The US turned off its best AI for the rest of the world in the time it takes to sign a letter. Europe should treat that as the most useful thing that has happened to it in years.

Let me say plainly what yesterday meant, because the polite version helps no one.

On 12 June the US government ordered Anthropic to shut off its two most advanced models, Fable 5 and Mythos 5, for every foreign national on Earth. Not just for people abroad. For foreigners living and working inside the United States too, including Anthropic’s own engineers, locked out of the systems they helped build. The order did not come from a safety regulator. It came through export-control authority, the same machinery that governs the shipment of weapons and advanced chips. Anthropic could not sort its users by passport, so it did the only thing it could. It turned the models off for everyone, three days after launching them. (Anthropic’s statement; Al Jazeera)

Read that timeline again. A model that hundreds of millions of people had already started to depend on was switched off, worldwide, by a single letter, in seventy-two hours. No vote. No warning. No appeal.

That is not a story about one company, or one administration. It is a story about where Europe actually stands. So let me think it through the way I have been turning it over since yesterday, point by point.

One. The frontier is American, and now it is policy.

US companies build the most capable models. The rest of us get them later: or, as of yesterday, not at all. We told ourselves that gap was a few months of market timing, an irritation we could plan around. It was always more than that, and now it is explicit. Access to the best intelligence on the planet is a lever, and the US has shown it is willing to pull it. The delay was never just a by-product of where the labs happen to sit. It is an instrument, and we are on the wrong end of it.

Two. Exclusive access compounds, and it compounds against us.

A model that is both far stronger and available to only one side does not help that side a little. It helps it everywhere, every day, on every problem its companies touch. The American firm gets the better tool and we get the announcement. Stack that edge across a few product cycles and you no longer have a gap you can close with effort. You have European companies, whole industries, quietly outrun by competitors whose only real advantage was the better machine and the permission to keep it. We could lose entire sectors, not in a fight we entered and lost, but in one we were never allowed to join.

Three. And even if we win, the hardware closes the loop.

Say we do the hard thing. We build the datasets, we train the models, we get genuinely good. Those models still run on chips we do not make. Nvidia designs roughly 90% of the world’s AI silicon and will not sell us its best, because the advanced parts are rationed by the same export rules that just silenced Anthropic. So we would own the data and the models and still be renting the metal underneath them: straight back to points one and two, one layer down. Who really owns what runs on someone else’s machine?

Now the part that is hard to say, because saying it makes the problem enormous.

You cannot fix this one layer at a time. “Data sovereignty”, the comfortable phrase everyone in Brussels reaches for, is not sovereignty. It is a label we stick on the act of integrating someone else’s systems and hoping they stay switched on. If the models can be revoked, the chips withheld, the cloud compelled, then owning the data is owning the one layer that matters least. Sovereignty is not a layer you can buy. It is the whole stack, or it is theatre. Chips, models, data, infrastructure. All of it, or none of it counts.

I know how that sounds. Slow. Expensive. Close to impossible. We are talking about a project measured in a decade, not a budget cycle, and Europe is not famous for acting as one country for ten years at a stretch. But here is what I keep returning to. We are not starting from zero, and we insist on behaving as if we are.

The single most irreplaceable machine in the entire AI supply chain is not American. It is Dutch. Every advanced chip on Earth, including every Nvidia part the US is currently rationing, is printed by a lithography machine that exactly one company in the world knows how to build. That company is ASML, in Veldhoven. The Americans cannot make the best chips without us either. We are not a bystander in this supply chain. We own its narrowest chokepoint; and we have never once used it as leverage, because we have trained ourselves to be the customer in every layer instead of the owner of the one layer that cannot be replaced.

And it is not only the chokepoint. A few streets from the same Dutch supply chain, in Eindhoven, a company called Axelera AI is designing its own AI silicon: its Europa processor launched in 2025, it has raised more than 450 million dollars, and its next chip rides on Europe’s own RISC-V sovereignty programme. One company does not make a continent independent. But it is proof, sitting right there, that the talent and the design know-how are not a thing we have to invent from scratch. We are not starting from zero. We keep pretending we are.

