Artificial intelligence has a flag too

Until recently, we had never imagined that artificial intelligence models might have a nationality.

The interface was the same in Madrid, Berlin, or São Paulo. The bill arrived every month. The model responded from a cloud with no visible geography. Its limits seemed technical: context, latency, cost, hallucinations. Its rules appeared to be private decisions made by the provider. Artificial intelligence was presented as a global service, more like a digital subscription than a piece of strategic infrastructure.

On June 12, 2026, the United States government ordered Anthropic to suspend foreign nationals’ access to Fable 5 and Mythos 5 after the company disclosed a potential method for bypassing their safeguards. Shortly after its latest major release, Anthropic complied with the order and disabled access to the model for all non-US users.

The sequence matters more than the technical incident. A company built the model, designed its classifiers, selected the authorized users, and publicly explained why it considered the risk acceptable. It took a letter from the state to establish who had the final say.

Fable 5 did not merely become unavailable. Artificial intelligence acquired a flag.

The fence and the territory

Fable 5’s architecture rested on a simple premise: the dangerous knowledge was already inside the model, so safety had to determine who could access it. The classifiers acted as a fence. High-risk queries were redirected. Accredited researchers could receive additional permissions. Anthropic served as the gatekeeper.

During the brief period in which the model was available worldwide, the initial debate focused on the height of that fence: whether it could be climbed, how much effort it would take, and who could do it. That was a necessary technical question, but an incomplete one.

The suspension has revealed that the fence was built inside sovereign territory. Anthropic can decide who crosses it only as long as that decision does not conflict with the authority of the US government. It took two days for that conflict to emerge. The classifiers, the consortium, and the private evaluations are no longer the highest layer of the system.

This does not mean a private company should hold sovereign authority over offensive capabilities. It means the public framework capable of replacing that authority is still rudimentary. There was no visible technical process, no published proportionality criteria, no known review period, and no appeals mechanism for affected users. The only available tool was the shutdown.

This is not yet a governance regime. It is state authority applied at crisis speed.

From controlling chips to controlling capability

There are precedents for this mechanism. The United States once treated cryptography as munitions, coordinated restrictions on advanced technology during the Cold War, and now uses semiconductor export controls as an instrument of national security and industrial policy.

Fable 5 belongs to that tradition while also changing what is being controlled.

A chip is physical. It can be identified, tracked, and stopped at a border. Distributed software can be copied, which eventually limited the effectiveness of controls on cryptography. A cloud-hosted model combines elements of both: its parameters may be reproducible information, but its commercial capabilities remain centralized. Users do not receive the model. They receive tokens.

That makes access an extraordinarily convenient lever. There is no need to intercept a shipping container or remove software that has already been installed. Permissions can simply be changed in a service controlled by a handful of companies.

Export controls no longer have to apply only to physical goods. They can apply to cognitive capability rented one query at a time: reasoning, scientific research, defense analysis, or code generation, granted, measured, and withdrawn remotely.

The difference is geopolitical. Whoever controls the model does more than sell a tool. They manage a reserve of capability that other countries integrate into their businesses, universities, and public institutions without controlling the ultimate terms of supply.

Europe regulates demand

For Europe, the suspension carries a more uncomfortable lesson.

The European Union has built the world’s most ambitious AI regulatory framework. It can impose obligations on systems deployed in its market, require documentation, and define prohibited uses. That is one form of power: the power to regulate demand.

Yet the most advanced models are concentrated in a small number of US companies, trained with infrastructure, capital, and ecosystems that Europe cannot immediately reproduce. As long as access continues uninterrupted, that dependence looks like a purchasing decision. When Washington intervenes, it reveals itself for what it has always been: a political relationship.

Europe has the regulations. The United States has many of the models, the cloud infrastructure that serves them, and jurisdiction over the companies that can switch them off.

The distinction recalls other forms of infrastructure that were presented as global until a crisis exposed their center of power. It happened with financial networks, semiconductor supply chains, and cloud services. Operational neutrality lasts only until it conflicts with a national security priority.

AI is no exception. It has civilian and military applications, concentrates economic value, and reshapes the distribution of knowledge and productivity. It is exactly the kind of technology that states turn into an instrument of foreign policy once it becomes important enough.

Sovereignty is not self-sufficiency

The easy answer would be to demand a European Fable. It would also be the slowest answer.

Building domestic models requires compute, energy, capital, data, talent, and a market capable of sustaining the investment. Europe should work toward that goal, but technological sovereignty cannot depend on first achieving parity with every US lab.

Organizations that use frontier models as critical infrastructure should treat jurisdiction as a vendor risk, just as they treat financial stability, security, and service continuity. They need architectures that allow providers to be replaced, graceful degradation plans, data portability, and processes that can continue operating without depending on a single API.

Contracts and European regulation should make clear what happens when an export ban is imposed, what information customers will receive, and how they will be able to migrate. No European contract can override a US government order, but it can prevent a strategic dependency from being disguised as an ordinary subscription.

Industrial policy must invest not only in a continental champion, but also in distributed capacity: compute, open models, research centers, cloud providers, and standards that reduce the cost of switching systems. Sovereignty does not mean producing everything within national borders. It means retaining options when another actor decides to close a door.

The meter and the kill switch

When AI began moving away from flat-rate pricing, the change appeared to be economic. Usage would be metered. Companies would have to track tokens, allocate budgets, and decide which tasks justified the cost.

The suspension of Fable 5 adds the missing piece: alongside the meter is a kill switch. The provider controls part of it. The government with jurisdiction over that provider controls another. The foreign customer controls neither.

The debate over artificial intelligence can no longer be limited to how much a model costs, what safeguards it includes, or who receives privileged access. It must also ask which state can interrupt access, through what process, and in defense of whose interests. The suspension of Fable 5 has shown that the model was already part of a system of control that was neither technical nor private, but geopolitical.

Europe can continue refining the rules that govern consumption. But if artificial intelligence is to become essential infrastructure, Europe will also have to confront supply, substitutability, and power.

Models do not arrive from a cloud without territory. They come from specific companies, data centers, jurisdictions, and flags.