There is no side to take on frontier AI

By Richard Stirling

Within four days of each other, the US and the G7 put out documents pointed in seemingly opposite directions on how to handle the AI frontier. 

The US order, signed on 2 June, treats the AI frontier as a security problem. It introduces a classified process to grade the most capable models for cyber capability, and asks labs to volunteer early sight before release and accept a government hand in who gets access. Days earlier, in Paris, the G7 approved a vision on AI openness: open weights, interoperability, shared language, diffusion as wide as it could manage.

They are not two answers to the same question, because they are not regulating the same thing. Both say they are managing risk; the instruments point different ways. The US seeks to gate the frontier; the G7 to spread it. So when governments that are neither the US nor the G7 ask me which way to lean, there is no side to take. The two documents aren’t having the argument they appear to be having.

For governments that are neither, the useful question is not which document to align with. It is the one from my earlier piece about how ‘Sovereign AI is the wrong framing’: which links of your own chain you take, hedge, and build. And the links are not moving the same way.

Inference gets cheaper every month. A dependency you have to take this year you can hedge next year, and own the year after if you still want it. There is no prize for owning it early; paying to secure it now is buying at the top of the market.

Frontier training is going the other way. It is concentrating, not spreading, and for almost every country it is a take whether they like it or not. The ambition to own that link is the most expensive way to learn this.

Which leaves the two links that never commoditise: the talent to make these choices, and your own ability to run AI inside government. These do not get cheaper, do not concentrate somewhere else, and cannot be bought in finished. Most “sovereign AI” money goes to compute you could have rented and away from the two things you can’t. That is the decision worth getting right, and it is rarely the one being argued about.

So redo the map every year. The cheap links and the chokepoints keep trading places, and a chain you mapped once is already out of date.

Grid image by Claudio Schwarz on Unsplash