NVIDIA, Form 8-K on the guarantees for the Ohio campus, August 17, 2026: residual value guarantees starting in 2028, which OpenAI is obligated to reimburse.↩
- 364.xAI, “Series E”, January 6, 2026. It does not publish NVIDIA’s amount.↩
- 365.Business Insider, “Jensen Huang's Nvidia Ouster Removes Major Threat to Anthropic IPO”, November 2026.↩
- 366.Michael Cembalest, J.P. Morgan Asset Management, “The Icarus Files: What's next for Nvidia and AI”, Eye on the Market, November 14, 2026, accessed November 18, 2026.↩
- 367.NVIDIA, model card for Nemotron 3 Nano 30B-A3B, December 15, 2025.↩
- 368.NVIDIA, model card for Nemotron 3 Super 120B-A12B, March 11, 2026.↩
- 369.NVIDIA, model card for Nemotron 3 Ultra 550B-A55B, June 4, 2026.↩
- 370.Linux Foundation and PyTorch Foundation, OpenMDW, license for models introduced in 2025, accessed November 17, 2026.↩
- 371.NVIDIA, model card for Nemotron 3.5 Super VL 120B-A12B, October 12, 2026.↩
- 372.Reuters, “Exclusive: US DOJ is inquiring about Nvidia's purchase of Hugging Face,” October 28, 2026, published by KFGO.abc
- 373.Federal Trade Commission of the United States, premerger notification program, accessed November 17, 2026.↩
- 374.Reuters, “US DOJ probes Nvidia's licensing deal with AI startup Groq, NYT reports”, published by The Star, September 10, 2026.↩
- 375.MLex, September 2026, on a Reuters report.↩
- 376.Mistral AI, announcement of its Series C, September 9, 2025.↩
- 377.TechCrunch, October 9, 2025, which names NVIDIA among the investors.↩
- 378.Cyberspace Administration of China, summons of NVIDIA over the H20, July 31, 2025.↩
- 379.State Administration for Market Regulation of China, preliminary finding on NVIDIA, September 15, 2025, and conditions of the Mellanox purchase, April 16, 2020, conditions 1 and 2. We did not find a final ruling.ab
- 380.White House, fact sheet on the economic and trade agreement with China, November 1, 2025.ab
- 381.NVIDIA, “The Future of AI Is Open and Proprietary”, March 25, 2026, signed by Kari Briski, with Huang’s quote at GTC.↩
- 382.Jensen Huang, GTC Washington keynote, October 28, 2025, transcript from The Singju Post.↩
- 383.CSIS, “Nvidia's Jensen Huang: Securing American Leadership in AI”, conversation with John Hamre, December 3, 2025, with transcript.↩
- 384.NVIDIA, press release on Europe’s AI infrastructure, June 11, 2025.↩
- 385.NVIDIA, blog post on the World Governments Summit in Dubai, February 12, 2024.↩
- 386.Jensen Huang on the Special Competitive Studies Project’s podcast Memos to the President, July 14, 2025, transcript from The Singju Post.ab
- 387.NVIDIA, live blog of the GTC Berlin keynote, October 21, 2026.↩
- 388.Ben Thompson, “An Interview with Nvidia CEO Jensen Huang About Chip Controls, AI Factories, and Enterprise Pragmatism”, Stratechery, May 19, 2025.↩
- 389.Reuters, “Nvidia's Jensen Huang: ‘China is going to win the AI race,’ FT reports”, published by The Standard, November 6, 2025, on the Financial Times interview and Huang’s statement on X.↩
- 390.Dwarkesh Patel, interview with Jensen Huang, April 15, 2026, with transcript.ab
- 391.CBS News, Jo Ling Kent’s interview with Jensen Huang, September 20, 2026.ab
- 392.Legal Information Institute, Cornell University, “Products liability”, accessed November 17, 2026.↩
- 393.NVIDIA, press release on the Open Agent Safety Platform, September 28, 2026.↩
- 394.Axios, “Exclusive: Nvidia's Jensen Huang defends Chinese AI amid Kimi panic”, July 22, 2026, interview by Mike Allen on July 21; full video, minutes 19:03 and 27:45.↩
- 395.Lionel Laurent, Bloomberg Opinion, column on NVIDIA’s “sovereign AI”, June 18, 2025, published by the Taipei Times on the 21st.ab
- 396.Ben Thompson, “Nvidia's Risky Business”, Stratechery, August 11, 2026.↩
- 397.NVIDIA, GTC 2026 live updates, March 16, 2026. The source gives no dates for Feynman.abcd
- 398.Groq, “Groq and Nvidia Enter Non-Exclusive Inference Technology Licensing Agreement to Accelerate AI Inference at Global Scale”, December 24, 2025.↩
- 399.U.S. Energy Information Administration, note of February 20, 2026.↩
- 400.pv magazine, “China adds 315 GW of solar in 2025”, January 28, 2026.↩
- 401.World Nuclear Association, “Plans for New Reactors Worldwide”, updated November 16, 2026.↩
- 402.Elon Musk, post on X, January 31, 2026.↩
- 403.Ember, Global Electricity Review 2026, “Electricity demand and supply trends”, 2026.↩
- 404.Ed Zitron, “Dead Money”, September 29, 2026.↩
- 405.Nathan Lambert, “Teaching Everyone to Fish for Tokens”, Interconnects, August 17, 2026.↩
- 406.Eric Xu, speech at Huawei Connect, September 18, 2025.abcde
- 407.Citadel Securities, “Nvidia's Jensen Huang on AI & the Next Frontier of Growth”, October 14, 2025, at 47:50.↩
- 408.Reuters, “China bans foreign AI chips from state-funded data centres, sources say”, published by The Standard, November 5, 2025.↩
- 409.The Standard, “China subsidizes power for tech firms to boost domestic AI chip development, FT reports”, November 4, 2025, on the Financial Times report.↩
- 410.Bureau of Industry and Security, guidance on General Prohibition 10 and Chinese advanced-computing integrated circuits, May 13, 2025.↩
- 411.The Next Web, “AI chipmaker Cambricon's sales miss points to tight chip supply in China”, October 23, 2026, on a Bloomberg report.ab
- 412.The Next Web, “Huawei's answer to Nvidia isn't a faster chip, it's a much bigger machine”, September 18, 2026, on Huawei Connect 2026 and statements to Bloomberg. We did not find the speech on Huawei's website.ab
- 413.Reuters, “Huawei's new AI chip finds favour with ByteDance, Alibaba, which plan to place orders, sources say”, March 27, 2026, published by The Standard. Anonymous sources.↩
- 414.TrendForce, “Huawei debuts Atlas 350 on Ascend 950PR with in-house HBM, touting 2.8x H20 performance”, March 23, 2026. The comparison with the H20 is Huawei's.ab
- 415.Yicai Global, “Chinese AI chip giant Cambricon posts first annual profit since listing”, March 13, 2026, on the company's annual report.↩
- 416.Caixin Global, “Alibaba unveils new AI chip to rival Nvidia's China offerings”, January 29, 2026.↩
- 417.Euronews, “China's bet on home-grown chips pays off with Moore Threads blockbuster debut”, December 5, 2025.↩
- 418.Caixin Global, “Chinese GPU maker MetaX doubles revenue amid push for domestic chips”, March 26, 2026, on the company's annual report.↩
- 419.Fortune, “AI chip designer Biren's shares surge 76% on debut in Hong Kong”, January 2, 2026.↩
- 420.The Standard, “Tencent-backed Enflame IPO draws 6,109 times online demand”, September 2, 2026.↩
- 421.Caixin Global, “Baidu's chip unit Kunlunxin files for Hong Kong IPO to tap AI investment boom”, January 2, 2026.↩
- 422.Reuters, “In China, the AI bubble has already burst”, October 26, 2026.ab
