The Story

Say it out loud and it sounds paranoid. Could the US government switch off another country’s artificial intelligence overnight? Then you look at how deeply the models are being wired into everything, and the paranoia starts to feel like planning.

A closed model like Claude or GPT is a rental. You reach it through an API that lives on American servers, under American law and one company’s terms of service. Pull that thread and whatever you built on top stops working the same afternoon.

Why It Matters

Washington has already shown it treats compute as a lever. It has restricted advanced chips to China, built a tiered framework deciding which countries get how much AI hardware, and leaned on allies to fall in line. A model’s access sits inside that same machinery. An export rule, a sanction, a sudden reclassification: any one of them could cut a foreign government or company off from the frontier model it had quietly made load-bearing.

And load-bearing is exactly what these systems are becoming. Tax agencies, hospitals, banks and defence ministries are threading large language models into core workflows. The more useful the model, the deeper it goes, and the more it hurts when someone else’s hand is on the switch.

The Distinction That Matters

There are two kinds of model, and they carry opposite risks — a distinction the panic almost always skips.

Open-weight models, the Llama and Mistral and DeepSeek family, can be downloaded and run on your own machines. Once the file is on your servers, no one can revoke it. The trade is that you are often a step behind the very best systems, and you inherit the cost and burden of running and securing them yourself.

Closed frontier models give you the sharpest capability and a clean API, and in exchange you accept a dependency you do not control. So it comes down to which risk you would rather carry: the capability you give up by running your own, or the control you give up by renting someone else’s. Every government and serious company is now quietly making that call.

What’s Next

“Sovereign AI” is about to graduate from conference slogan to budget line. Expect national models, state-funded compute, local hosting mandates, and procurement rules that demand a fallback if the foreign tap closes. France has Mistral, the Gulf states are pouring money into their own systems, India and others are building. Some of it is genuine resilience. A good deal of it is industrial policy wearing a security badge.

The countries that come out of this well will not be the ones with the cleverest chatbot. They will be the ones who assumed, from the start, that the switch was never theirs to hold.