For years we bought AI models like cloud services. You checked quality, price, latency, security, compliance, vendor stability. You picked a provider and built your products, workflows and operations on top. Underneath sat the assumption: a model available today would be available tomorrow.
That assumption died on a Friday afternoon.
On 13 June, a US national-security directive forced Anthropic to cut off two of its frontier models for every foreign national — overnight, including its own engineers on visas. Hundreds of millions of users dropped on the strength of a Friday afternoon letter, sent at 5:21pm. You can argue about the pretext. The precedent is the part that matters. Access to frontier AI is no longer decided only by the companies building it. Governments now have a hand on the switch.
Most AI governance debate has been about safety: should a model ship, should it be open, should it be regulated. That is no longer the live question. The live question is who controls access. Once a technology becomes infrastructure, it turns political. We watched it with energy, with semiconductors, with the cloud. Intelligence is next.
So a new line appears on the procurement checklist. Next to security, privacy, lock-in and compliance, you now weigh geopolitical risk. Questions that sounded academic a year ago are Monday-morning questions. Can a capability be withdrawn after we have built on it? Restricted by geography, or by nationality? This is business continuity. Most organisations spent years making sure they did not depend on a single cloud. Almost none made sure they did not depend on a single source of intelligence. That gap is now visible.
The operational reality gets awkward fast. Picture an engineering team spread across Europe, the US and India, where access suddenly forks by passport. The restriction does not stop at the model. It can ripple through every tool that serves it — the coding assistant, the agent, the law firm’s research system. Questions that used to sit with HR and legal land on the architecture team’s desk. Two engineers on the same team, different tools, because of where they were born. For a global organisation, that is a problem with no clean owner.
The precedent is contagious. If one government treats model access as a lever, others copy it. China can. Europe can. Blocs form around strategic interest. Instead of one global market competing on capability, you get several, competing on access and jurisdiction. The US paid a price too. Its one real advantage as a supplier — predictable, rule-of-law, safe to build on — took a hit in a single Friday letter. Governments should be able to act. A model showing real cyber or biosecurity danger is a fair reason to intervene. The problem is timing. Blocking a release is one thing. Pulling a capability after thousands of organisations have built on it is something else entirely.
This is where the European question changes shape. For years the debate about providers like Mistral was about quality. The first question is now how much single-country dependence a European bank, hospital, railway or ministry can afford to carry — the way we once asked it of a single cloud. The US models keep their lead; that was never the question.
The organisations that handle this well will build a sovereignty stack on purpose. US frontier models where you need raw capability. European providers for sovereignty-sensitive work. Open-weight models for continuity you control. An orchestration layer that lets you switch when the ground moves. It is the same logic countries use for energy. Countries are removing a single point of failure, the same way they do with energy. Optionality like that is bought and maintained.
The last decade taught us that data is strategic. That Friday taught the harder lesson: access to the intelligence itself can be taken away — and that changes how AI gets governed, bought and built. Once access is geopolitical, sovereignty stops being a government word and becomes a boardroom one. So every leadership team now owns a question it did not have last year. What happens if our most important AI capability disappears tomorrow?