The AI Trusted Partners Problem

What a working lunch in Évian tells you about the week access to AI stopped being something you buy.

As I write this, a working lunch is underway in Évian-les-Bains. The G7 heads of state are at the table. So are the people who build the models the G7 now spends its summits worrying about: Sam Altman of OpenAI, Demis Hassabis of Google DeepMind, Dario Amodei of Anthropic. Seated alongside them are Arthur Mensch of Mistral and Aidan Gomez of Cohere, Europe's and Canada's frontier hopefuls.

Mensch and Gomez have been invited to rooms like this before. What is new is not the invitation. It is the function of the chair. A month ago, a European challenger at a G7 AI table was the courtesy seat, the polite nod to sovereignty so the people who care about it felt seen. This week the courtesy seat became the hedge. They are the answer to the question every government in that building started asking last Friday.

On Friday, Anthropic switched off worldwide access to its newest model, Fable 5. The model had been public for three days. The shutdown followed a US export-control directive ordering the company to bar any foreign national, inside or outside the United States, from using them. The stated reason was national security. Mythos (Fable) is built to find flaws in code, and the concern is that the same capability that hardens a bank's systems can be turned to breaking them.

Note what Anthropic did not do. It did not build a two-tier system that lets approved users in and keeps foreign nationals out. It turned the models off for everyone. That choice is the part worth sitting with, and I will come back to it.

A model does not need to be a munition to be treated like one

A technology becomes a controlled strategic capability not when a statute says so but when a government behaves as though it already is. The moment a state asserts the right to decide who may use a model according to their nationality, the model has changed category, whatever the export schedule formally lists.

The category change does not complete in Washington. It completes in a procurement meeting in Munich or Riyadh or Singapore, in the moment someone there accepts that the switch exists and that a foreign government holds it. After this week, that acceptance is general. It is the thing that does not reverse when the models come back online.

The chip playbook does not fit

The instinct in Washington is to treat this as a chip problem and reach for the chip playbook. Export controls worked on advanced semiconductors because the chokepoint was physical and slow to copy. You cannot reproduce an ASML lithography machine or a TSMC fabrication process by reading about it. The knowledge is tacit, the capital enormous, the lead time measured in years. Deny the asset and the other side is genuinely stuck.

A frontier model is not that kind of asset. The architecture is published. The talent moves between labs and across borders. The scarce inputs are compute and capital, and those are precisely what the parties now reaching for alternatives are not short of. Restricting access to something that can be rebuilt does not wall off a lead. It puts a clock on the lead and hands every excluded party the budget justification to start the countdown.

The selective gate the company refused is being rebuilt one tier up

This is where the Évian lunch and the scheme being negotiated around it come together. According to the Financial Times, a US delegation led by Commerce Secretary Howard Lutnick is working with European diplomats on a "trusted partners" arrangement: a route for close allies, whether companies or whole countries, to reach the American models now barred from export. The framing is cybersecurity, the pitch is giving trusted allies the tools to defend against China.

Read that against Friday and the shape of it is clear. The company declined to gate access by nationality and shut the models off rather than operate that system. Governments are now assembling the same nationality-gate themselves, one level up, and calling the gate a partnership. The selective-compliance layer that was too compromising for a private firm to run is being reconstructed as a diplomatic club.

That is the real shift, and it is larger than any single ban. Access to frontier AI has stopped being a product you buy and become a status you are granted. You are inside the club or outside it. Outside, you are funding a sovereign alternative. Inside, your access is contingent on staying aligned with Washington across administrations that have already shown how quickly their policy turns. The membership card is not a guarantee. It is a bet on the issuer.

It would be too easy to file all of this as an American own goal, and I do not think that is the whole picture. The trusted-partners plan is also a piece of statecraft. It takes a blunder and tries to convert it into leverage: access as a reward for loyalty, extendable to the aligned and withheld from the wavering. That is more deliberate than self-harm and more dangerous. But it does not rescue the position. Inside the club or outside it, the same lesson has been delivered to every ministry and every boardroom watching: the availability of this layer is a political variable now, not a commercial one. That lesson does not unlearn.

The same failure I work on, run at the scale of nations

This is the part I have spent a year writing about, arriving from a direction I did not expect. I work on what I call trust architecture, the human layer of AI adoption, on the proposition that trust is the variable that decides whether a deployment lands or stalls. I usually make that argument inside a single organisation, to a leader who has bought the tools and cannot understand why the team quietly routes around them. The Évian lunch is the same argument at the scale of nations.

An organisation that builds something load-bearing on a foundation a third party can withdraw does not have a capability problem. It has an architecture problem. This week the world discovered, in public, that it had built load-bearing infrastructure on a foundation whose availability turns on US foreign policy. The responses are already on the record. Mark Carney has named the danger of leaning on a small number of American providers. The European Commission's technology chief has called the measures discriminatory and wants them gone. European capitals are moving to fund domestic alternatives at a speed that would have been politically impossible a month ago. At the summit itself, trust has become an explicit agenda item, with leaders including Narendra Modi pressing for transparency and trust frameworks instead of another round of who holds the cleverest model.

Sovereignty has stopped being an ideological preference and become a line on a risk register.

There is a deeper pattern under this, and it is one I learned to recognise somewhere other than a tech conference. Dominant powers tend to erode not through sudden collapse but through a particular manoeuvre. They begin converting the soft power that made them dominant, the plain fact that everyone wanted what they had, into hard power, the requirement that you comply in order to keep getting it. The thing people chose freely becomes the thing they must qualify for. The predictable result of weaponising what everyone wanted is that everyone starts looking for a version that cannot be taken away.

That is what spending trust as leverage looks like. It is not the end of American leadership in AI. The capability lead is real and may hold for a while. It is the moment the leader starts drawing the trust down faster than it can be rebuilt, on the assumption that the lead alone will keep the relationship intact. It usually does, right up until the alternatives mature.

The lunch will end and the communiqué will be careful. Access to the models may even be restored quietly within the week. None of that reaches what actually changed, because what changed is not a policy but an assumption. Every government, enterprise, hospital and university outside the United States now operates on the knowledge that a capability built into their work can be switched off by a decision taken somewhere they hold no vote. That knowledge does not flip back when the switch does. It leaves anyone deploying a frontier model carrying a question they were not carrying a week ago: not whether the model is good enough, but whether the relationship underneath it is stable enough to build on. In Évian today, that is the only question on the table that matters.

About Gail

Gail Weiner runs Simpatico Studios from Bristol, where she works on the human layer of AI adoption under the heading of trust architecture. She spent two decades in tech climbing from analyst to the C-suite, which is a polite way of saying she has watched a great many expensive deployments fail for reasons nobody thought to write into the business case. She is South African by birth, which means she learned to read state capture and institutional hollowing in real time, long before they became fashionable lenses for everyone else. She also runs an AI-native publishing house with twelve titles to its name, and shares an office with a small, elderly cat named Peanut, who outranks her and is fully aware of it.