- 423.Forbes México, “Restricción de Trump a centros de datos en China avanza pese las gestiones de cabildeo”, November 17, 2026, with information from Reuters.↩
- 424.Reuters, “US says China's Huawei can't make more than 200,000 AI chips in 2025”, June 12, 2025, on the testimony of Jeffrey Kessler, Under Secretary for Industry and Security at the U.S. Department of Commerce, before the South and Central Asia subcommittee of the House of Representatives.↩
- 425.Skadden, “Commerce Department's New Export-Related Restrictions Inhibit Semiconductor Design by and Manufacturing for Huawei”, May 2020, on the Department of Commerce rule of May 15, 2020. The Department of Commerce's website blocked the request.↩
- 426.Epoch AI, “Will Huawei catch up to Nvidia by 2030?”, accessed November 16, 2026.abc
- 427.Bureau of Industry and Security, “Addition of Entities to the Entity List, Revision of Entry on the Entity List, and Removal of Entities From the Entity List”, 85 FR 83416, Federal Register, December 22, 2020, in effect since December 18, 2020.↩
- 428.TechInsights, “SMIC N+3 confirmed: Kirin 9030 analysis reveals how close SMIC is to 5nm”, December 11, 2025.↩
- 429.TechNode Global, SMIC's third-quarter 2026 results, November 13, 2026.↩
- 430.TrendForce, “China reportedly aims to boost 7nm/5nm output fivefold in two years, driven by SMIC and Hua Hong”, February 25, 2026, on a Nikkei report.↩
- 431.Anton Shilov, Tom's Hardware, “China to increase leading-edge chip output by 5x in two years, report claims: aims to lift 7nm and 5nm production to 100,000 wafers per month”, February 25, 2026, on the same Nikkei report, which specifies that the figure is wafer starts per month.↩
- 432.TrendForce, “U.S. reportedly moves to block chip tool shipments to China's No. 2 foundry Hua Hong as it seeks 7nm after SMIC”, April 29, 2026, on a Reuters report of April 28.↩
- 433.SEMI, report on 2025 global sales of semiconductor manufacturing equipment, April 8, 2026, published by I-Connect007.↩
- 434.TrendForce, “China's SMEE showcases lithography systems with 90% localized parts”, November 9, 2026, citing Reuters on Aishengna.↩
- 435.Bureau of Industry and Security, “Commerce strengthens export controls to restrict China's capability to produce advanced semiconductors for military applications”, December 2, 2024.↩
- 436.House Foreign Affairs Committee, “Chairman Mast: HFAC advances MATCH Act”, April 22, 2026.ab
- 437.Bureau of Industry and Security, “Extension of the Suspension of Expansion of End-User Controls for Affiliates of Certain Listed Entities”, Federal Register, November 9, 2026; summarized by Thompson Hine.ab
- 438.The National, “China extends pause on rare earth export curbs for two months”, November 9, 2026. We did not locate the announcement from China's Ministry of Commerce.ab
- 439.Bureau of Industry and Security, “Department of Commerce expands Entity List to cover affiliates of listed entities”, September 29, 2025.↩
- 440.Bureau of Industry and Security, “Expansion of End-User Controls To Cover Affiliates of Certain Listed Entities”, interim final rule, 90 FR 47201, Federal Register, September 30, 2025, in effect since September 29, 2025.↩
- 441.Bureau of Industry and Security, “One Year Suspension of Expansion of End-User Controls for Affiliates of Certain Listed Entities”, 90 FR 50857, Federal Register, November 12, 2025, in effect since November 10, 2025.↩
- 442.China's Ministry of Commerce, Announcement No. 61 of 2025, October 9, 2025, Article 1, in the CSET translation at Georgetown University.↩
- 443.Reuters, “China's CXMT reports quarterly profit after Asia's biggest IPO this year”, October 30, 2026, published by The Standard. The HBM3E plans and the yields come from anonymous sources.↩
- 444.The Register, “DeepSeek's new V3.1 release points to potent new Chinese chips coming soon”, August 22, 2025, citing DeepSeek's comment on WeChat.↩
- 445.DeepSeek, model card for DeepSeek-V3.1 on Hugging Face, August 2025.↩
- 446.InfoWorld, “Chinese AI firm trains state-of-the-art model entirely on Huawei chips”, January 14, 2026.↩
- 447.TechNode, “DeepSeek reportedly plans to train V5 model on Huawei Ascend chips”, November 4, 2026, citing STAR Market Daily. DeepSeek has not confirmed it.↩
- 448.SemiAnalysis, “Huawei AI CloudMatrix 384, China's Answer to Nvidia GB200 NVL72”, April 16, 2025.↩
- 449.TechCrunch, article on the chips used to train Chinese models, November 13, 2026. On the analysis by Will DePue.↩
- 450.Gregory C. Allen, CSIS, “DeepSeek, Huawei, Export Controls, and the Future of the U.S.-China AI Race”, March 7, 2025.↩
- 451.Alicia García-Herrero and Bertin Martens, Bruegel, “Stack battles”, June 22, 2026, copy published by SIPoTra.↩
- 452.CSIS, “About”, and Bruegel, “About”, accessed November 17, 2026.↩
- 453.RAND, “About”, accessed November 17, 2026.↩
- 454.Lennart Heim, RAND, commentary on the gap between Chinese and American models, May 2, 2025.↩
- 455.TechNode, “China's AgiBot leads global humanoid robot shipments in 2025, Omdia says”, January 9, 2026, with figures from Omdia.ab
- 456.South China Morning Post, “Nvidia's Huang calls China 'formidable' in robotics as company bets on physical AI”, March 21, 2026, on a podcast recorded at GTC in San Jose.↩
- 457.Google DeepMind, model card for Gemini Robotics On-Device 2, July 30, 2026.↩
- 458.Figure, “F.03 Arrives at BMW”, June 30, 2026.↩
- 459.NVIDIA, “NVIDIA Isaac GR00T N1.7: Open Reasoning VLA Model for Humanoid Robots”, blog post on Hugging Face, April 17, 2026, and the GR00T technical page, which lists the license of N1.5 and N1.6 as noncommercial. The same page warns that the license file for the N1.7 weights was still the noncommercial one at the time it was consulted.ab
- 460.NVIDIA, model card for GR00T-N1.7-3B on Hugging Face, with the license “nvidia-open-model-license”, accessed in November 2026.↩
- 461.NVIDIA, “How Cosmos 3 Helps Physical AI Think Before It Acts”, May 31, 2026.↩
- 462.The Robot Report, “NVIDIA launches GR00T 2, Isaac Lab 3.0 at CoRL”, November 2026.↩
- 463.Physical Intelligence, openpi repository, Apache 2.0 license, accessed in November 2026. It also carries the Gemma license, because its models are built on PaliGemma, from Google.↩
- 464.The AI Insider, “Physical Intelligence Opens Pi API Beta for Robot Model Training and Deployment”, October 14, 2026. No price is published.↩
- 465.Hugging Face, model card for SmolVLA, Apache 2.0 license; the model was presented on June 3, 2025.↩
- 466.Unitree, model card for UnifoLM-VLA-0 on Hugging Face, with the CC BY-NC-SA 4.0 license, and its repository, which dates the release of the weights to January 29, 2026.↩
