On 12 June 2026, the United States Department of Commerce issued an export-control directive to Anthropic, one of the world’s leading AI developers. The instruction was to suspend access to its two most advanced models, Fable 5 and Mythos 5, for all foreign nationals, inside the United States and abroad, including the company’s own foreign employees. Anthropic had no reliable way of checking its users’ nationality in real time, so it switched off both models for every customer, everywhere, that same evening. Three days after their launch.
The suspension lasted almost three weeks. On 1 July, Fable 5 began returning to users worldwide. Mythos 5, which had been reserved for a small group of partners from the start, was restored only to certain American organisations approved by the government. Some chose to read this as the end of the incident. Yet restoring access only underscored the problem. A service that European businesses, public bodies and researchers use every day can be interrupted by an American national-security decision, without Europe being consulted or even targeted. Access withdrawn by one letter, restored by another.
To understand what has just happened to us, it helps to reread a text written twenty-five centuries ago.
An ancient lesson in power
In 416 BC, the Peloponnesian War had been under way for fifteen years, with Athens, at the head of the Delian League, ranged against Sparta and its allies. Melos, a small island in the Aegean and a Spartan colony, had initially remained neutral. Athens landed and ordered the island to submit and pay tribute. What followed was the exchange Thucydides recorded as the Melian Dialogue. The Melians asserted their right to remain neutral; they appealed to justice, to fairness and to the mercy owed to a peaceful city that was hopelessly outmatched. The Athenians dismissed the argument: justice, they replied, applied only between equals in power. Melos refused to yield. Athens besieged the city, starved it into surrender after several months, executed the men of fighting age, sold the women and children into slavery, and settled five hundred colonists in their place.
One sentence from that episode has travelled down the centuries:
“The strong do what they can and the weak suffer what they must.”
Twenty-five centuries later, it describes Europe’s position in the race for artificial intelligence with scarcely a word changed.
We are the Melians
The AI value chain, from semiconductors and computing power to models and platforms, is held at both ends by two rival powers, and Europe is neither of them. The United States controls the most advanced chips, the frontier models and most of the global cloud market. China controls the rare earths, the permanent magnets and a growing share of the hardware chain, and is closing the gap in software fast. Each has already shown that it can turn this position into a weapon: Washington through export controls on processors, and then, on 12 June, on a model itself; Beijing through licensing restrictions on rare earths and critical materials, restrictions whose impact European industry has already felt. Europe has real strengths, not least ASML, but has converted almost none of them into the ability to operate independently of foreign providers.
Faced with this, what does Europe do? It pleads. It appeals to law, to rules, to multilateralism and to regulation, which is exactly the Melian defence. Our first response to each show of force is another policy document. Nothing guarantees that the superpowers will not one day do to us what Athens did to Melos: not by force of arms, but digitally, economically and politically, by cutting off access or revoking a licence. 12 June was a dress rehearsal, and Europe was not even the target. We were collateral damage from an American domestic decision.
Remember how the story ends: no one came to save Melos. Sparta, from which its settlers had come, did not intervene. No one will come to save us either.
The switch
What 12 June exposed was not so much a government’s ill will as an architecture. Access to an AI model served remotely, through an API, can be withdrawn: whoever hosts it can switch it off for you, at any moment, on its government’s instruction, with no notice and no detailed explanation. A letter is enough. A locally deployed model, one whose weights, the parameters that make it up, have been published, downloaded and installed on your own servers, cannot be remotely disabled by its original provider, on one condition: the environment running it must itself be independent, with no licence to renew and no remote administration by the original supplier. Meet that condition, and no letter from Washington or Beijing will switch it off. This asymmetry, the asymmetry of the switch, must now shape Europe’s strategy.
We should be honest about its limits. Open-weight models run on processors that Europe does not make, and those processors are themselves subject to American export controls; China has experienced those restrictions first-hand. Resilience is not all or nothing; it is built up layer by layer. But there is a difference in kind, not merely in degree, between a tool that can be revoked in an evening and a fleet of machines already delivered and running without anyone’s permission.
