I am writing this on a Sunday, which is usually when I do my reading-the-news-properly hour. This week that hour was entirely eaten up by one story: Anthropic, on direct order from the US government, switching off access to its two most advanced models (Fable 5 and Mythos 5) for every non-American, including its own staff.
I don’t think the headline is really about Fable and Mythos. I think it’s about whether Europe has anything worth protecting in the first place, and that is a much less comfortable question than “we need our own model.”
So, what actually happened. The US government issued an export-control directive on national security grounds, with very little explanation attached. Anthropic complied while saying, quite openly, that it disagreed with the justification. The vulnerability the government seems to be worried about was, by Anthropic’s own account, minor. The kind of thing other models can also be made to do, with no special uplift from being “Mythos.” Whether or not you believe that account, it tells you something important: the decision to cut off access wasn’t really a response to a demonstrated catastrophic risk. It was a demonstration of who gets to decide.
That’s the part Europe should sit with.
Now, how much of this is bubble and how much is real? Here’s my honest take: the capability gap between Fable 5/Mythos and the models everyone outside the US can still use is, for almost all practical purposes, small. Most researchers, most companies, most of us writing emails and analysing data were never going to notice the difference day to day. The “we’ve lost access to the best AI” framing is doing a lot of emotional work that the actual technical reality probably doesn’t support, yet. What is real, though, is the precedent. If a government can decide overnight that an entire category of people cannot use a tool, regardless of what that tool actually does, then “access to AI” stops being a market question and becomes a foreign policy question. That shift is real, even if the immediate capability loss is small.
Watching the reactions scroll past, everyone is talking about this as if it’s the first time a government has ever decided a piece of technology was too dangerous to export. It isn’t. Not even close.
In 1999, Apple’s Power Mac G4 crossed the symbolic 1-gigaflop threshold and, under a Cold War-era export law originally written to keep supercomputers out of the Eastern Bloc, got classified as a munition. Banned from export to over 50 countries. Apple’s response was to lean into it completely. Steve Jobs announced from the stage that the US government had classified their new computer as a supercomputer and a weapon, and Apple ran a TV ad with tanks rolling up to surround a Power Mac on a desk. They called it “the world’s first personal supercomputer.” The restriction lasted a few months before the threshold was quietly adjusted and the whole thing became a footnote. For context, the iPhone in your pocket today does roughly a thousand times that gigaflop, and nobody is sending in the tanks. Same decade, different angle: Netscape and Microsoft were legally barred from exporting browsers with full-strength, 128-bit encryption outside the US and Canada. So they shipped a deliberately weakened “export version”, 40-bit encryption, to the rest of the world, including most of Europe. Within a couple of years, researchers in Sweden, the UK, and France had cracked that weakened version in days using a handful of ordinary workstations, which rather undercut the entire point of the restriction. The rules were loosened by 1999–2000, after years of industry lobbying. The pattern, every time, looks the same: a government draws a line based on a specific technical threshold, the company complies (sometimes loudly, sometimes with a wink), the rest of the industry’s products keep shipping completely unaffected, and within months to a couple of years the line gets redrawn once it looks as outdated as it always was. Fable 5 and Mythos 5 will very likely follow that arc — probably faster, given how quickly this category moves.
Which brings me to the “just build our own model” reflex, which I’ve now seen about fifty times on LinkedIn since yesterday. Here is the thing that nobody wants to say out loud: we already have Mistral. It’s good. It’s improving fast. It’s explicitly built around European data residency, EU AI Act compliance, sovereign cloud partnerships, everything the sovereignty crowd says they want. And yet, walk into most European universities, ministries, or companies, and what’s actually running on people’s laptops? Not Le Chat. The gap isn’t a model gap (at least not entirely). It’s an adoption gap, a procurement gap, a habit gap. All together, at the same time. Building a second or third European model while everyone keeps defaulting to American tools because the integrations are smoother or the habit is older doesn’t fix sovereignty. It just gives us another underused option to point to in op-eds.
And this is where the China comparison gets interesting, because I think people are reasoning about it backwards. China’s big platforms thrive on a domestic market of 1.4 billion people, one language, one regulatory regime, and a government that actively steers procurement towards them. The “Chinese market alone is enough” logic works because it is, structurally, one market. Europe’s combined population is comparable, but it isn’t one market — it’s 27 regulatory regimes, dozens of procurement cultures, and a default assumption that the “real” tools come from somewhere else. So when people ask “why isn’t Europe’s market enough for Europeans,” the honest answer is: it could be, but only if Europe started treating itself as a single buyer with a preference. The way it does, occasionally, for industrial standards. Right now it behaves like 27 small, slightly anxious customers of someone else’s product.
As for Anthropic, I think they’re in a genuinely difficult spot, and it’s worth being fair about it. They built their public identity on “these models are powerful enough to need careful gatekeeping,” and now a government has taken that framing and turned it into a literal gate, applied in a way the company itself says it disagrees with. And even keeping their own developers, greatest minds away from their products. How they respond matters: complying while publicly contesting the rationale, as they’ve done, is one path. It preserves the relationship while keeping the disagreement on record. But it also exposes a tension that every frontier lab will eventually face. If you spend years telling everyone your product is potentially dangerous, you lose some standing to argue later that a government is being unreasonable for treating it that way. I don’t think this “destroys” Anthropic, but I do think it’s a preview of a dynamic that isn’t going away. Model access increasingly sitting downstream of geopolitics rather than just product strategy.
So what are the actual pathways for Europe? I see three, and I don’t think they’re mutually exclusive.
- One: genuinely coordinate demand. Government and large-enterprise procurement that defaults to European tools, the way Mistral’s partnerships with Dassault and others are already trying to do, but at much larger scale.
- Two: negotiate, rather than hope treat tiered access to frontier models as something to be formally discussed between governments, not something that gets decided unilaterally and discovered via a press release.
- Three, and this is the one closest to my own research: stop trying to win the frontier race on someone else’s terms, and double down on the things Europe is actually positioned to do differently. Sustainability, human-centred design, sector-specific tools built around real constraints rather than spectacle.
None of these are quick. All of them require Europe to actually use what it builds, not just fund it and then keep buying somewhere else out of habit. That, more than any export control, is the thing within our control.
If you want to talk through any of this — especially the adoption-gap point, which I think deserves its own post — let me know. I suspect I’ll be coming back to this topic more than once this year.
AI #Europe #DigitalSovereignty #Anthropic #Mistral #AIPolicy #FutureOfAI #TechGeopolitics
Didem Gürdür Broo 