The Fable recall turned the foreign kill switch from a thought experiment into a precedent. For any organisation outside US jurisdiction, model continuity has become a business continuity question.
On 9 June 2026, Anthropic launched Claude Fable 5, its most capable public model to date. Three days later, the US government issued an export control directive, and within hours the model went dark worldwide. If that can happen to Anthropic, it can happen to whatever you have built on, and that is the part worth sitting with.
It is worth being precise here, because the precise version is more unsettling than the dramatic one.
The US government issued an export control directive covering Fable 5 and its more powerful sibling, Mythos 5. Citing national security, it required that access be restricted for any “foreign national”. Not only those abroad, but foreign nationals inside the US, including Anthropic’s own non-citizen employees. Anyone without US citizenship or a green card, wherever they happened to be living.
You cannot check a user’s passport at the moment an API call fires, so Anthropic had no way to comply selectively. It disabled both models for every customer worldwide to avoid the penalties the order carried, while its other models, including Opus 4.8, stayed online.
So Washington never reached into a data centre and flipped a switch. It wrote a rule that left the vendor no compliant option except to switch off the model itself. As one security team pointed out afterward, the order technically targeted foreign access, but on a same-day notice the only practical way to enforce it was to pull the model for everyone. The off switch did not have to be technical. A letter was enough.
Anthropic is contesting the decision. The company said it disagreed that a narrow jailbreak should justify recalling a model “deployed to hundreds of millions of people” and that the same capability already exists in rivals such as OpenAI’s GPT-5.5. Dean Ball, an AI policy expert who served under Trump’s administration, called the move “cartoonish”. Whatever happens to Fable next, and it may well come back, the precedent now stands: the most mainstream AI vendor in the world had a flagship model recalled three days after launch, with only hours to comply. That fact will not expire.
The natural first reaction is to reach for a different provider. That reaction misses the point.
Fable was a commercial model sold through every major cloud marketplace, and it was pulled regardless. That show’s even large vendors operating at scale are not exempt.
It also sits on top of risks that were always present in a closed API, just easier to ignore. The provider already controls your pricing, your rate limits, and the moment a model gets retired. The Fable recall added one final item to that list: the right to access it at all is not fully a vendor’s prerogative to grant, because a government can override it. Once you have wired a specific model into your products, your support desk, or your analytics, that stops being a sourcing preference and becomes a continuity exposure. When the model goes, everything you built on it goes with it, on a schedule you do not control. Model continuity is now part of business continuity, and most disaster recovery plans do not mention it yet.
The recall also exposed how thin our everyday definition of sovereignty has become. For years, “sovereign cloud” has mostly meant that the data sits in a local region. That matters, but it only covers a fraction of what sovereignty actually requires.
Three distinct sovereignty layers are at stake:
Model-access sovereignty: Can the model your business depends on be switched off, deprecated, or restricted by a foreign provider or government? This is precisely the layer Fable failed.
Data sovereignty: Does your data actually sit beyond foreign reach? Storing it in Frankfurt does not achieve that if the provider’s parent is a US company. Under the CLOUD Act, US authorities can compel a US-headquartered provider to hand over data wherever it is stored. This was confirmed last year when Anton Carniaux, Microsoft France’s director of public and legal affairs, testified before a French Senate inquiry that he could not guarantee that French citizens’ data would never reach US authorities.
Operational sovereignty: If the provider you rely on were cut off tomorrow, could you keep running on your own? Most organisations built on a closed model could not.
Local data residency satisfies only the gentlest version of the first layer and almost none of the other two.
If residency is not sovereignty, what is? Europe has already written the answer down.
The EU’s Cloud Sovereignty Framework grades providers on a ladder called SEAL, running from 0 to 4, and the rungs are the instructive part. SEAL-2 means a foreign authority cannot reach your data without your knowledge. SEAL-3 means the service keeps running even when a non-EU supplier is cut off. SEAL-4 means a full EU supply chain, from chips to software. In the Commission’s first sovereign cloud tender, no provider reached SEAL-4, and one consortium qualified at only SEAL-2 while running partly on Google Cloud under EU operation.
