Summary
On June 12, 2026, the US government forced Anthropic to switch off a model that had been live for three days. No outage, no unpaid invoice — just an order from a foreign executive. That shifts the debate about European AI somewhere new, away from price and quality: to a class of risk no contract covers for you. Sovereignty here doesn't mean ideology, it means managing risk.
At 5:21 p.m. Eastern, a model that had been alive for exactly three days vanished for its customers — not because of an outage, and not because of an invoice, but because another country's government ordered it.
Three days and a kill switch
Anthropic launched Fable 5 and the underlying Mythos 5 on June 9, 2026. Open access was supposed to run until June 22. Instead, after three days, came an order and the lights went out.
The US Commerce Department — under Secretary Howard Lutnick — issued an export control order, leaning on national security authority. Anthropic was to cut off access to Fable 5 and Mythos 5 immediately. For all customers. For foreign employees. Now.
The other models stayed untouched. This was a surgical strike on one specific product. And it's the first time a government has forced a commercial AI product offline.
What actually happened (and what didn't)
First, what it wasn't. It wasn't an outage — nothing crashed. Nobody owed an invoice. No customer abused the product. And it had nothing to do with price.
The pretext was a jailbreak method for Fable 5. Anthropic reads it as a minor, previously known vulnerability — one you can find in other public models too. No universal jailbreak, no new class of hole. One sentence on the architecture: Fable 5 is the public version with safety classifiers, Mythos 5 the underlying model with cyber safeguards removed for vetted professionals.
Anthropic complied. But it disagrees, and out loud. It talks about thousands of hours of red-teaming and about safeguards "substantially more effective than any previously deployed model." It argues that if this standard were applied across the board, it "would halt all new frontier model deployments across all providers." And it's calling for a transparent legal process grounded in technical facts, not an overnight order.
This is important to read correctly. It's not a "evil USA" story where a company collaborates with a regime. The company itself is pushing back. The problem isn't Anthropic, and it isn't any particular administration either. The problem is the process — a mechanism that lets a single signature cut off a product thousands of others are building on.
A few days after the order, Trump barred federal agencies from using Anthropic's technology. That's a separate matter, a different authority, a different reach — just context for the mood all of this played out in.
A risk no contract covers for you
When you pick an AI provider, you run the usual variables. Latency. Price per million tokens. Quality on your workloads. SLA. Maybe data residency and GDPR. You've got all of that in your head, and usually in the contract too.
What you don't have in that contract is the government of the country where the provider sits.
Jurisdictional dependence is a class of risk all its own. Not technical — the model works. Not commercial — the provider wants to keep selling. Political. Your contract with Anthropic was valid the whole time on June 12. It didn't help. Because above it sits an authority your vendor can't override and that you yourself have no relationship with and no leverage over.
Almost nobody factors this risk into their architecture. We build failover for when a datacenter goes down. A hedge for when a vendor raises prices. An exit plan for when one goes bankrupt. But "what if a foreign government orders access cut off" isn't in the risk register. Until June 12 it was a hypothesis for paranoids. Now it's a precedent with a timestamp.
And it's not just frontier models. US tech controls over 70% of European cloud infrastructure (2025). Same logic, bigger surface.
Europe builds power plants, not cars
What Europe can do and what it can't.
It can do infrastructure and regulation. EuroHPC JUPITER in Jülich is the first European exascale supercomputer, inaugurated in September 2025 — real iron, not PowerPoint. InvestAI throws 200 billion EUR into the game, 20 billion EUR of it in a fund for AI gigafactories. And the AI Act has been in force since August 2024, with the main provisions including high-risk applying from August 2, 2026. We finished the regulation before the product it's supposed to regulate.
What Europe can't do is the frontier model and the speed. Private AI investment in the US reached roughly 285.9 billion USD in 2025 (Stanford AI Index). Across the EU countries, around 50 billion USD.
That's not a gap, that's an order of magnitude.
Mistral AI is the European flagship and is preparing Mistral Compute as sovereign infrastructure. Good model. But at the top of reasoning and coding it trails the American ones — it wins on price/performance and the sovereignty narrative, not on absolute quality. Germany's Aleph Alpha pulled out of the frontier-model race after half a billion dollars, because it can't sustain one, and pivoted to a layer on top of foreign models. And OpenEuroLLM, a consortium of 20 European institutions coordinated by Jan Hajič of Charles University, has a budget of 37.4 million EUR. Training a single frontier model costs orders of magnitude more.
Europe builds power plants and writes traffic laws. Someone else still makes the cars.
Sovereignty isn't autarky
Before this turns into a fantasy about Europe cutting itself off from the world, one hit of reality: even that European iron runs on NVIDIA. JUPITER too. No gigafactory changes that. If you wanted a "purely European stack" top to bottom, you'd hit the chip and stop.
That's not what sovereignty is about. It isn't autarky, and it isn't "buy European because European" either. It's about fallback and spreading risk. Not keeping a critical thing on a single kill switch — especially a switch controlled by someone you don't vote for.
The same engineering thinking as with anything else. A single point of failure is bad, even when it isn't hardware or the network this time, but jurisdiction.
What I'm doing about it on Monday
This isn't a call to revolution. It's a few lines you can ship right now.
Start with a multi-provider abstraction. Don't hardwire your code to one API. A thin layer between your application and the model, behind which you can swap providers without rewriting half the codebase — that's basic hygiene today, not a luxury. Once you have it, "switching vendors" is a config change, not a quarter-long project.
The second layer is an open-weight fallback. Models you can self-host can't be cut off by any order — you've already downloaded the weights. Mistral today, OpenEuroLLM down the road. They don't have to be better than the best you've got in the cloud. They just have to be good enough that an outage at your primary provider doesn't take you offline.
And for critical paths, where being cut off means an existential problem, self-host. Not everything — that would be its own kind of stupidity. Just the part whose going dark sinks you.
And in the background, keep an eye on what's growing under your feet. EuroHPC, gigafactories, European compute that in two or three years might make sense as a real alternative. You can't put it into production today, but it's worth knowing when it tips.
The question is no longer whether the European model is good enough. For most of your workloads it is — at the top of the frontier not yet, but that's not where most products live. The question is whether you can afford to let the kill switch on your infrastructure be held by someone you don't vote for and have no contract with. And you can answer that as early as Monday — with one line in your provider config.
