On 12 June 2026, the US Commerce Department directed a major American lab to suspend all access to its two frontier models by any foreign national worldwide. It was the first documented use of export-control authority against a commercially deployed AI model, and it settled an argument that European policy circles had been having for years. Model access is not a procurement decision. It is foreign policy. And if you are a European company, the policy is written somewhere else.
The precedent is set, in both directions
Nobody should have been surprised. The United States spent four years tightening export controls on the chips that train models before extending the logic to the models themselves. China requires approval for AI model exports and blocks the leading American chatbot at home. Italy banned DeepSeek. Washington's June directive was not a new category of policy; it was the first time the category was applied to a deployed frontier model rather than a datacenter component.
Which means the scenario this continent quietly assumed was theoretical has already run once. The next exercise does not need to target citizens of everywhere. It can target a region, a sector, a sanction list, or simply become collateral damage in a trade dispute. Every European product that calls an API for its intelligence contains a component another government, or another company under government pressure, can switch off. The only question is whether your business plan survives the switch.
How Europe got here
No single law outlawed European labs. That is what makes the outcome so robust and so hard to reverse. What happened was cumulative: GDPR set the tone that novel data use is presumed guilty until regulated innocent. The AI Act layered engineering obligations onto anyone training general-purpose models, with penalties up to €35 million or 7% of global revenue, and its general-purpose rules went live this August while the high-risk regime waits in the wings. Procurement rules reward documented compliance over documented capability. Capital markets that fund compute ten-year runs exist in exactly two places, and neither is the EU. Each decision was defensible in isolation. The sum is a jurisdiction where the rational move is to regulate AI rather than build it, and then to be proud of how good the rules are.
Europe does have champions. Mistral ships credible open-weight models, pushes inference infrastructure inside the region, and has become the default answer to every sovereignty question asked in a Brussels panel. EuroHPC exists, gigafactory money has been announced, and the EU AI Act's defenders note, correctly, that nothing in the text forbids anyone from training a frontier model. But credibility is not capability. The best European models trail the frontier, the announced compute is a rounding error against American hyperscaler capex, and after the June cutoff even the friendly coverage read the situation correctly: sovereignty by vendor substitution is appealing precisely because the alternative does not exist.
What being a renter actually means
Renting has a price list, and Europe is reading it line by line.
Pricing is set abroad and can be repriced at will, and there is no credible exit because there is no credible alternative at the frontier. Capability is gated: new modalities and features land in other markets first, or arrive hemmed in by terms that treat the EU as a compliance burden rather than a market. Deplatforming risk is one directive away, as June demonstrated, and it does not even take malice; a sanctions action against a vendor's customer segment takes out everyone in it. Data gravity runs backwards: European data flows to foreign inference for processing, and the regulation that was supposed to prevent exactly this kind of dependency carefully does not. And in a squeeze, the EU gets pressure from both directions, with Washington restricting what its models may do and Beijing requiring approval for what its models may leave. Two foreign rulebooks, no home one.
The usual reply is that interoperability and diplomacy will keep the pipes open. Maybe. But note what the EU brings to that negotiation: market access, which cuts both ways, and no reciprocal capability whatsoever. You cannot credibly threaten to keep your frontier models at home when you do not have any.
What a European team does now
We are a European studio, so this is the part we treat as engineering rather than commentary. None of the following requires the political situation to get worse. All of it assumes it might.
Abstract the provider, not the ambition. Build against a thin interface so any model can be swapped in an afternoon, and rehearse the swap. The consolidation piece we wrote earlier applies with interest: single-platform convenience is a single point of geopolitical failure.
Hold an open-weights fallback on European inference. Open weights are the only weights that cannot be recalled, revoked, or geo-fenced after download. Run quarterly evaluations of the open frontier against your actual workloads, not the benchmark leaderboard, and know in advance which products degrade how far on the fallback. For regulated workloads, EU-hosted inference from day one; retrofitting data residency is a rewrite.
Own your data path. Prompts, completions, embeddings, and fine-tune sets in storage you control, with export runs in the disaster-recovery plan like any other backup. If your vendor is directed to cut you off, your history is the asset you will still have.
Read the contract like a geopolitics document. What happens to your fine-tuned weights if the provider is legally compelled to suspend your region? Notice periods, termination triggers, and where arbitration happens are no longer boilerplate. The June cutoff produced a wave of customers discovering their terms had no answer to that question.
What would actually fix it
Frontier labs are not founded by directives. They follow compute, patient capital, and early anchor customers, historically the state: defense contracts and national research programs seeded every American lab that now defines the field. Europe has the ingredients in principle, universities, engineering talent, industrial customers, and even announced compute programs. What it lacks is the risk appetite to fund training runs that fail and the procurement machinery to be someone's first customer. Until one of those changes, every European AI company is a tenant, and the lease terms are set in Washington and Beijing.
The AI Act gave European companies obligations. It did not give them capacity, and capacity is the thing the era actually runs on. Rules without labs means the future arrives as an API: metered, geo-fenced, and revocable by people who have never read a page of the rulebook we are so proud of. Sovereignty is not declared in a regulation. It is compiled into weights, and right now nobody on this continent is compiling.
Start from what the August obligations actually require of a small team, then read our take on when self-hosting the model is worth it and exiting a consolidated platform.