- 467.Beijing Academy of Artificial Intelligence, “RoboBrain 2.5”, arXiv, January 2026, with its repository and its Hugging Face model card, which give the Apache 2.0 license and the Qwen3-VL base model.↩
- 468.Yicai Global, “Unitree Robotics Files for USD608 Million IPO in Shanghai”, and Gasgoo, “Unitree Robotics IPO Reaches Key Milestone”, 2026, with figures from the prospectus. The figures are the company's own.ab
- 469.Gasgoo, “IDC: AgiBot Leads Global Humanoid Robot Shipments With Over 8,600 Units in H1”, October 2026, with figures from IDC.↩
- 470.36Kr, article on the Unitree prospectus, 2026, with the company's own figures.↩
- 471.Nicholas Gordon, Fortune, “Unitree surges 460% in trading debut”, August 19, 2026.↩
- 472.Gasgoo, “Unitree's First Post-IPO Earnings: Reported Net Profit Up 270%, Recurring Net Profit Down 19%”, October 30, 2026.↩
- 473.UBTech, 2025 annual results, filed with the Hong Kong stock exchange on March 31, 2026.↩
- 474.Matt Gonzales, TechRepublic, “Nvidia-Backed Agility Robotics Goes Public at $2.5B”, November 5, 2026.↩
- 475.Crunchbase News, “Sector Snapshot: Robotics Startups On Fire As Venture Funding Surges To Record Numbers In 2026”, June 22, 2026.↩
- 476.Skild AI, “Skild AI Raises $1.4B, Now Valued Over $14B”, Business Wire, January 14, 2026.↩
- 477.K&L Gates, “FCC Adds Foreign-Produced Advanced Robotic Devices to the Covered List: Five Things to Know”, August 3, 2026.↩
- 478.TechCrunch, “Pentagon says Alibaba, Baidu, BYD and Unitree support China's military”, June 8, 2026.↩
- 479.House Select Committee on the Chinese Communist Party, “Effort to ban Chinese robots moves forward in House-passed NDAA”, July 22, 2026.↩
- 480.Víctor Mayoral-Vilches, Andreas Makris and Kevin Finisterre, “Cybersecurity AI: Humanoid Robots as Attack Vectors”, arXiv, September 2025.↩
- 481.Outline of China's 15th Five-Year Plan, Xinhua version of March 13, 2026, Chapter 5, Section 2.↩
- 482.TechNode, “Humanoid robot maker Galbot raises RMB 2.5 billion”, March 2, 2026, and Cailian, article on Galbot's funding round, March 2, 2026.↩
- 483.Bloomberg, “China Warns of Bubble Risks in Booming Humanoid Robots Arena”, November 27, 2025.↩
- 484.NVIDIA, “NVIDIA Blackwell-Powered Jetson Thor Now Available, Accelerating the Age of General Robotics”, August 25, 2025.↩
- 485.South China Morning Post, “Chinese humanoid robotics firms Unitree, UBTech, AgiBot adopt Nvidia's new 'robot brain'”, August 26, 2025.↩
- 486.NVIDIA, “NVIDIA Announces NVIDIA Isaac GR00T Reference Humanoid Robot for Academic Research”, May 31, 2026. The announcement does not state the price or whether the body designs are published.↩
- 487.Regulation (EU) 2026/1744, July 8, 2026, published on July 24, recital 42 and Article 3, which amends Article 8 of Regulation (EU) 2023/1230.↩
- 488.Regulation (EU) 2026/1744, July 8, 2026, published on July 24, Article 1, which amends, among others, Articles 2, 5, 25(4), 50, 56(6), 75(1) and 113 of the EU AI Act.abcdefghi
- 489.Regulation (EU) 2023/1230, on machinery, Article 54 and Annex I, Part A. The application date, January 20, 2027, is the one in the corrigendum published in the Official Journal L 169 of July 4, 2023.↩
- 490.Directive (EU) 2024/2853, Articles 5, 6 and 10 and recital 42.ab
- 491.American National Standards Institute, “ISO 10218-1:2025 Robots and Robotic Devices Safety”, accessed in November 2026.↩
- 492.International Organization for Standardization, ISO/CD 25785-1, at committee draft stage, accessed in November 2026.↩
- 493.SESEC, “China's First Standards System for Humanoid Robots and Embodied Intelligence”, April 1, 2026, on the presentation of February 28, organized by technical committee MIIT/TC08 of China's Ministry of Industry and Information Technology.↩
- 494.IBM, blog of the Quantum Developer Conference 2026, November 2026, and Nighthawk press release, November 11, 2026. The claims of advantage are the company's own.ab
- 495.IBM, “IBM Sets the Course to Build World's First Large-Scale, Fault-Tolerant Quantum Computer at New IBM Quantum Data Center”, June 10, 2025.↩
- 496.Quantum Computing Report, “USTC's Zuchongzhi 3.2 achieves below-threshold QEC milestone”, and The Quantum Insider, “China demonstrates quantum error correction using microwaves, narrowing gap with Google”, December 26, 2025, on the paper published in Physical Review Letters on December 22, 2025.↩
- 497.The Quantum Insider, “Microsoft and Quantinuum report on major gains in quantum error correction”, June 13, 2026.↩
- 498.Hua-Liang Liu et al., “Gaussian boson sampling with 1,024 squeezed states in 8,176 modes”, Nature 653, 687-692, May 13, 2026.↩
- 499.Chinese Academy of Sciences, article on Jiuzhang 4.0, May 14, 2026, on the paper published in Nature (DOI 10.1038/s41586-026-10523-6).↩
- 500.South China Morning Post, “Does China's Jiuzhang 4.0 computer herald the age of quantum supremacy?”, May 16, 2026, published by The Star. The estimate in years is the team's own.↩
- 501.Google Quantum AI, “A verifiable quantum advantage”, October 22, 2025.↩
- 502.Emily Conover, Science News, “Quantum 'echoes' reveal the potential of Google's quantum computer”, October 22, 2025.↩
- 503.IBM, “IBM and The University of Chicago Demonstrate Quantum Advantage, Establishing Trusted Quantum Computation on Logical Circuits”, July 30, 2026.↩
- 504.Simon Martiel et al., “Sampling hard circuits with verifiably high fidelity”, July 28, 2026, revised on September 2, 2026. It is a preprint.↩
- 505.Marin Ivezic, PostQuantum, “IBM's Three Quantum Advantage Claims, Fact-Checked”, August 1, 2026.↩
- 506.Morteza Aghaee et al., Microsoft Azure Quantum, “Interferometric single-shot parity measurement in InAs-Al hybrid devices”, Nature 638, 651-655, February 19, 2025.↩
- 507.Matteo Rini, “Microsoft's Claim of a Topological Qubit Faces Tough Questions”, Physics 18, 68, American Physical Society, March 21, 2025, and Fortune, “Microsoft's quantum computing breakthrough questioned by experts”, February 20, 2025.↩
- 508.Henry F. Legg, “On the robustness of topological gap detection via transport”, Nature, June 24, 2026.↩
- 509.Microsoft Quantum, “Reply to: On the robustness of topological gap detection via transport”, Nature 654, E27-E28, June 24, 2026.↩
- 510.Catherine Bolgar, Microsoft Source, “Majorana 2, made more reliable with Microsoft Discovery agentic AI”, June 2, 2026.abc
- 511.Microsoft, “Majorana 2 Explained: The road to 1000x improvement”, video, November 16, 2026.abc
- 512.Morteza Aghaee et al., “20 Second Parity Lifetime in an InAs-Pb Tetron Device”, arXiv, June 2, 2026. It is a preprint.↩
- 513.Davide Castelvecchi, “Microsoft upgrades controversial quantum chip: researchers are still sceptical”, Nature 654, 308-309, June 3, 2026.↩