So here is what the wake-up call is really asking of us. Not another strategy paper. Not twenty-seven national AI plans politely contradicting each other. One strategy, run like the defence programme it actually is. Bring the talent home: the European chip designers in California, the researchers training frontier models for other people’s missions, the founders who left because the capital and the compute were somewhere else. Fund sovereign compute the way we fund motorways. Build open-source models that belong to us, not fine-tuned copies of weights that an American or Chinese lab decided, this quarter, to let us borrow. And use the leverage we already hold at the bottom of the stack instead of apologising for having it.

None of this is comfortable. All of it is cheaper than the alternative. The alternative is the world we watched yesterday, made permanent. A Europe that builds elegant applications on top of an intelligence it does not own, cannot govern and is not allowed to keep, running on machines it is forbidden to buy, available right up until the morning someone in another capital decides otherwise.

We own the stack, or we keep renting our future from people who have just shown us, in writing, that they will switch it off the moment it suits them.

A note on the facts

Everything in this piece is real, including the part that reads like a thriller.

• The 12 June shutdown happened. Claude Fable 5 and Claude Mythos 5 became available on 9 June 2026. Three days later, on 12 June, the US government directed Anthropic to suspend access to both for any foreign national, whether inside or outside the United States, including Anthropic’s own foreign-national employees. Unable to filter its users by nationality, the company disabled both models for everyone, worldwide, three days after release. The order came through national-security export authority; the letter gave no detailed justification, and Anthropic said it disagreed with the action. (Statement on the US government directive to suspend access to Fable 5 and Mythos 5, Anthropic; US orders Anthropic to disable AI models for all foreign nationals, Al Jazeera; Anthropic Pulls Its Most Powerful AI Models After U.S. Bars Foreign Access, Time)
• The frontier is American and increasingly governed by export law. The US already restricts the sale of its most advanced AI chips through the “AI Diffusion Rule” and its successors, a tiered regime that blocks flagship parts like the H100, H200 and Blackwell from most of the world by default. The same export-control machinery governs both chips and the software that runs on them. (United States export controls on AI chips and semiconductors, Wikipedia; AI Export Controls: Navigating Chip Restrictions Globally, Introl)
• Nvidia designs the overwhelming majority of the world’s AI silicon. Estimates cluster around 80 to 90% of the AI-accelerator market: Bloomberg data put Nvidia at roughly 86% of AI data-center revenue in late 2025, and the company held about 92% of discrete GPUs in the first half of 2025. (Charted: The Battle for AI Data Center Revenue, Visual Capitalist; Data center / AI chip revenue of Nvidia, AMD & Intel 2025, Statista)
• The narrowest chokepoint in the supply chain is Dutch. ASML, in Veldhoven, is the world’s sole supplier of EUV lithography machines, the only tools that can print chips at 7 nm and below. Every advanced chip on Earth, including every Nvidia part the US rations, depends on a machine exactly one company knows how to build. (ASML Has a Monopoly on the Most Important Machine in Tech, The Motley Fool; ASML: The Key Bottleneck In The Global Semiconductor Supply Chain, Seeking Alpha)
• Europe does design its own AI silicon. Axelera AI, based in Eindhoven, builds energy-efficient AI inference chips. Its Europa AIPU launched in October 2025, the company has raised more than 450 million dollars (including a 250 million round in 2026), and its Titania chiplet is funded by the EU’s EuroHPC programme as part of DARE, the open-source RISC-V sovereignty initiative. The caveat: Axelera targets inference at the edge, not the large-scale training where Nvidia dominates. (Axelera AI, Wikipedia; Dutch AI inference chipmaker Axelera AI raises $250M, Tech.eu; Axelera AI Secures up to €61.6 Million Grant, Axelera)
• Europe answers with twenty-seven plans, not one. As of December 2024, 24 of the 27 EU member states had adopted national AI strategies, and the OECD itself flags the result as fragmented policy with limited cross-border coordination. (Progress in Implementing the EU Coordinated Plan on Artificial Intelligence, OECD)