Two months ago I wrote a long piece called “The Mistral Mystery: How Much is the European AI Champion Worth?”. The round was a rumour then, reported at €20 billion. It has closed at €21 billion.
Management has said more since. This is an update on what we now know, and on what we still do not.
It is an examination, not a recommendation.
The bigger story of the last few months is not the money. It is what has happened to Mistral's sovereignty pitch.
Start with what has not changed. No second European lab is competing with Mistral head-on. Others exist, and some do good work. None is anywhere near €21 billion. The next largest is worth a fraction of that. The United States has OpenAI, Anthropic, Google, Meta and SpaceXAI pushing at the same problem at once. China has DeepSeek, Alibaba's Qwen, Moonshot, ByteDance and Z.ai. Europe has Mistral, and then a gap.
What has changed is who Mistral is standing with.
In July it expanded its partnership with Microsoft. Microsoft is funding Mistral's European data-centre build-out, taking capacity from those centres for its own cloud customers, and distributing Mistral models through Azure, Foundry and Copilot Studio. In August, Mistral opened its platform to third-party open models. The first was GLM-5.2, from the Chinese lab Z.ai. And this morning's round is led by Samsung.
All three came within seven weeks of each other. For investors, that puts a question mark over the promise Mistral was built on.
Some have called this a surrender. The French tech podcast Silicon Carne ran an episode on the Microsoft deal titled “Mistral, la capitulation”, and a second one asking whether hosting a Chinese model was a betrayal.
A year ago the situation was different. Europe's only frontier lab was funded by Europe's largest company. ASML put €1.3 billion into a €1.7 billion round, took around 11% and a seat on the strategic committee. It was a strategic investment and not a financial one: ASML builds the lithography machines that every advanced chip in the world is printed on, and the two companies said they would work together on chip design. Europe's champion, backed by Europe's champion, working on the one part of the AI supply chain Europe already owns.
This morning's round is wider. Samsung leads. The EU's Scaleup Europe Fund and PSG Equity co-lead. Advent, BlackRock funds and the Grand Duchy of Luxembourg come in as new investors. ASML did not sell. It invested again, and so did Nvidia.
So is the sovereignty promise dead? It is worth working through that slowly. Sovereignty is Mistral's moat, and that makes it one of the most important questions an investor can ask about this company.
The case for an alternative to American and Chinese AI models is obvious. In June the US Commerce Department ordered Anthropic to cut off access to its two most capable models, Fable 5 and Mythos 5, for any foreign national anywhere in the world. That included Anthropic's own foreign employees. The order was lifted two weeks later. The fortnight is not the point. The point is that the switch exists, it sits in Washington, and it was thrown once with no notice. Four months earlier the administration had ordered federal agencies to stop using Anthropic, and the Defense Secretary had designated the company a supply-chain risk. No American company had ever been given that label before.
The worry about Chinese models runs in the other direction, towards data. It is not baseless either. Anthropic reported that DeepSeek and two other Chinese labs had tried to extract capabilities from Claude.
So the demand for a model that is neither American nor Chinese is real, and it is not only European. Cohere in Canada, AI21 in Israel, Falcon and G42 in the UAE, Sakana in Japan, LG and Naver in Korea are making the same argument to their own governments. Cohere and Aleph Alpha signed a transatlantic partnership in April. Analysts call them the middle powers. Mistral is the largest of them.
Sovereignty sounds simple until you try to build it. The obvious version is a European model, trained and run on European infrastructure, owned and financed by a European company. Clean. It has also never existed.
Every serious AI system on earth runs on Nvidia. Nvidia designs its chips in California. TSMC makes them in Taiwan. Europe's contribution to that chain is ASML, in Veldhoven, which builds the lithography machines that print the transistors and has no competitor anywhere in the world. That is one link out of three, and it sits at the start of the chain rather than the end. Mistral's own data centre in Essonne runs 13,800 Nvidia GB300s. The building is French. The power is French and mostly nuclear. The silicon is not.
That said, the training has not all happened on French soil either. Mistral has rented compute from CoreWeave, an American AI cloud provider, since signing its first contract for H100s in 2023, moving on to H200 and then GB200 clusters. Mistral's own chief technology officer, Timothée Lacroix, said so in a CoreWeave case study: “[Our models] were trained 100% on CoreWeave infrastructure. I think not being on that kind of infrastructure would've delayed us by at least a few months.” CoreWeave operates data centres in Europe as well as in the United States, so this does not tell us where the training physically took place. It does tell us the infrastructure was rented from an American company, and that until recently Mistral had no alternative.
So the strict definition was never available to anyone. Mistral has proposed a different one.
Sovereign AI, the company wrote last month, is “AI that keeps data, intelligence, compute, and operations under the customer's control”, where “models can be adapted and owned on open weights” and “training and inference can run on infrastructure and in jurisdictions the customer chooses”.
Read the subject of that sentence. It is the customer, not Europe. Sovereignty here means you pick your weights, you pick your jurisdiction, and you keep what you build on top. It says nothing about where the company is registered or where the chips were made.
That definition describes what Mistral has spent the last year becoming, which is no longer a research lab. It is a full stack.