- 514.Thomas Claburn, The Register, “Boffin claims Microsoft's supposed quantum leap does not compute due to 'basic Python errors'”, June 24, 2026.ab
- 515.Joseph Howlett, Scientific American, “Microsoft's upgraded Majorana quantum computing chip fizzles with physicists”, June 2, 2026.abc
- 516.Hao Zhang et al., “Retraction Note: Quantized Majorana conductance”, Nature 591, E30, March 8, 2021.↩
- 517.Hannah J. Manetsch et al., “A tweezer array with 6100 highly coherent atomic qubits”, Nature, September 24, 2025.↩
- 518.Dolev Bluvstein et al., “A fault-tolerant neutral-atom architecture for universal quantum computation”, Nature, November 10, 2025.↩
- 519.Licenses declared in their GitHub repositories: Qiskit, Cirq, CUDA-Q, PennyLane, Q# and QPanda, accessed November 17, 2026.↩
- 520.IBM, IBM Quantum Platform plans, accessed November 17, 2026.↩
- 521.Physics World, “New findings shorten the road to cryptographically relevant quantum computers”, May 26, 2026, on the paper by Ryan Babbush, Craig Gidney, Dan Boneh and others, from Google, “Securing Elliptic Curve Cryptocurrencies against Quantum Vulnerabilities: Resource Estimates and Mitigations”, arXiv, March 30, 2026.↩
- 522.Trail of Bits, “We beat Google's zero-knowledge proof of quantum cryptanalysis”, April 17, 2026.↩
- 523.Marin Ivezic, PostQuantum, “Google's Secret Quantum Circuits for Breaking ECC Reproduced and Improved in Two Months”, June 2, 2026, on the work of André Schrottenloher, of the Inria center in Rennes.↩
- 524.Google DeepMind and Google Quantum AI, “A scalable and real-time neural decoder for topological quantum codes”, presentation of AlphaQubit 2, arXiv, December 8, 2025, revised on March 11, 2026. It is a preprint.↩
- 525.Volodymyr Sivak et al., Google Quantum AI, “Reinforcement learning control of quantum error correction”, Nature 655, 879-884, July 8, 2026.↩
- 526.The Quantum Insider, “Google study shows quantum computer can learn from its own errors while it computes”, July 10, 2026, on the paper in Nature (DOI 10.1038/s41586-026-10759-2).↩
- 527.NVIDIA, “NVIDIA NVQLink”, October 28, 2025.↩
- 528.The Quantum Insider, “NVIDIA launches Ising”, April 14, 2026.↩
- 529.NVIDIA, model card for Ising-Calibration-1-35B-A3B and model card for Ising-Calibration-1.5-31B on Hugging Face, accessed November 17, 2026.↩
- 530.Tom's Hardware, “Nvidia releases open AI models for quantum computing tasks”, April 15, 2026.↩
- 531.Honeywell, Quantinuum's US$600 million round, September 4, 2025.↩
- 532.Amy Thomson, Bloomberg Law, “Nvidia Backs PsiQuantum in Round at $7 Billion Value”, September 11, 2025.↩
- 533.CNBC, published by NBC San Diego, “Quantum stocks like Rigetti plunge after Nvidia's Huang says the computers are 15 to 30 years away”, January 8, 2025.↩
- 534.NBC News, “Nvidia CEO Huang says he was wrong about timeline for quantum”, March 20, 2025.↩
- 535.Hsin-Yuan Huang, Michael Broughton, John Preskill et al., “Quantum advantage in learning from experiments”, Science, June 2022.↩
- 536.Junyu Liu, Minzhao Liu, Jin-Peng Liu et al., “Towards provably efficient quantum algorithms for large-scale machine-learning models”, Nature Communications, 2024.↩
- 537.Haimeng Zhao, Alexander Zlokapa, Hartmut Neven, Ryan Babbush, John Preskill, Jarrod R. McClean and Hsin-Yuan Huang, “Exponential quantum advantage in processing massive classical data”, arXiv, April 8, 2026, revised on October 1. It is a preprint.↩
- 538.Torsten Hoefler, Thomas Häner and Matthias Troyer, “Disentangling Hype from Practicality: On Realistically Achieving Quantum Advantage”, Communications of the ACM, vol. 66, no. 5, May 2023.↩
- 539.Axel Ciceri, Philip Intallura, Manuel Proissl et al., “Enhanced fill probability estimates in institutional algorithmic bond trading using statistical learning algorithms with quantum computers”, arXiv, September 22, 2025. It is a preprint.↩
- 540.HSBC, “HSBC demonstrates world's first-known quantum-enabled algorithmic trading with IBM”, September 25, 2025.↩
- 541.Scott Aaronson, “HSBC unleashes yet another "qombie": a zombie claim of quantum advantage that isn't”, Shtetl-Optimized, September 25, 2025.↩
- 542.Oliver Knitter, Sang Hyub Kim, Martin Roetteler et al., “Measuring Accuracy and Energy-to-Solution of Quantum Fine-Tuning of Foundational AI Models”, arXiv, May 4, 2026. It is a preprint by authors from IonQ and QuantumBasel.↩
- 543.Hsin-Yuan Huang, Michael Broughton, Norhan Eassa, Hartmut Neven, Ryan Babbush and Jarrod R. McClean, “Generative quantum advantage for classical and quantum problems”, arXiv, September 10, 2025. It is a preprint.↩
- 544.Ewin Tang, “A quantum-inspired classical algorithm for recommendation systems”, arXiv, July 2018, published in the proceedings of STOC 2019.↩
- 545.Quantinuum, “Quantinuum and RIKEN Expand Partnership to Accelerate the Commercialization of Quantum-AI Hybrid Supercomputing in Japan”, PR Newswire, November 10, 2026.↩
- 546.NVIDIA, “NVIDIA Launches Ising, the World's First Open AI Models to Accelerate the Path to Useful Quantum Computers”, April 14, 2026.↩
- 547.Bureau of Industry and Security, interim final rule on quantum computing and semiconductors, September 6, 2024, classifications 4A906, 4D906 and 4E906.↩
- 548.House Select Committee on the Chinese Communist Party, passage of the Remote Access Security Act, January 12, 2026, and Congress.gov, actions on H.R. 2683, roll call vote 13, 369 to 22.↩
- 549.H.R. 2683, Remote Access Security Act, text passed by the House of Representatives on January 12, 2026, Section 2.ab
- 550.Bureau of Industry and Security, additions to the Entity List, May 14, 2024.↩
- 551.National Institute of Standards and Technology, “Department of Commerce announces letters of intent with 9 companies”, May 21, 2026.↩
- 552.Executive Order 14413, “Ushering in the Next Frontier of Quantum Innovation”, June 22, 2026, published on June 25.↩
- 553.Morgan Bazilian and Brad Handler, Payne Institute, “Washington's Quantum Investments”, November 17, 2026.↩
- 554.Outline of the 15th Five-Year Plan adopted by the National People's Congress on March 12, 2026, in the full reproduction by ifeng.↩
- 555.MERICS, “China Tech Observatory: Quantum”, December 2024.↩
- 556.Global Times, “China's latest superconducting quantum computer opens application services to global market”, November 25, 2025, and Tianyan Quantum Group, “Tianyan: Cloud services with quantum advantage”, arXiv, December 11, 2025.↩
- 557.National Institute of Standards and Technology, NIST IR 8547, “Transition to Post-Quantum Cryptography Standards”, October 2026, Tables 2 and 4.ab
- 558.National Institute of Standards and Technology, publication of the first three post-quantum standards, August 13, 2024.↩
- 559.Google, “Tracking the cost of quantum factoring”, May 23, 2025.↩
