The diplomatic split now forming between Washington and Beijing over AI turns on a technical distinction most coverage leaves unexplained: the difference between open-weight and proprietary models. It is worth understanding, because it determines what each side is actually offering other countries.

What “weights” are

A modern AI model is, at bottom, an enormous list of numbers. Training adjusts those numbers — the weights — until the system produces useful outputs. The finished model is essentially that numerical file, which can run on suitable hardware anywhere.

A proprietary model keeps that file private. You reach it over the internet, through the developer’s servers, and pay per use. The company controls who has access, what the model will refuse to do, when it is updated and when it is withdrawn.

An open-weight model publishes the file. Anyone can download it, run it on their own machines, adapt it, and use it without asking permission or reporting usage.

One clarification, because the terms get mixed up. Open-weight is not the same as open-source. Open-source software conventionally means the recipe is public — here, the training data and methods. Most “open” AI models publish only the finished weights, not what went into them. You can use the result; you cannot reproduce or fully audit it.

Why this is geopolitical

Export controls on advanced chips rest on an assumption: that capability follows hardware, so restricting the chips restricts what a rival can build.

Open weights weaken that assumption. Training a frontier model needs vast computing power, but running one afterwards needs far less. A country that cannot train its own system can still download a capable one, and once a file is published it cannot be recalled. There is no mechanism for un-releasing weights.

That is why Beijing’s promotion of open-weight models functions as a strategy rather than a licensing preference. It offers countries something that does not depend on continued goodwill from a foreign supplier — no metered access, no risk of being cut off during a dispute. For a state that has watched cloud services and chip supplies used as leverage, that is a substantive offer.

The genuine trade-off

Neither approach is straightforwardly better, and the argument between them is not settled among researchers.

The case for open weights: independent scientists can examine and test the system rather than taking a company’s word for its properties; smaller countries, universities and firms are not locked out by price; capability is not concentrated in a handful of corporations; and work can proceed without an internet connection to a foreign server.

The case against: safety measures built into a downloadable model can be stripped out by anyone with the file and modest expertise. A proprietary model can be monitored, rate-limited or shut off if misused. An open one cannot be recalled — a point that matters more as models become capable in areas like cyber operations, where autonomous capability is a documented concern.

The honest summary is that open weights distribute both capability and risk, and you cannot choose one without the other.

The inconsistency on both sides

Neither bloc is a pure case. American companies have released open-weight models alongside their proprietary flagships. And Reuters reported in July that Beijing is considering restrictions on overseas access to its leading models — which would sit awkwardly beside a pitch built on openness.

So the choice being put to 35 governments is less about software licensing than about dependency: whether to accept access on an ally’s terms, or capability with a different set of strings that have not yet been fully specified. Framed that way, “choose deliberately” is advice worth taking literally.

Sources

Reuters reporting on the Pax Silica letter and on Chinese model access restrictions; published technical documentation from major model developers; academic literature on open-weight release and safety.