At the bottom is compute. Mistral Compute sells GPU capacity in Europe. European Compute Units let enterprises commit for several years in advance, and the anchor group is named: Amadeus, ASML, Capgemini, Caisse des Dépôts and CMA CGM. Mistral says it is aiming for one gigawatt of its own capacity by 2030. Separately, it is a partner in a 1.4GW campus in the Paris region, a joint venture with Nvidia, the French state bank Bpifrance and MGX, the Abu Dhabi investment fund. Construction starts in the second half of this year and operations are due from 2028.
Above that is inference you can place. Regional endpoints let a customer pin processing to Europe or the United States. A priority tier adds custom rate limits and an uptime guarantee. Mistral says it is the only European lab offering both.
Above that is model choice, including models Mistral did not build. The platform now offers customers a selection of models rather than only Mistral's own, and the first addition is GLM-5.2, from the Chinese lab Z.ai. Mistral has said more open models will follow. GLM-5.2 runs on the same infrastructure, under the same regional controls and the same service commitments as Mistral's own models.
Above that is software. Studio for building agents, Forge for training custom models, Vibe for the assistant. Then model customisation and a delivery methodology, which is consulting.
And underneath all of it sit the contracts. The French Ministry of Defence signed a framework agreement in January that runs to 2030 and covers the armed forces, the CEA, ONERA and SHOM, on French infrastructure only and no commercial cloud. There is a defence partnership with Helsing in Germany and a public-sector programme with SAP for France and Germany. In April the European Commission awarded €180 million of sovereign cloud under explicit sovereignty criteria for the first time. In June it proposed the Cloud and AI Development Act, which sets four tiers of cloud sovereignty and reserves the top defence tier for suppliers that are EU-owned and EU-controlled. That act is still a proposal. Adoption is targeted for late 2027.
What I take from all this is that Mistral knows what it cannot win. It cannot beat the American labs on capability. It cannot beat the Chinese labs on cost. So it is competing somewhere else.
The value in AI does not sit only in the model. It sits in the infrastructure underneath, the architecture around it, the control the customer keeps, the contracts, the legal requirements and the independence. That is the ground Mistral has chosen.
Having shareholders and partners on several continents does not weaken that position. It protects it. A company that depends on no single government is a real alternative to a fully American or a fully Chinese stack. The definition of sovereignty has moved. The dependence risk has not, and that is the problem Mistral is still solving for its customers.
It shows the limits too. Putting a Chinese model on the shelf is an admission. Mistral's own models are not the best answer for every customer, and the company does not pretend otherwise.
Now to the numbers, which are thin.
A month ago I wrote that Mistral says far less about its business than its American peers, who publish their annualised revenue almost as a sport. What we knew then was this. Annualised recurring revenue passed $400 million in January, against roughly $20 million a year earlier. Management guided to more than $1 billion by the end of 2026.
This morning confirmed the guidance and added little else. Arthur Mensch said Mistral is on track to pass $1 billion in annual recurring revenue before the year is out, and that he expects “to be beating” that figure “if everything happens as they are trending”.
That is the whole disclosure. No margin. No losses. No cash burn. No split between the enterprise business and consumer subscriptions. No customer numbers. For a company that has just raised €3 billion, including public money from the European Commission's new Scaleup Europe Fund, managed by EQT, and from the Grand Duchy of Luxembourg, that is not very much.
One caution on the headline figure. Annualised recurring revenue is a run rate, not audited revenue. It takes the most recent month, or perhaps only the last few weeks, and multiplies it out, which flatters any business growing quickly and says nothing about whether the revenue recurs. Mistral has never published a full-year revenue figure.
That said, the direction is impressive. Going from $400 million to $1 billion in a year is roughly two and a half times growth.
Here is a rough sum, and it is mine rather than the company's. If Mistral is running at about $800 million today and reaches $1.6 billion twelve months from now, revenue over the next twelve months averages around $1.2 billion. Against a $24 billion valuation that is about 20 times forward revenue. This assumes enterprise value and equity value are the same, which ignores both the $830 million of data-centre debt and the €3 billion of cash just raised, but it is close enough for a back-of-envelope number.
Twenty times is a familiar figure. OpenAI priced near it at $300 billion, then at $500 billion, then at $852 billion this year, each time against forward revenue at the moment of the round.
Anthropic's January round, at $350 billion, appears to have priced on the same 20 times forward revenue. What followed was a growth acceleration few observers had modelled. Its run rate reached $65 billion by July. Against numbers like that, the May round at $965 billion implies a multiple closer to the mid-teens. That is the exception rather than the rule, and it comes from a specific situation.
So is Mistral fairly valued, overvalued or undervalued at the $24 billion announced this morning? For a company growing this fast and disclosing this little, that is close to unanswerable.
One thing is clear. Mistral is worth nearly twice what it was a year ago, when the September 2025 round priced it at €11.7 billion, or about $14 billion. So it seems that revenue has grown faster than the valuation. The company has not been repriced upward on hope. It has grown into a similar multiple.
And the sovereignty promise, which still holds, has changed shape. Call it a soft pivot. Whether that is a company adapting well to an environment that moved under it, or a quiet retreat from what it promised its customers and its investors, is for you to decide.