- 560.Madelyn Cain et al., “Shor's algorithm is possible with as few as 10,000 reconfigurable atomic qubits”, March 30, 2026. It is a preprint.↩
- 561.Dark Reading, “Google sets 2029 deadline for quantum-safe cryptography”, March 2026.↩
- 562.IonQ, third-quarter 2026 results, Exhibit 99.1 to its Form 8-K, November 5, 2026.↩
- 563.Quantinuum, third-quarter 2026 results, November 5, 2026.↩
- 564.24/7 Wall St., “Quantum computing's next winners may not be quantum stocks”, November 10, 2026.↩
- 565.Yann LeCun on the Lex Fridman Podcast No. 416, March 7, 2024, transcript published on the 8th.↩
- 566.Liang Wenfeng, interview with 暗涌 Waves, “疯狂的幻方”, May 23, 2023, copy of the WeChat article on gwern.net.↩
- 567.National Assembly of France, record of Arthur Mensch's hearing before the commission of inquiry on digital dependencies, May 12, 2026.ab
- 568.Dario Amodei, “The Adolescence of Technology”, January 26, 2026.↩
- 569.Fortune, June 11, 2025, on Huang's press conference at VivaTech.↩
- 570.Demis Hassabis, “A Framework for Frontier AI and the Dawning of a New Age”, July 14, 2026.ab
- 571.FINRA, “About FINRA”, accessed November 17, 2026.↩
- 572.36Kr, “揭秘DeepSeek:一个更极致的中国技术理想主义故事”, 暗涌 interview with Liang Wenfeng, July 22, 2024.↩
- 573.Marc Andreessen, “Why AI Will Save the World”, June 6, 2023.ab
- 574.Cyberspace Administration of China and six other agencies, Interim Measures for the Management of Generative AI Services, published on July 13, 2023, in effect since August 15, Articles 2, 7 and 17.abcdef
- 575.Xinhua, speech by Xi Jinping at the World AI Conference, July 17, 2026.↩
- 576.The Paper, April 2024, on Robin Li.↩
- 577.Joel Spolsky, “Strategy Letter V”, Joel on Software, June 12, 2002.abcd
- 578.CNET News.com, “IBM to spend $1 billion on Linux in 2001”, December 12, 2000, reproduced by Linux Today. We did not find IBM's press release.↩
- 579.IBM, “IBM Unveils Next Chapter of watsonx with Open Source, Product & Ecosystem Innovations to Drive Enterprise AI at Scale”, May 21, 2024.abcd
- 580.IBM, announcement of Granite 3.0, 2024.abcd
- 581.IBM, “IBM Announces Availability of watsonx Granite Model Series, Client Protections for IBM watsonx Models”, September 28, 2023.ab
- 582.Microsoft, “The next chapter of the Microsoft-OpenAI partnership”, October 28, 2025.abc
- 583.Microsoft, Form 10-Q for the quarter ended September 30, 2026.abcd
- 584.Alibaba Cloud, “AliViews: Eddie Wu on Alibaba's Q1 earnings”, August 21, 2026, on the quarter from April to June 2026.↩
- 585.U.S. Copyright Office, Copyright and Artificial Intelligence, Part 3: Generative AI Training, pre-publication version, May 2025, with the comments of Andreessen Horowitz, Hugging Face and Meta.abc
- 586.The Verge, May 18, 2020, with the statement by Steve Ballmer from 2001 and the one by Brad Smith at an MIT event.abc
- 587.Free Software Foundation, “About”, accessed November 17, 2026.↩
- 588.Meta, Llama 3.1 Community License, clauses 2 and 5(c).ab
- 589.Free Software Foundation, “What is free software?”, current definition, accessed November 13, 2026.↩
- 590.Open Source Initiative, history of the OSI, accessed November 13, 2026.↩
- 591.Richard Stallman, “Why Open Source Misses the Point of Free Software”, Free Software Foundation, accessed November 13, 2026.↩
- 592.Roger Parloff, “Microsoft Takes On the Free World”, Fortune, May 28, 2007, with statements by Brad Smith and Horacio Gutiérrez, of Microsoft; and an entry by the same author of May 14, 2007.↩
- 593.Microsoft, “Microsoft and Novell Announce Broad Collaboration on Windows and Linux Interoperability and Support”, press release, November 2, 2006.↩
- 594.Free Software Foundation, GNU General Public License, version 3, June 29, 2007, Sections 5 and 6, and Brett Smith, “A Quick Guide to GPLv3”.↩
- 595.Free Software Foundation, GNU General Public License, version 3, June 29, 2007, Section 11.↩
- 596.Brett Smith, “A Quick Guide to GPLv3”, Free Software Foundation.↩
- 597.Apache Software Foundation, Apache License, version 2.0, Section 3.↩
- 598.Apache Software Foundation, Apache License, version 2.0, Section 4.↩
- 599.Microsoft, “Microsoft joins Open Invention Network to help protect Linux and open source”, October 10, 2018.↩
- 600.Reddit, S-1/A prospectus filed with the U.S. Securities and Exchange Commission, March 2024.ab
- 601.U.S. District Court for the Northern District of California, Kadrey v. Meta Platforms, case 23-cv-03417-VC, order on partial summary judgment, June 25, 2025.abcd
- 602.Bloomberg Law, “Anthropic, Authors' $1.5 Billion Deal Receives Final Approval”, July 2026.↩
- 603.U.S. District Court for the Northern District of California, Bartz v. Anthropic, case C 24-05417 WHA, order on fair use, June 23, 2025.ab
- 604.U.S. Court of Appeals for the Third Circuit, Thomson Reuters v. Ross Intelligence, No. 25-2153, September 29, 2026, with footnote 7 on the Department of Justice's statement of interest in In re: OpenAI (S.D.N.Y., September 1, 2026).abcde
- 605.James Gibson, “Risk Aversion and Rights Accretion in Intellectual Property Law”, Yale Law Journal, vol. 116, p. 882, 2007.↩
- 606.Michael Weinberg, “AI licensing deals bad”, August 26, 2024.↩
- 607.U.S. District Court for the Southern District of New York, complaint in The New York Times Company v. Microsoft Corporation and OpenAI, case 1:23-cv-11195, December 27, 2023.↩
- 608.United States Copyright Office, “About”, accessed November 17, 2026.↩
- 609.High Court of England and Wales, Getty Images v. Stability AI, [2025] EWHC 2863 (Ch), November 4, 2025.↩
- 610.Chambers and Partners, “Copyright and AI remain in focus for 2026 with Getty appeal given the green light”, January 20, 2026.↩
- 611.Munich Regional Court I, press release on the judgment in GEMA v. OpenAI, 42 O 14139/24, November 11, 2025.abc
- 612.Regulation (EU) 2024/1689, Articles 2(12), 25(4), 51, 52(2), 53(2) and 54(6) and Recitals 103 and 104, checked against the Psiquepol + AI corpus.abcdefghijk
- 613.European Commission, explanatory notice and template for the public summary of training content, July 24, 2025, note 4.↩
- 614.Japan, Copyright Act, Article 30-4, official English translation.↩
- 615.Original complaint in Kadrey v. Meta Platforms, July 7, 2023, paragraphs 25 and 26.↩
- 616.Hugo Touvron and others, “LLaMA: Open and Efficient Foundation Language Models”, Meta, February 2023.↩
- 617.EleutherAI, “The Common Pile v0.1”, June 2025.↩
- 618.Open Source Initiative, frequently asked questions on the open source AI definition, accessed November 13, 2026.↩
- 619.OpenAI, written evidence to the House of Lords Communications and Digital Committee, December 5, 2023, in the Internet Archive copy.↩
- 620.Nikhil Kandpal and others, “The Common Pile v0.1”, June 5, 2025.↩
