Europe's Fable: The Off Switch Was Never Ours
Someone in Europe shipped a product in early June with Fable 5 sitting underneath it.
Not a toy integration. Months of work had been built around it: prompts, evaluations, routing logic, latency budgets, cost projections, demos, internal promises. The usual quiet bargain you make when you build on a frontier model, that the thing answering on Friday will still be there on Monday.
Then it was not.
On June 12, Anthropic disabled foreign access to Fable 5 and the larger Mythos 5 after the U.S. government placed both models under export controls. Not a price change, not a deprecation window, not even the familiar enterprise ritual of a migration guide and a vaguely apologetic email. A government order.
The official reason was national security. The reported trigger was concern that the safeguards around the models could be bypassed. Anthropic pushed back, arguing it had not been shown evidence of a universal jailbreak and that the issue had been described too narrowly and too vaguely for the company to respond properly. It did not matter. The order held, and the frontier went dark for people outside the United States.
So this was not Europe being blocked from a competitor, and it was not a sanctions case in the ordinary sense. These were customers, teams, companies paying full price for infrastructure they had already wired into their products, who woke up to find the model in their stack had a political jurisdiction hidden inside it.
What got me was not that the Americans did it. Of course they did. A frontier model is strategic infrastructure now, and governments fence off strategic infrastructure. What got me was how completely Europe just had to eat it. There was no appeal path, no sovereign override, no "we are allies, surely this does not apply to us" clause that meant anything in practice. If your product depended on Fable on Thursday, Washington had edited your roadmap by Friday.
I have spent the better part of a year building on models I do not control, so I am not writing this from a clean distance. While you build, the dependence never feels geopolitical. It feels like normal engineering, latency and context windows and pricing, code written around providers, a test suite, a release. Open weights help, and they matter, but reading the code was never the same as holding the switch. You can audit a model all day long and still not stop another country from deciding who gets to call the strongest systems and who is left waiting. I used to say, half as a joke, that Europe was renting the nervous system of its own future. On June 12 it stopped sounding like a line from an essay.
The CERN we already built
This is the part that stings, because it is not a story about Europe being asleep at the wheel. Europe woke up more than a year ago. In February 2025, at the Paris AI Action Summit, Ursula von der Leyen said it from the main stage:
This unique public-private partnership, akin to a CERN for AI, will enable all our scientists and companies, not just the biggest, to develop the most advanced very large models needed to make Europe an AI continent.
A CERN for AI. Good phrase.
And to Europe's credit, the thing did not stay purely rhetorical. OpenEuroLLM was launched that same month: around twenty institutions, companies and supercomputing centres, coordinated by Jan Hajič in Prague and co-led by Peter Sarlin of AMD Silo AI in Finland. The brief is exactly the kind of thing no single European country can really pull off alone, open and multilingual and shared, foundation models for public and commercial use.
The budget is 37.4 million euros.
I had to read that number twice.
Because the same Commission also announced InvestAI, a plan to mobilise 200 billion euros for AI, with 20 billion of it for up to five AI gigafactories, each one planned around roughly 100,000 of the most advanced chips, about four times what today's largest data centres run. So the money for concrete, power, GPUs and industrial theatre exists, and it exists in volume. The shared model, the one asset on the entire list that would have meant something on June 12, is running on grant money.
That is the strange European imbalance. We get infrastructure when it looks like a building, and sovereignty when it comes with a ribbon-cutting. Funding the actual model as if it were infrastructure too, that part we are still not comfortable with. But it is infrastructure, and more than the building in some ways. A data centre without a sovereign model is just a place where somebody else's intelligence can run faster.
The machines, at least, already know how to cooperate. MareNostrum 5 in Barcelona, LUMI in Finland, Leonardo in Italy, JUPITER in Germany, the first exascale computer Europe has ever had. All of them exist through cross-border EuroHPC cooperation. The hardware sorted out collaboration years before the model layer did.
