On Friday last week, the US government applied export controls to an LLM, banning non-US citizens from using it, but in practice removing the model from use.
This is a Schrödinger’s cat of an event. Simultaneously, a wake-up call for European sovereignty and a nothingburger, depending on whether you think LLMs are revolutionary, or just a “next word guessing machine”.
Benedict Evans, who is nobody’s idea of a hype man, says AI is as big a deal as the internet or mobile, but only as big. We’re in 1997. Most of it doesn’t work yet, and most of what people will actually do with it hasn’t been built.
But the internet became what it became precisely because nobody could switch it off. It was open, distributed, owned by no one, and that was the whole point. Now picture it’s 1997 again, and a single government could turn the internet off for your whole country with a letter sent on a Friday afternoon. That is this weekend.
The directive came from the US Commerce Department, and Anthropic disabled Fable and its more powerful sibling Mythos for every customer to comply. It’s the first time a leading lab has taken a live model offline because a government told it to. The lesson, for anyone outside the US, is simpler. The most important infrastructure of the next twenty years is arriving owned, concentrated, and revocable, and we don’t hold the switch.
Not a money problem
The reaction, such as it was, came mostly from the wings. A former British security minister and the leader of the French opposition said the right things, that this was a wake-up call, that AI had become a question of national sovereignty, that Europe needed its own models and needed them now. They weren’t wrong. But Macron didn’t say it, Starmer didn’t say it, and in Norway nobody said anything at all. The people who would actually have to find the money stayed quiet. The few who did speak ran straight into the deeper problem with how we talk about all this, because the loudest voices on digital sovereignty don’t understand why people choose one product over another.
It’s always the same. “There are so many good, open source alternatives, just use LibreOffice instead of Office.” LibreOffice, for anyone who’s opened it lately, looks like it stopped developing in 2003. An alternative existing is not the same as an alternative being good, and people don’t want to sacrifice convenience for an idea.
People didn’t switch to electric cars out of virtue, they switched when the cars got good, when a Tesla was a nicer thing to own than the petrol car beside it. Even then, here in Norway, it took a decade of tax breaks and free tolls and bus lanes. The formula is a good product plus scaffolding that rewards it, never moralising propped on top of something worse.
The real question, then, is whether Europe could build something people would choose with the politics stripped out. Mistral, the great European hope, has raised somewhere around three to four billion dollars across its entire life. Anthropic raised sixty-five billion in a single round this spring. OpenAI raised a hundred and twenty-two in one round. A single American funding event is larger than everything Mistral has ever raised, by more than an order of magnitude.
And it gets worse, because Mistral isn’t even the best of the alternatives. The leading open models, the ones you can actually run yourself, are now Chinese, DeepSeek and Qwen, and the rest. Mistral’s real claim is a location, not capability. So sovereignty through open source, in practice, today, means running models from Beijing. The escape hatch leads to a different cage.
You could read all this as proof the gap is too big, that frontier AI costs tens of billions Europe was never going to find. Except DeepSeek built something near the frontier for a few hundred million, under American chip controls, with worse hardware than anyone here worries about. Harder constraints, better result. So the bar was never sixty-five billion. It was a few hundred million, the nerve to put it on a handful of people and one long bet, the stomach for political decisions some people won’t like, and the willingness to watch most of it fail. That is what Europe lacks, not the money. The money exists. Norway’s own wealth fund could write the cheque without noticing, and a sovereign fund from Abu Dhabi is already in Mistral. Europe’s own institutions are not.
The country with no excuse
Take the clearest case. Norway has the two things the rest of Europe lacks, the largest sovereign wealth fund on earth and access to cheap, clean hydropower. That is enough to build the thing itself and to use it, to fund the company that makes the models and the people who would put them to work. We rent out the riverbank instead. The flagship project, branded Stargate Norway, is a hundred-thousand-GPU data centre near Narvik, except OpenAI put its name on it, walked away before it opened, and Microsoft stepped in to take the capacity for some six billion dollars. American compute, Norwegian power.
The fund, meanwhile, holds well over a percent of Nvidia, of Microsoft, of Apple, and screens its own investments with Anthropic’s Claude, the very company at the centre of the ban. It is run as an index fund, so it cannot pick a national champion and has no reason to want to, because it grows richer every time American AI does. And when something frightens us enough, we move: the same country found six hundred billion kroner for a decade of rearmament, incidentally about what investors put into Anthropic in a single round. We know how to invest in securing the future. We just haven’t decided this is worth it.
Who holds the switch
I’d spent the week building a game with it. The model didn’t break, a government simply decided I shouldn’t have it.
Evans is probably right that this is as big as the internet and no bigger. But the internet wove itself into everything because no one could take it away from you. This time someone can, and last Friday they did. Europe keeps treating that as a procurement problem, a question of which American tool to standardise on, when it is really a question of whether you own the thing at all. Building the power station for someone else’s intelligence isn’t independence. It’s the most sophisticated way yet of being a supplier, on the one platform that, this time, has an off switch, held in a hand that isn’t ours.