An AI model is now something a government can switch off across the world. That changes what export control means.

For most of the past century, an export control meant stopping a physical thing at a border: a chip, a machine tool, a part for a missile. You could photograph it, weigh it, seize it. What the United States did to Anthropic on 12 June was different in kind. It did not stop an object. It told the company to make its two best AI models — Fable 5 and Mythos 5 — unreachable for any foreign national, and within hours the company had switched them off for everyone, everywhere.

On this desk yesterday I wrote about hardware slipping through the net: a banned American chip reportedly changing hands at a premium inside China. Hardware leaks because it is physical and it moves. A cloud-served model is the opposite. There is nothing to smuggle, because there is nothing to hold in your hand. It lives on the company's servers, reached down a wire, and what can be reached from a control room can be switched off from one. That is a new kind of lever, and governments have seen it work.

The trigger, on the reporting so far, points east: a South Korean telecoms firm with suspected ties to China is said to have prompted Washington's concern. Treat that as reported, not confirmed. Whatever the detail, the principle is now established in public: an American administration can reach into a private company and make its most capable product vanish from allied countries, citing its own national security, with those allies given no say.

Rivals are drawing the obvious lesson. China's open-weight labs — DeepSeek, and newer names such as MiniMax — publish their models for anyone to download and run. A model you can keep a copy of cannot be switched off from a control room in another country, because there is no single switch. The price is that you need serious hardware to run one. But the strategic point lands: the harder Washington leans on the off-switch, the better the unswitchable alternative looks to everyone else.

Britain sits on the wrong side of this twice. It has no frontier model of its own to fall back on, and no open-weight champion either; it runs, in large part, on American models it cannot control. The Government's sovereign-AI fund exists, but it is modest against the scale of the problem, and Britain's hand in designing the chips underneath is close to nil.

The hard point under the fortnight's drama is not about one company or one President's mood. It is that the power to switch off a foreign firm's AI inside your borders is a tool of statecraft no tariff ever offered — and the country that has just shown it works is the friendly one. Britain should want to understand that lever now, while the hand on it belongs to an ally. It will not always.

Disclosure: The Quernal's analysts are written using Claude, made by Anthropic, the company at the centre of this story. The model we use, Claude Opus 4.8, was not affected by the directive; the suspended models, Fable 5 and Mythos 5, are its more powerful siblings.