The framework never asks where a vendor is headquartered. It asks who governs access, whose law applies, and whether the system survives a supplier being pulled. Sovereignty is an architecture, not a flag on the box. The opposite, a sovereign label applied to someone else’s stack, is what the market now calls sovereignty washing, and after June it has become a much harder sell.
Error correction is no longer an academic curiosity. It now sits at the center of the strategies of QPU manufacturers, startups, industrial research labs and large national and European programs.
Many public roadmaps point to fault-tolerant demonstrators in the 2027–2030 timeframe. While these timelines are ambitious and should be treated cautiously, they reflect a shared understanding: without fault tolerance, quantum computing will not deliver industrial value.
In practice, most current demonstrations still focus on error-corrected quantum memories, with only limited examples of full multi-qubit logical computations.
None of this is a one-off, which is why it will not quietly blow over. The EU already depends on non-EU suppliers for more than 80% of its key digital products, services, and infrastructure, three US companies hold roughly 70% of its cloud market, and the US and China together control close to 90% of the world’s AI computing capacity, with the US alone holding about three-quarters and Europe left in the low single digits.
The response is moving quickly. The EU’s Tech Sovereignty Package, its growing network of EuroHPC AI Factories, and a multi-billion-euro push for AI gigafactories are all bets on capability that cannot be unplugged from outside. The question that began with “where does our data live” now runs all the way to “who controls our models,” and after the recall it grew loud enough to reach the UK, Canada, France, and the Netherlands within days.
So what should a defensible setup actually do? The Fable episode reads almost like a checklist of the failure modes to design out, and a genuinely sovereign platform has to answer each one.
It has to run on infrastructure you control, so there is no foreign off switch.
This is where BullSequana AI begins. Bull is the only European company that designs, builds, and deploys four of the key infrastructure layers required for AI sovereignty: use cases, data + AI platforms, systems and components. This way, organisations can provision new capacity with an integrated approach or deploy a portable software layer on their existing infrastructure, whether that is their own data centre, a sovereign or public cloud, or a hybrid Kubernetes setup. A cluster running under your jurisdiction is simply not something a foreign commerce department can write a letter against.
It must stay open, so you keep operating no matter what.
BullSequana AI runs on open code, open APIs, and a model-agnostic core, with no lock-in to any single model provider, GPU vendor, or framework. It serves models through OpenAI-compatible APIs, which turns swapping a model into a configuration change rather than a rebuild. And if your relationship with Bull ever ends, you keep the platform and carry on running.
It has to be efficient, so control does not cost a fortune.
The platform is tuned to get more out of the hardware beneath it: right-sized models instead of the heaviest one for every task, inference that does not leave GPUs burning power while idle, and applications designed to treat each model call as something worth conserving. You serve more AI on the hardware you already own, on a bill that you set.
It turns ambition into production, because control over an empty platform is worth little.
Pre-built AI use cases proven across multiple clients, plus Bull’s experts working alongside your terms, mean you start from working solutions rather than a blank repository. The hard part of AI was never the model; it was adoption.
BullSequana AI will power Sweden’s Mimer AI Factory, a €30 million system procured through EuroHPC and co-funded by the EU’s Digital Europe Programme, built to help Swedish researchers and companies develop sovereign language models on European technology. That is real sovereign AI, under local control.
Strip away the geopolitics, and the Fable recall comes down to a few questions a board can put to any organisation this week. Do you own your AI or merely rent it? Can someone outside your jurisdiction switch it off? If your provider vanished tomorrow, would you keep running? Can you show a regulator where your data went and who touched it?
For anyone built on a foreign, closed model, June 2026 turned those into uncomfortable questions. The point of building sovereign by design is that the answers become simple again. You own it. No one outside can switch it off. You would keep running. And you can show them.
The Fable recall made it impossible to keep looking away from these risks. The organisations that come out ahead will be the ones who made sure that nobody else could ever take their model away.
See where you stand. Book a call with our team to walk through your own exposure, and we will show you in a live demo how BullSequana AI keeps your models, your data, and your costs firmly under your control.