When Washington controls access to models: export controls and cognitive sovereignty

Until now, the geopolitics of AI played out on silicon: GPU quotas, restrictions on lithography equipment, entity lists. In June 2026, a line was crossed. The US Department of Commerce temporarily applied export controls to Anthropic's most advanced models — Mythos and Fable — imposing access restrictions on foreign nationals. The same month, an executive order tasked federal agencies with designing an evaluation framework for "covered frontier models". For the first time, it is access to the model itself — not the chip, not the exported weights, the service — that becomes an instrument of foreign policy.

A precedent that matters more than its scope

The measure is presented as temporary and targeted. No matter: it is the mechanism that counts. It establishes that a government can, overnight, condition access to a cutting-edge cognitive capability on the user's nationality or the geography of their organisation. European companies building their products, their R&D processes or their internal tools on American frontier APIs are discovering that their intellectual supply chain contains a single point of failure — and that this point is in Washington.

From software dependence to cognitive dependence

We knew how to reason about hyperscaler dependence: it is an infrastructure problem, painful but migratable. Dependence on frontier models is of another nature. As AI inserts itself into product design, analysis, software development or decision-making, it is the organisation's capacity for intellectual production that becomes hostage to a provider subject to a third country's law. Cutting off access does not cut off a service: it degrades the company's collective speed of thought. This is what one might call cognitive sovereignty — and it does not yet appear in any European regulatory framework.

The possible responses, without naivety

First response: the model portfolio. Never couple a critical process to a single frontier provider; maintain a tested switchover capability towards open models (European or not) even if less capable. A loss of quality can be measured; a loss of access is simply endured. Second response: local inference for sensitive workloads — open-weights models now reach a level largely sufficient for a great share of enterprise uses, and once the weights are downloaded, no foreign decision can withdraw them. Third response, a collective one: the European gigafactories and the support for European model makers only make sense if demand follows. Every enterprise contract signed with a European provider is, literally, an act of industrial policy.

The June episode may be remembered as the moment Europe understood that the question was no longer "is our data protected?" but "does our capacity to think with machines belong to us?". The answer, today, is no. It can become yes — provided models are treated as what they now are: critical infrastructure.