Austria's idea, floated this weekend in a letter to the European Commission, has the appeal of all the best simple ideas. Europe keeps getting cut off from the most powerful AI by decisions made in Washington — most recently on 12 June, when a US export-control order forced Anthropic to switch off its top models for everyone outside the United States. So why not stop depending on access you can lose overnight, and instead bring the company that makes the models inside European borders, under European law?

It is a genuinely interesting move. It is also, in the maximalist form people will read into it, something close to impossible — and understanding exactly why is the most useful thing to take from it.

American law doesn't stay behind when the company moves

The instinct behind "host the lab in Europe" is that jurisdiction is about geography: move the operation onto EU soil and it falls under EU rules. For US export controls, that instinct is wrong in three separate ways, and you only need one of them to be true.

The technology carries its nationality with it. US export law — the Export Administration Regulations, run by the Commerce Department's Bureau of Industry and Security — attaches to the origin of a technology, not its current location. US-origin technology stays subject to those rules wherever in the world it is moved. Anthropic's existing model weights are US-origin technical data; shipping a copy to Frankfurt doesn't strip the US claim off them. Washington has even reached directly at the weights themselves once: in early 2025 it created a dedicated export classification for the most advanced closed-model weights — then rescinded the measure that May, before it ever took effect, leaving the question of weight-specific controls unsettled. The broader argument doesn't need that flourish, though. The weights are US-origin technical data, and US-origin technical data carries the rules with it wherever it goes.

The company is American by incorporation, not by address. For export-control purposes a firm is a "US person" based on where it is incorporated. Anthropic is a Delaware public-benefit corporation — a US person — and its US-citizen staff are US persons individually, too. Opening a subsidiary in Vienna or Paris doesn't change that; a subsidiary is a corporate-structure move, not a change of nationality. The parent and its people remain bound everywhere they operate.

Even foreign-made copies get caught. Through what are called the foreign-direct-product and de-minimis rules, items made abroad can still fall under US control if they are the direct product of US-origin technology or tooling, or contain enough US-origin content. A frontier model built on US research, US-origin code, and US-designed chips is extremely hard to scrub clean of every one of those hooks.

Put together, these are why the 12 June directive could reach not just foreign users but foreign-national Anthropic employees inside the United States — releasing controlled technology to a foreign person counts as an export even if it never leaves the building. Relocating that same activity to Europe doesn't dissolve the logic; if anything, an EU operation staffed by EU nationals working on US-origin weights is exactly the kind of transfer the rule is written to capture.

There is one theoretical door left open: a genuinely separate, foreign-incorporated entity, building a model from scratch, abroad, with non-US staff and zero US-origin technology, tooling or chips, and no US weights transferred in. That really would sit outside the American regime. But notice what it describes — a clean-room European model. It is no longer "hosted Anthropic." The legal escape exists only for something that would not be Anthropic's models at all.

So what could "hosting" honestly mean?

The word is doing a lot of undefined work, and it helps to read it as a spectrum.

At the achievable end: an EU legal entity (most US labs already have one), European data residency for inference, a sovereign-compute footprint serving the models from European data centres, even European capital taking an equity stake. All of these are real, and most are probably already on offer in some form.

The trouble is that none of them touches the thing the Fable episode actually exposed. Data residency answers a privacy question. EU compute answers a "where does it physically run" question. An equity stake answers a "who profits" question. None of them answers the question that matters here: who holds the off-switch. If a US government order can disable the model worldwide, it does not help that the servers are in Hesse or that a European pension fund owns a slice of the company. The version of "hosting" that would neutralise that — moving the models out from under US jurisdiction altogether — is precisely the version that no company can deliver on its own.

Why Anthropic, and not just "back Mistral harder"?

The subtext of Austria's letter is a capability gap Europe doesn't much like to say out loud. The continent has a credible and fast-growing champion in France's Mistral, reportedly valued in the tens of billions and building its own compute; it has genuinely sovereign open models like Switzerland's Apertus and national projects such as the Netherlands' GPT-NL; and the EU has committed serious money to "AI gigafactories." But the gigafactories are compute, not labs — and they would still run on US-designed silicon. And Europe's own models, strong as some are, are generally regarded as sitting a notch below the very top US frontier that Anthropic, OpenAI and Google occupy. Wanting Anthropic specifically is an admission that backing the home team harder hasn't yet closed that gap.

The honest verdict

Here is the part worth holding on to. Even a willing Anthropic could not unilaterally exit US jurisdiction — the technology, the incorporation and the derivative-product rules all follow it. The only thing that lifts that ceiling is the US government choosing not to assert its claim: a government-to-government arrangement, a licensing accommodation, a policy carve-out. Europe can build the most attractive welcome mat imaginable — compute, capital, a legal home, a values match — and it still can't grant itself jurisdiction that Washington has not released. Austria has opened a diplomatic conversation, not announced a deliverable.

If you want a measure of how hard this is, look at the closest real precedent. America's cloud giants have spent billions building "sovereign" European clouds — AWS switched on a fully separate EU-staffed European Sovereign Cloud in January; Microsoft and Google have their own versions. Those are the most serious attempts anyone has made to wall US technology off inside Europe, and even they only achieve data and operational separation — and critics still argue a US-headquartered provider remains ultimately reachable by US law. If that debate is unsettled for where your files live, it is a different order of difficulty for who controls a frontier model's weights.

None of which makes Austria's letter pointless. Its real significance isn't legal, it's political: a European government saying out loud that being a price-taker on access to the most important technology of the decade is no longer acceptable. That sentiment is going somewhere. It just can't get there on a subsidiary and a welcome letter — only on a deal Washington has to want to sign.