Study · Psiquepol + AI

Open and Closed Models: The Dispute That Shapes AI

A study by Psiquepol + AI, published together with “The State of Artificial Intelligence Public Affairs Worldwide”. Each can be read on its own, and the two complement each other. November 2026 edition. Data cutoff: November 17, 2026. NVIDIA’s quarterly figures are from the third quarter of its fiscal year 2027, August to October, unless otherwise noted.

This study is about who makes artificial intelligence models, who funds them, which are published and which are held back, and how the law treats each kind. “The State of Artificial Intelligence Public Affairs Worldwide”, published at the same time, deals with what limits access to AI: chips, energy, data centers, export controls and public opposition. When a fact is already in that study, we do not repeat it and instead refer to its section and its note, using the numbering of the published version. This study draws on the legislation and treaties gathered by the Psiquepol + AI platform and on the information we monitor, and each note gives its source.

In the AI supply chain, a rule written by a company weighs as much as a country’s law. That rule applies in every country where its author operates. It is enforced without a prior trial, simply by cutting off service, and it reaches governments and foreign companies that no national law can reach. When Microsoft cut off its services to an Israeli military unit, Yotam Berger, a doctoral candidate in juridical science at Stanford Law School, wrote that the case gives big tech companies a practical capacity to regulate surveillance, and warned of the risk that a private company acts as the de facto regulator of those programs [1]. Peter Henderson and Mark Lemley observe, for their part, that labs enforce their terms by suspending accounts rather than by going before a judge [2].

The private rule includes a model’s license, a lab’s usage policy, a cloud’s terms of service, the clauses a large company imposes on its suppliers and the long-term contracts through which the capacity to make chips and memory is allocated. Added to it are international treaties and recommendations, the laws of each region, country and state, the regulators of each sector, technical standards and the industry’s own commitments. Each of those rules leaves winners and losers.

Artificial intelligence is developed along two paths, closed models and open models. The difference between them decides who gets paid for AI, who controls it, what a government can monitor and what it can ban. A model is the program that generates text, images or code from a request. A closed model can only be used through its maker’s service, and that developer decides who gets in, what it charges, what it monitors and when it changes or withdraws the model.

An open model, by contrast, publishes its weights, which are also called parameters. The weights are the numerical values that hold everything the model learned during its training, the stage in which it processes enormous amounts of data on thousands of chips over weeks or months. A large model has billions of weights. With them, anyone can download the model, modify it and run it on their own hardware without going through its developer. Once it is published, no one can withdraw it.

The line between open and closed follows business as much as the rivalry between the United States and China. Those who earn by selling access to a model end up on one side, and those who earn when the model spreads end up on the other. China bets almost entirely on open. In the United States, the government, NVIDIA, Meta, Microsoft, Google and OpenAI back open weights. In July, a public letter signed by technology companies called for support for open weights (see “The open camp in the United States”), and of the large labs only Anthropic stayed off it. xAI did not sign it under its own name, but SpaceX, which has owned it since February, did.

No law in force in a major power bans publishing weights. Restrictions fall on the rival’s models, such as DeepSeek’s in U.S. federal systems, and on the largest closed models. On November 16, the U.S. Department of Commerce released a rule that eliminates the license that had been required to export the weights of the largest closed models, a control the department had already announced it would not enforce. The same rule requires a license to provide compute, that is, to let others use a data center’s chips remotely, to companies in China and in other countries under arms embargo. It takes effect on its publication in the Federal Register, the federal government’s official journal, on November 19.

In June, the same department had ordered Anthropic to cut off foreigners’ access to two of its models, and it lifted the order at the end of that month. China, for its part, has asked since October that the weights of its largest models be registered before they are exported, except those that are already public. The three measures are developed in “How the law treats each model”.

Kuuntele Spotifyssa

How to navigate: the contents list takes you to each section and marks the one you are reading in purple. Bracketed numbers, such as [12], take you to their note in “Notes and sources”; in the note, the ↩ arrow returns to the place in the text and, if the note is cited more than once, each letter or number next to it returns to one of the citations.