- 621.World Intellectual Property Organization, press release PR/2024/921, July 3, 2024.↩
- 622.OpenAI, response to the OSTP request for information on the AI action plan, March 13, 2025.↩
- 623.Kwon Nam-hoon, “Where will AI profits flow?”, Korea JoongAng Daily, August 28, 2026.abc
- 624.ASML, third quarter 2026 results, Form 6-K, October 14, 2026, with the financial statements.ab
- 625.The Next Web, on OpenAI’s third quarter revenue and losses, November 12, 2026; the revenue is according to The Information, and the loss according to documents cited by The Wall Street Journal. OpenAI does not publish its accounts. Margin calculated from those figures.↩
- 626.CoreWeave, Form 10-Q for the third quarter of 2026.↩
- 627.Alphabet, third quarter 2026 earnings release, Exhibit 99.1.ab
- 628.Oracle, Form 10-Q for the quarter from June to August 2026.ab
- 629.Meta, Form 10-Q for the third quarter of 2026.↩
- 630.Rhodium Group, “Examining China's AI Financing”, September 17, 2026.↩
- 631.Torsten Slok, Apollo, “In AI, the 41% Depends on the -59%”, August 7, 2026.↩
- 632.Psiquepol + AI, “The State of Artificial Intelligence Public Affairs Worldwide,” section “The bubble risk,” note 461.↩
- 633.Jeremy Kahn, Fortune, on the circular financing between NVIDIA and OpenAI, September 28, 2025, with the opinion of Stacy Rasgon, of Bernstein.↩
- 634.Ben Lorica, “Open Models Will Absorb Most of the AI Spend”, Gradient Flow, July 22, 2026.↩
- 635.Epoch AI, language model inference price trends.ab
- 636.Gundlach and others, arXiv 2511.23455, November 2025.ab
- 637.Anthropic, API pricing, accessed November 13, 2026.↩
- 638.OpenAI, API pricing, standard rate and short context, accessed November 13, 2026.↩
- 639.Google, Gemini API pricing, accessed November 13, 2026, with the price in effect from December 1.ab
- 640.DeepSeek, API pricing, accessed November 13, 2026.↩
- 641.Z.ai, 2026 interim report, Hong Kong Stock Exchange, September 24, 2026.ab
- 642.Knowledge Atlas Technology (Zhipu), global offering prospectus, Hong Kong Stock Exchange, December 30, 2025.↩
- 643.MiniMax, revision of the caps on connected transactions, Hong Kong Stock Exchange, August 26, 2026. Share of revenue calculated with the annual report.↩
- 644.DeepSeek-AI, “DeepSeek-V3 Technical Report”, December 2024, training costs section.↩
- 645.The Next Web, on DeepSeek’s funding round, May 6, 2026.↩
- 646.Meta, Form 10-K for fiscal year 2025, January 29, 2026.↩
- 647.Meta, transcript of the third quarter 2026 earnings call, October 2026.↩
- 648.Vercel, “AI Gateway Production Index”, October 15, 2026, with data from September.↩
- 649.Ontario Teachers' Pension Plan, investor in Fireworks, press release on its Series D, July 16, 2026.↩
- 650.Ai2, “NSF and NVIDIA award Ai2 a combined $152M”, August 14, 2025.↩
- 651.United States National Science Foundation, “About NSF”, accessed November 17, 2026.↩
- 652.Ai2, “About us”, accessed November 17, 2026.↩
- 653.Forrester, “SaaS As We Know It Is Dead”, February 11, 2026.ab
- 654.Lex Fridman Podcast No. 501, conversation with David Heinemeier Hansson, August 26, 2026. Quotes are from the video’s automatic transcript.abcde
- 655.United States Bureau of Labor Statistics, Occupational Employment and Wage Statistics (OEWS), occupations 15-1252 and 15-1251, May 2021 to May 2025, accessed through its public API.↩
- 656.United States Bureau of Labor Statistics, Occupational Outlook Handbook, software developers, March 2025, October 2025 and November 2026 editions in the Internet Archive.↩
- 657.Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, “Canaries in the Coal Mine?”, Stanford Digital Economy Lab, August 2026 revision.↩
- 658.Microsoft, 2024 annual report, Form 10-K, human capital.↩
- 659.CNBC, “Microsoft is cutting 3% of workers across the software company”, May 13, 2025.↩
- 660.CNBC, “Microsoft laying off about 9,000 employees in latest round of cuts”, July 2, 2025.↩
- 661.Amy Coleman, Microsoft, “The latest in our company transformation”, July 6, 2026.↩
- 662.Satya Nadella, “Recommitting to our why, what, and how”, July 24, 2025.ab
- 663.Fortune, “Microsoft's cloud just hit a new milestone: Azure crosses $100 billion in annual revenue”, July 29, 2026, on Microsoft’s earnings release.↩
- 664.Microsoft, earnings release for the first quarter of fiscal year 2027, October 28, 2026.↩
- 665.Microsoft, “Open Source AI Editor: First Milestone”, VS Code blog, June 30, 2025.↩
- 666.Microsoft, “DeepSeek R1 is now available on Azure AI Foundry and GitHub”, January 29, 2025, and gpt-oss on Azure AI Foundry, August 5, 2025.↩
- 667.Anthropic, knowledge-work-plugins repository, with the legal plugin added on January 29, 2026 under the Apache 2.0 license.↩
- 668.Cloud Wars, transcript of the BG2 podcast with Satya Nadella, December 2024.↩
- 669.CNBC, on the drop in software stocks, February 6, 2026.↩
- 670.diginomica, on Klarna, Salesforce and Workday, March 7, 2025.↩
- 671.The Register, on Salesforce’s results, February 27, 2025.↩
- 672.Salesforce, annual report for fiscal year 2026, Form 10-K, employees as of January 31, 2026.↩
- 673.Bain & Company, “Will Agentic AI Disrupt SaaS?”, Technology Report, September 23, 2025.↩
- 674.Daniel Stenberg, “The end of the curl bug-bounty”, January 26, 2026.↩
- 675.Daniel Stenberg, “A new breed of analyzers”, October 10, 2025.↩
- 676.The Star, with Bloomberg, “Samsung bans staff's AI use after spotting ChatGPT data leak”, May 2, 2023.ab
- 677.Cisco, 2024 data privacy benchmark study, January 2024, 2,600 privacy and security professionals in 12 countries.ab
- 678.NBC News, on JPMorgan Chase’s AI assistant, 2024.abc
- 679.Fortune, on Goldman Sachs’s internal AI assistant, June 24, 2025.abc
- 680.The Korea Times, “Samsung Electronics to introduce ChatGPT, Gemini, Claude in operations”, June 11, 2026, with the company’s statement.abc
- 681.KPMG and the University of Melbourne, “Trust in AI” 2025, U.S. results, survey from November 2024 to January 2025.ab
- 682.Procter & Gamble, “Expectations of P&G Suppliers Using Artificial Intelligence (AI) in their P&G Work”, November 2024.abc
- 683.Kenvue, data safeguards requirements annex for suppliers, January 2026.ab
- 684.Clifford Chance, “AI principles for suppliers”, April 22, 2025.ab
- 685.White House Office of Management and Budget, memorandum M-25-22, April 3, 2025.abc
- 686.European Commission, public buyers community, “Updated EU AI model contractual clauses”, accessed November 17, 2026.↩
- 687.European Commission, public buyers community, model clauses for the procurement of high-risk AI, February 2025 version, working draft without an official Commission position.↩
- 688.European Union, Regulation (EU) 2022/2554 on digital operational resilience for the financial sector, Articles 2 and 64.↩