And what runs on that shared Spanish supercomputer? ALIA-40B. Eight months of training, 40 billion parameters, public money, fluent in Spanish, Catalan, Galician and Basque. As a Spaniard, and especially as a Galician, I want that to exist. I am glad it exists. It is the kind of public AI infrastructure Europe should have started building years earlier.
But it also shows the problem. Spain has ALIA, France has Mistral, Germany has Aleph Alpha, Italy has Domyn, the Netherlands has its own efforts. Everyone gets a flag, everyone gets a press release. Meanwhile the American frontier does not care about our flags. It arrives as a product. Same grid, same power bills, a different national champion bolted onto each one.
A tier behind
I am not going to pretend any of this comes for free, because that is exactly how Europe ends up with one more summit, one more declaration, and one more underfunded consortium asked to perform a miracle.
Going sovereign in Europe today means accepting that, at least for a while, we may live a tier below the American frontier. There is no clean way around that. ALIA is a public-sector base model, not a Mythos killer. OpenEuroLLM is necessary, but it is not going to erase the gap in a quarter. Mistral is the best private shot the continent has, and even Mistral is still chasing the labs in front rather than clearly leading them.
Doing one thing together is also slower than letting every capital sprint on its own. It drags in the fights nobody enjoys: whose language comes first, whose national champion gets the contract, who gets the compute this quarter, which country gets to say it led the European model.
But the alternative is what just happened.
Europe keeps behaving as if fragmentation is a cultural inconvenience rather than a strategic weakness. It is not. Fragmentation is the reason we can have world-class researchers, serious public compute, good private companies, multilingual datasets, industrial users and endless political speeches about sovereignty, and still be left refreshing a status page when the United States changes an export rule.
And to be clear, the gigafactories are worth building. Europe needs the chips, the energy planning, the data-centre capacity, and a Commission with any sense builds the model and the buildings at the same time. The problem is the order of operations, which Europe has somehow managed to get backwards. Pour the foundations, furnish the place, underfund the tenant. A gigafactory with no European frontier model to train in it is just a very expensive data centre that you will, in the end, rent straight back to the American labs you were trying to climb out from under.
Fund the model, not only the buildings
The team that lost Fable in June did not need a speech about sovereignty or a future gigafactory. It needed a model sitting under a jurisdiction that would not pull the plug over a weekend, good enough to ship on, and actually available.
That model does not have to be the best on Earth on day one. This is where Europe tends to freeze. If it cannot lead it hesitates, and if it cannot dominate it funds a pilot and calls that ambition. But sovereignty does not begin at number one on a benchmark. It begins the day your hospitals, newspapers, public administrations, schools, startups and software teams have a serious domestic option that cannot be revoked by a foreign export-control decision overnight.
A slightly worse model that stays available can be worth more, strategically, than a brilliant one that vanishes the moment someone else's risk calculus shifts.
That is not an argument for isolation. Europe was never going to be independent in the childish sense, and it does not need to be. Chips, research, capital, cloud, open-source projects, security work, all of that stays international. The point was never autarky, the point is weight. Europe has to be heavy enough that the switch stops living entirely in someone else's office.
The pieces are not imaginary. The supercomputers are running, the consortium exists, the languages and the universities and the companies are all already here. The political vocabulary has been rehearsed so often that everyone in Brussels can say "sovereignty" in their sleep. The one thing missing is the decision to fund the shared model as if the future turns on it. Because it does.
OpenEuroLLM is expected to release its first models on July 31. Watch that date. Not because the first release will magically close the frontier gap, it will not, but because the reaction will tell us whether Europe understood what June 12 actually meant. If Brussels treats it like one more research milestone, we have learned nothing. If it treats it like the first real piece of a long-delayed European stack, then maybe the lesson landed.
The off switch was never ours. The only question left is whether Europe is finally serious about building one it can actually reach.