Mistral AI Didn’t Betray Sovereignty. It Shrank It Down to a European Endpoint

After Mistral’s September funding round, the French press put the paper value of Cédric O’s stake in the company at around €90 million, and the activist lawyer Juan Branco seized on the figure. You don’t have to owe Branco anything to take a hard look at the record. Mistral is not the European counter-model we were sold. It is a company bound up with political power at every level: in its cap table, in the negotiation of its regulation, in government procurement. And since this summer, it has been a platform that hosts other people’s models, Chinese ones included. The pivot makes sense. The story told around it does not.

What I wrote, and what has changed

I owe my readers an explanation first. In January, I hailed the framework agreement signed with France’s Ministry of the Armed Forces as a patriotic deal at last (in French). In February, I called the shift toward infrastructure risky but smart, then saw the acquisition of Koyeb as confirmation of that thesis. In March, $830 million in debt financing still left me cautiously optimistic.

I stand by those pieces. They assessed an industrial trajectory, and on that front my view hasn’t changed. What has changed is that the trajectory has reached its destination, and the destination is not the one we keep being told about.

Cédric O, from cap table to AI Act

April 2023. Mistral has just been incorporated. Cédric O, France’s junior minister for digital affairs until 2022, buys in through his consulting firm, Nopeunteo: 17,610 shares at one euro cent each, or €176.10. On September 8, 2026, a €3 billion round led by Samsung values Mistral at €21.3 billion, which is where the €90 million figure comes from.

That number calls for rigor. It ignores dilution across successive rounds and the tranche Cédric O acknowledged selling as early as 2023; the shares aren’t listed, and his current position isn’t public. But even cut in half, the ratio between the stake and its value bears no resemblance to an ordinary investment.

A few months after he bought in, in the fall of 2023, the AI Act entered its final stretch. The proposed obligations on foundation models posed a direct threat to European labs too small to absorb heavy compliance costs. France dug in its heels, and Cédric O, then an adviser to and shareholder in Mistral, was on the front line arguing for lighter rules.

I have criticized Thierry Breton often enough that quoting him here is no act of flattery. On this point, he saw clearly: in his words, Mistral was defending “its business,” not the public interest. In the French Senate, lawmakers raised the question of a conflict of interest; the government replied that it had listened to all stakeholders. The final text, as I have argued, regulates by inputs and makes disclosure a confidential affair.

France’s ethics watchdog for public officials, the HATVP, had placed conditions on his move into the private sector, and to my knowledge no proceedings have been opened. Legally, then, there is nothing to object to. The question is political: the camp that talked about sovereignty had, at the negotiating table, one of its own who had become a shareholder in the startup billed as Europe’s answer to OpenAI. The state supplied the narrative; the company collected the valuation.

The state as first customer

Since then, sovereignty has become a sales pitch. A framework agreement with the Armed Forces in January 2026. A conversational assistant rolled out across the civil service. A Franco-German partnership with SAP for public services.

I have said it before and I stand by it: buying domestic is not a scandal, it is industrial policy. The trouble starts when public procurement stands in for proof of competitiveness. These contracts embed Mistral in government ministries faster than in the global market, and they create a dependency: let the majority change or priorities shift, and the order book shrinks. A company whose first customer is the state ends up optimizing its relationship with the state rather than its product.

The pivot, and what no one wants to say about it

The promise of 2023 was a lab that could hold the frontier against the Americans. In 2026, Mistral is no longer there. Closed models from OpenAI and Anthropic sit at the top of the leaderboards. Chinese open-weight models have closed most of the gap at a fraction of the price, as I laid out in Good Enough, Fifty Times Cheaper (in French).

So Mistral has moved to a platform model: data centers in Sweden and in Bruyères-le-Châtel, south of Paris, a serverless layer inherited from Koyeb, regional endpoints, agents, an enterprise offering. On August 11, it opened its infrastructure to third-party models. The first is GLM-5.2, from the Chinese lab Z.ai, served unmodified, in preview, under the same regional controls as Mistral’s own models.

This is the logical next step of what I have been writing for months. When the model becomes a commodity, value moves to the infrastructure, and hosting the best open-weight model available is a rational choice. What is not rational is to keep presenting that catalog as technological sovereignty.

A European endpoint guarantees where the computation runs and where the data stays. It says nothing about who designed the model, what it was taught, or what it was taught to avoid. With Nvidia chips underneath, what remains is legal sovereignty, not technological sovereignty. The latter is built into the architecture, not onto the label.

Qwant, the French search engine, had already worn the word out by selling independence built on results that came from elsewhere. Same word, same emptiness, bigger budget.

Mistral’s engineers are real, and there is nothing fictional about their work. What is fictional is the story of a startup whose work supposedly bothers only a few “petty thugs,” to borrow Arthur Mensch’s phrase, when it is also a product of France’s political and financial establishment.

Between Mistral and Azure, I will still choose Mistral. But from now on I will judge it on what it delivers: benchmarks, cost per task, how open the weights really are, how independent the stack is. Not on how many ministers have called it a source of French pride.