- 689.United States, 45 CFR 164.502(e), disclosures to business associates, in the Electronic Code of Federal Regulations (eCFR), current text accessed through its API on November 17, 2026.↩
- 690.United States, 45 CFR 164.504(e)(2), contents of business associate contracts, in the Electronic Code of Federal Regulations (eCFR), current text accessed through its API on November 17, 2026.↩
- 691.State Bar of California, practical guidance on generative AI, approved May 14, 2026.ab
- 692.Shijie Wu and others, “BloombergGPT: A Large Language Model for Finance”, March 2023.ab
- 693.Samsung, Samsung Developer Conference Korea 2024 and Gauss2, November 28, 2024.ab
- 694.NTT, announcement of tsuzumi 2, October 20, 2025.ab
- 695.OpenAI, “Data controls in the OpenAI platform”, accessed November 13, 2026.ab
- 696.Anthropic, commercial terms, in effect since June 17, 2025.↩
- 697.Google Cloud, zero data retention in Vertex AI, updated November 12, 2026.abc
- 698.AWS, Amazon Bedrock frequently asked questions, accessed November 13, 2026.ab
- 699.AWS, abuse detection in Amazon Bedrock, accessed November 17, 2026.ab
- 700.Matt Bornstein, Guido Appenzeller and Martin Casado, Andreessen Horowitz, “Who Owns the Generative AI Platform?”, January 19, 2023.↩
- 701.Andreessen Horowitz, second survey of large-enterprise technology executives, June 10, 2025. The firm invests in AI labs.↩
- 702.Andreessen Horowitz, third survey of large-enterprise technology executives, January 30, 2026, survey of 100 Global 2000 executives. The firm invests in AI labs.↩
- 703.Bureau of Industry and Security, “Framework for Artificial Intelligence Diffusion”, January 15, 2025, 90 FR 4544, pp. 4554 and 4555.abcd
- 704.Bureau of Industry and Security, announcement of the rescission of the AI diffusion rule, May 13, 2025.↩
- 705.Export Administration Regulations, 15 CFR § 742.6(a)(13) and (b)(14), § 734.9(l) and note 2 to classification 4E091 of Supplement No. 1 to Part 774, text in effect as of November 9, 2026, checked against the Psiquepol + AI corpus.abcd
- 706.Government Accountability Office, decision B-337935, May 12, 2026, summarized by Holland & Knight, June 5, 2026.ab
- 707.United States, 31 U.S.C. § 702, subsection (a).↩
- 708.Government Accountability Office, decision B-337935, May 12, 2026, full text: the petition from Senator Elizabeth Warren, dated November 19, 2025 (note 2), and the summary of the Congressional Review Act (section “Congressional Review Act”). Accessed November 17, 2026.↩
- 709.Unified Regulatory Agenda, RIN 0694-AJ90, fall 2025.↩
- 710.Psiquepol + AI, “The State of Artificial Intelligence Public Affairs Worldwide,” section “What the state can demand,” notes 638 and 639.↩
- 711.Psiquepol + AI, “The State of Artificial Intelligence Public Affairs Worldwide,” section “The gap with China,” note 656.↩
- 712.China, Data Security Law, adopted June 10, 2021, in effect since September 1, 2021, Article 21.↩
- 713.China’s Ministry of Commerce, spokesperson’s answers on Announcement No. 48, October 16, 2026, including local deployment abroad.abc
- 714.Regulation (EU) 2024/1689, Articles 53(1), points (a) and (b), 54(1) and 55(1).abc
- 715.European Commission, General-Purpose AI Code of Practice, transparency chapter (commitment 1) and copyright chapter (measure 1.4), checked against the Psiquepol + AI corpus.↩
- 716.European Commission, guidelines on the obligations of general-purpose model providers, document C(2025) 7719 of November 19, 2025, whose content it approved on July 18, §§ 60, 80 and 83. They are not binding.ab
- 717.Illinois, SB 315, Public Act 104-0538, signed July 6, 2026, in effect January 1, 2027, with audits from January 1, 2028, checked against the Psiquepol + AI corpus.abc
- 718.Illinois, SB 315, Artificial Intelligence Safety Measures Act, Public Act 104-0538, sections 5, 10(a) and (d), 18 and 99, in the enrolled text preserved by the Internet Archive, because the Illinois General Assembly’s site does not allow automated reading. Accessed November 17, 2026.↩
- 719.H.R. 10981, American AI Security Act, introduced by Josh Gottheimer and Mike Lawler on November 10, 2026 and referred to the Committees on Energy and Commerce and on Intelligence, section 2(a), (c), (e), (f), (h) and (i)(7). Checked against the Psiquepol + AI corpus.abcd
- 720.Cyberspace Administration of China and three other agencies, Provisions on the Administration of Algorithmic Recommendations in Internet Information Services, published December 31, 2021, in effect since March 1, 2022, Articles 24 and 25.↩
- 721.Tony Peng, “Friendlier China's Regulation on Generative AI”, Recode China AI, July 17, 2023.↩
- 722.Federal Senate, final text of PL 2338/2023, December 2024, arts. 30, § 1, and 73, pending in the Chamber of Deputies, assigned to committees on September 2, 2026. Checked against the Psiquepol + AI corpus.↩
- 723.Psiquepol + AI legal corpus, accessed November 17, 2026: South Korea’s AI Basic Act (46 articles), Japan’s Act on the Promotion of Research and Development and the Utilization of AI-Related Technologies (28) and Italy’s Law 132/2025 (28), with no mention of open source or of weights.↩
- 724.Directive (EU) 2024/2853, Articles 2(1), 2(2), 4, 8 and 22 and Recital 14, checked against the Psiquepol + AI corpus.abcde
- 725.Philipp Hacker and Matthias Holweg, “The regulation of fine-tuning: Federated compliance for modified general-purpose AI models”, Computer Law & Security Review, vol. 60, 2025.ab
- 726.Faegre Drinker, “Transposing the EU New Product Liability Directive: A Member State Progress Report (October 2026 Update)”, October 30, 2026.↩
- 727.California, SB 1047, enrolled text, September 3, 2024, and veto message, September 29, 2024.↩
- 728.United States, 47 U.S.C. § 230, subsection (c)(1).↩
- 729.Congressional Research Service, “Section 230 Immunity and Generative Artificial Intelligence,” LSB11097, December 28, 2023.↩
- 730.Crowell & Moring, August 3, 2026, on Forrest v. Meta and Bouck v. Meta.ab
- 731.Dan Mangan and Ashley Capoot, CNBC, “California sues OpenAI over Hugging Face, other AI model loss of control”, November 16, 2026.↩
- 732.Public Law 119-60, fiscal year 2026 defense act, sections 1532, including its subsections (a)(2) and (a)(3)(B), and 6604, December 18, 2025.abc
- 733.Fortune, “Pentagon accuses Alibaba, Baidu, BYD of being Chinese military companies”, June 8, 2026, on the update of the section 1260H list.ab
- 734.Department of Homeland Security, order to exclude and remove DeepSeek products, July 24, 2026, published in the Federal Register on October 20, 2026, 91 FR 66346.ab
- 735.Congressional Research Service, IF13336, October 27, 2026.abcd
- 736.Crowell & Moring, October 21, 2026.abc