Neither vassalage nor autarky
Does it follow that Europe should aim for technological self-sufficiency? No, and that is the second lesson of Melos. The island perished because it believed itself stronger than it was. Europe is not on the level of the AI superpowers and will not get there by decree. Rebuilding the entire chain at home, from advanced foundries to frontier models and hyperscale cloud, is a political aspiration rather than a workable industrial strategy. Accepting permanent dependence, on the other hand, means remaining Melian for ever: relying on the goodwill of the strong and calling it a partnership.
This warning should not be mistaken for a rejection of American capital, technology or partnerships. Europe has no alternative to them. It struggles today to raise capital at the scale and speed that infrastructure demands. When a European company finds in America the funding it cannot find at home, it is not betraying Europe; it is securing its survival. The transatlantic partnership has, in its time, built some of our finest industrial assets. Whether to work with the strong is not in question: we do, and we must. The question is what we are left with on the day the strong change their mind.
There is a third way, the way of clear-sighted middle powers: trade and cooperate with both blocs, and ensure that neither bloc can unilaterally cut off essential services. Resilience is not hostility. The ability to keep operating if access is withdrawn is what gives the weaker partner a stronger negotiating position.
The three pillars of resilience
Three tasks, in ascending order of importance.
01Models independent of the original provider
Competitive open-weight models hosted on European soil. They need not be developed in Europe; the global open-weight ecosystem is thriving, and it would be absurd to do without it. But they must be evaluated, fine-tuned, deployed and maintained here, on infrastructure that can operate without remote authorisation from the original supplier, and that has been tested to prove it. The gap with the best closed models is real. It matters less than a different question: do we have a foundation that no one can switch off? For a large share of business and public-sector uses, that foundation already exists. The longer-term task is to maintain and strengthen the ability to evaluate, adapt and replace these models as the best ones evolve. Otherwise today’s resilience will become tomorrow’s technological disadvantage.
02Inference capacity under European control
Backed by our low-carbon electricity, one of the few inputs in the chain where Europe, and France in particular, holds an advantage that geography protects. Large shared facilities have their place. Modular data centres, distributed across Europe, can be deployed in months rather than years, provided they are built on sites that are already equipped and connected and where power is available. They also strengthen continuity of service, provided applications are replicated across sites and failover is tested. This is the infrastructure that gives open-weight models somewhere to run.
03The public sector as an anchor customer
Use public procurement to provide a stable market for this capacity. EU institutions and national governments spend billions of euros on digital services each year. For critical applications, public buyers should require demonstrable portability and the ability to keep operating independently of the original provider. Contractual assurances alone are not enough. Buyers should require evidence: appropriate usage rights, documented dependencies, recoverable model weights and configurations, backup capacity in place, and failover tests carried out under real conditions. This resilience has a cost, from initial evaluation to day-to-day operation. For inference capacity serving critical public-sector applications, it is better funded through long-term procurement contracts for services that public bodies need than through subsidies. France already knows how to do this: SecNumCloud sets technical, operational and legal requirements designed to protect sensitive data, including against extraterritorial legal exposure. The next step is to apply the same principles to AI models. Public demand has built entire industries before; it can build this one, and send every supplier a message the market understands at once: here, the switch does not work.
The lesson of Melos
The Melians were not naïve. Faced with overwhelming force, they appealed to everything they could: justice, the gods, the uncertainties of war and the hope of Spartan intervention. Their arguments failed to save them. Their mistake was not believing in justice. It was having no means to defend it.
The story does not end there. The following year, Athens, drunk on its own strength, launched the Sicilian expedition, the disaster that would cost it first its fleet and then its empire. Unchecked force always turns on itself in the end. But that was no consolation to Melos: the punishment of the strong has never brought the weak back to life.
Europe has law, regulation and principles, and it is right to have them. It also has something Melos never had: time. The Melian Dialogue begins when the strong are already in the harbour and the weak have nothing left but arguments. We are not there yet. Open-weight models exist, the electricity is available, and European governments control their own procurement decisions. What remains is to act and commit the resources needed to sustain that effort.
There is still time to avoid the Melian Dialogue. But its defining feature is that, once it begins, it is already too late.