- 737.Gregory Huh, Haynes Boone, on the DeepSeek exclusion and removal order, October 23, 2026.↩
- 738.Governor of Virginia, Executive Order 46, February 11, 2025 (archived copy); for Texas, Fox 7 Austin, January 31, 2025, and for New York, Spectrum News, February 10, 2025.↩
- 739.Bureau of Industry and Security, addition to the Entity List, January 16, 2025, 90 FR 4617.↩
- 740.United States National Security Agency, Federal Bureau of Investigation and Cybersecurity and Infrastructure Security Agency, joint advisory AA26-251A, September 8, 2026.↩
- 741.S. 321, Decoupling America's Artificial Intelligence Capabilities from China Act of 2025, January 29, 2025, section 2(2).ab
- 742.H.R. 10670, China FIREWALL Act of 2026, October 1, 2026, which refers to 42 U.S.C. 18741(a).abc
- 743.United States, 42 U.S.C. § 18741, subsection (a)(5)(C).↩
- 744.H.R. 4142 and S. 2177, No Adversarial AI Act, June 25, 2025.ab
- 745.Psiquepol + AI, “The State of Artificial Intelligence Public Affairs Worldwide,” section “The gap with China,” note 659.↩
- 746.Italy’s Data Protection Authority (Garante per la protezione dei dati personali), press release on DeepSeek, January 30, 2025.↩
- 747.South Korea’s Personal Information Protection Commission, corrective recommendation to DeepSeek, April 24, 2025.↩
- 748.Berlin Commissioner for Data Protection, report of the DeepSeek app to Apple and Google, June 27, 2025.↩
- 749.Government of Australia, federal directive PSPF 001-2025, February 4, 2025, at protectivesecurity.gov.au, which did not respond when we tried to access it; we consulted it in the Victorian government directive that applies it, updated February 14, 2025.↩
- 750.Czech National Cyber and Information Security Agency, warning about DeepSeek products, July 10, 2025, and English press release, with the government resolution of July 9.↩
- 751.Taiwan’s Executive Yuan, press release of February 3, 2025, and Administration for Cyber Security, notice of February 20, 2025.↩
- 752.European Data Protection Board, “EDPB Chair”, accessed November 17, 2026.↩
- 753.European Data Protection Board, minutes of the plenary of February 11, 2025.↩
- 754.European Commission, guidelines on the transparency obligations of Article 50 of the AI Act, C(2026) 7702 final, November 5, 2026, paragraphs 23, 24 and 27.ab
- 755.European Commission, code of practice on transparency of AI-generated content, June 10, 2026, measures 1.1 and 1.2.ab
- 756.United States, TAKE IT DOWN Act, Public Law 119-12, May 19, 2025.↩
- 757.United States Federal Trade Commission, “FTC Begins Enforcing the TAKE IT DOWN Act”, May 2026.↩
- 758.California, SB 1000, Chapter 861 of 2026, approved September 30, 2026, legislative counsel’s summary.↩
- 759.California, SB 1000, September 30, 2026, Bus. & Prof. Code § 22757.1(e) and § 22757.3, in the same reproduction.↩
- 760.Latham & Watkins, on California’s 2026 AI laws.↩
- 761.United Kingdom Home Office, circular 004/2026 on the Crime and Policing Act 2026, section 99.↩
- 762.United Kingdom Home Office, factsheet on intimate image abuse.abcd
- 763.Lorna Woods, Online Safety Network, “The Crime and Policing Act 2026: An Online Safety Perspective”, May 14, 2026.↩
- 764.Cyberspace Administration of China and other agencies, Measures for Labeling AI-Generated Synthetic Content, March 14, 2025, in effect from September 1, Articles 2, 4, 5 and 10.↩
- 765.Internet Watch Foundation, “About us”, accessed November 17, 2026.↩
- 766.Internet Watch Foundation, Harm without limits, March 24, 2026.↩
- 767.Center for Countering Digital Hate, according to Decrypt, January 22, 2026.↩
- 768.European Commission, proceeding against X over Grok under the Digital Services Act, January 26, 2026.↩
- 769.Regulation (EU) 2022/2065, Digital Services Act, Articles 34, 35 and 74.↩
- 770.United Kingdom Competition and Markets Authority, AI Foundation Models: Update Paper, April 11, 2024.↩
- 771.United States Federal Trade Commission, Office of Technology, “On Open-Weights Foundation Models”, July 10, 2024, Internet Archive copy; the original page has redirected since September 2025.↩
- 772.TechCrunch, October 20, 2025.↩
- 773.United States Federal Trade Commission, report of the 6(b) study on AI partnerships, January 17, 2025, Internet Archive copy.↩
- 774.United Kingdom Competition and Markets Authority, European Commission, United States Department of Justice and United States Federal Trade Commission, joint statement on competition in generative AI, July 23, 2024.↩
- 775.Microsoft, Form 10-K for fiscal year 2026, filed July 29, 2026, note on investments in OpenAI.ab
- 776.CNBC, October 28, 2026, on a Reuters report.↩
- 777.Psiquepol + AI, “The State of Artificial Intelligence Public Affairs Worldwide,” section “Agents and cybersecurity,” notes 504 and 505.↩
- 778.Global Competition Review, September 8, 2026, and The Capitol Forum, November 5, 2026.↩
- 779.European Commission, IP/26/1276, interim measures against Meta over AI assistants’ access to WhatsApp, June 9, 2026.ab
- 780.S. 5105, Collaboration on Adversarial Threats and Security Risks Act, introduced by Adam Schiff and Jim Banks on July 23, 2026, sections 2 to 4; referred to the Senate Committee on the Judiciary.↩
- 781.Constitution of the United States, Article VI, National Archives transcript.↩
- 782.Nextgov, July 1, 2025.ab
- 783.Executive Order 14365, December 11, 2025, 90 FR 58499, sections 3, 5, 7 and 8.↩
- 784.Department of Justice, intervention in xAI’s lawsuit against Colorado’s law, April 24, 2026.ab
- 785.X.AI Corp. v. Weiser, 1:26-cv-01515 (D. Colo.), summary and docket at the Civil Rights Litigation Clearinghouse, order of April 27, 2026, accessed November 13.↩
- 786.Colorado, SB24-205, Consumer Protections for Artificial Intelligence, official bill summary.↩
- 787.Colorado, SB26-189, Automated Decision-Making Technology, signed May 14, 2026.ab
- 788.Letter from state legislators against the draft, June 16, 2026, about 203 signatures by our own count.↩
- 789.United States Federal Trade Commission, “Proposed Policy Statement Concerning the Suppression of Accuracy in Artificial Intelligence Systems”, July 1, 2026, sections I and IV and notes 12 and 43.↩
- 790.United States Federal Trade Commission, proposed statement on the suppression of accuracy in AI systems, July 1, 2026; its reading on the preemption of state laws, according to Arnold & Porter.↩
- 791.Anthropic, donation to Public First Action, February 12, 2026.↩
- 792.Anthropic, second donation to Public First Action, July 21, 2026.↩
- 793.Al Jazeera, November 4, 2026, with the AP count.ab
- 794.Time, November 4, 2026.↩
- 795.Harvard Business School AI Institute, “Open Weights and the New Economics of Inference”, August 3, 2026.abc
- 796.Meta, Llama Protections, accessed November 17, 2026.↩