Mistral raises three billion and Anthropic raises 65

On September 8, Mistral AI announced it had raised €3 billion, "the largest equity fundraising round ever completed by a European technology company," at a valuation above €21 billion.^[1] Samsung Electronics led the round, a name I did not expect to find in that seat, and the valuation nearly doubles last year's, when ASML came in at €11.7 billion.^[2] I read the number, I smiled, and then I got out the calculator. The smile did not survive the division.

Twenty Mistral rounds for one Anthropic round

Three billion euros is about $3.5 billion, and dollars are the currency the competition counts in. Anthropic raised $65 billion on May 28 at a $965 billion valuation, three months after raising $30 billion.^[3] OpenAI had closed a $122 billion round on March 31 at an $852 billion valuation, the largest in the history of venture capital, with Amazon, Nvidia, and SoftBank at the front of the line.^[4] Over the first half of the year, those two companies alone absorbed $217 billion, 43% of all the money invested in startups on the planet.^[5]

One Anthropic round is worth about twenty Mistral rounds, and the San Francisco company is worth forty times the Paris one. The European record is real, and it is measured on a European scale. Since its founding in 2023, Mistral has raised €5.7 billion in total,^[6] which Anthropic now books in under two months, its annualized revenue having reached $47 billion in May.^[3:1] Mistral is aiming for €1 billion in annual recurring revenue by the end of the year, up from €300 million a year ago,^[7] a fine curve on a chart whose vertical axis is not the same one.

Back in June I wrote that backing a national champion did not mean holding more summits but putting in enough money for it to even consider competing. Three billion is enough to exist, and the AFP wire says as much in its headline, which has Mistral raising money "to keep existing in the global AI race."^[2:1] Existing is an honorable goal, and a more modest one than the posters promised two years ago.

The money goes into concrete and the frontier can wait

Read the press release and you learn where the billions go, and it is not primarily into a model. Mistral describes itself as the builder of a full stack, "open-weight models, the infrastructure and the compute capacity they run on, and the products that bring them into production," serving 125 large enterprises in 20 countries, Airbus, ASML, and HSBC among them.^[1:1] The company has already committed €4 billion to data centers in France and across Europe, targets 200 megawatts of capacity in 2027 and a gigawatt in 2030,^[2:2] and an $830 million bank loan paid for more than 13,000 Nvidia graphics processors installed at Bruyères-le-Châtel, south of Paris.^[6:1] Microsoft already rents those machines for its European customers. The new money will "scale its compute capacity, build infrastructure, accelerate commercial growth, and expand its international footprint,"^[1:2] and Arthur Mensch speaks of "moving faster in research, products, and infrastructure,"^[2:3] research getting, in the sentence, the place the billions do not give it in the list.

Mistral disputes that reading. Audrey Herblin-Stoop, who runs its communications, answered the press's doubts the same day with one flat sentence, "We are not giving up on having frontier AI models,"^[8] frontier being the word for the best models of the moment. In July, Arthur Mensch admitted that Mistral did "not yet own the best language models" while having "constantly reduced that gap," and promised "a very exciting model to come this summer," in early access with a few partners.^[9] Summer is over, and the public has seen nothing. The biggest house model is still Mistral Large 3, from December 2025, and I could not find a date for the next one.

I can't say I'm surprised, and that is the part that stings. I used Le Chat, back when it was still called Le Chat,^[10] and I stopped a long while ago, without slamming any door, because the models stopped keeping up. What I ask of an assistant has grown over time, to include code, prototypes, and long documents, and Mistral answered less and less of it. I would be the first to pick the French option if the gap were a few months. It is measured in generations, and that is too much to give up for a flag.

The most interesting model on the platform comes from Beijing

On August 11, between two infrastructure announcements, Mistral opened its platform to other labs' open models, the ones anyone can download, "starting with Z.ai's GLM-5.2," and the sentence before it deserves a second read, since it says that "today's AI systems are no longer built on a single model but on an ensemble of capabilities."^[11] Z.ai is the brand of Zhipu, a Beijing lab, and GLM is, with DeepSeek and Qwen, one of the three big Chinese families of open AI that I listed in early September among the models you can download and cannot run, for lack of an air-conditioned rack in the living room. At Mistral it is served "without modifications," on the same infrastructure and with the same regional guarantees as the house models, at $1.40 per million input tokens and $4.40 per million output tokens.^[12]

Arthur Mensch supplied the doctrine to go with it, and it is a clever one. Sovereignty, he argues, depends less on where a model comes from than on who controls its deployment, its infrastructure, and its data, so a foreign model can be run without depending on its creator's servers.^[13] In other words, the sovereign asset is now the building. Three years after presenting itself as Europe's frontier lab, Mistral sells air-conditioned square meters with a Chinese model running inside, which LeMagIT sums up as "a sovereign AI layer."^[6:2]

I turned this over for a while and landed on the opposite of the easy irony. It is a good move. Mistral alone will not catch the frontier, its own executives say so, while the Chinese open models trail it by a few months and nobody outside a data center can run them. Offering them from a site south of Paris, with data that never leaves for Beijing and an invoice in euros, is exactly the product I was missing. I would never have typed my documents into a service hosted in China, and my graphics card will never swallow GLM. Between those two there was an empty chair, and Mistral just sat down in it.

Six million on one side and three billion on the other

The week before the round showed how much the French state weighs in this story. On September 2, Les Échos reported a €6 million agreement between the state and Mistral for 2026 and 2027, confirmed by the office of David Amiel, the minister for public administration and public accounts. It covers licenses and engineers seconded to the ministries, on three projects, cybersecurity, fraud detection, and high-volume litigation, and it is non-exclusive, so other vendors can compete on the same work.^[14] Six million over two years is a services contract, and it is two thousandths of the round announced six days later.

The same week, from San Francisco, where he was meeting OpenAI and Stripe, Roland Lescure, the economy minister, said that "if European AI comes down to Mistral, then we're done for," and that "we can't put all our eggs in one basket."^[15] He is right in principle, and it will take other baskets. For now, the only one available has just received €3 billion.

Public money is not absent, though, and that is what's new about this round. The Scaleup Europe Fund, a €5 billion vehicle set up by the European Commission and run by EQT, co-led the deal, one of its very first investments.^[16] The Grand Duchy of Luxembourg took a stake, Bpifrance came back in, and LeMagIT notes that two thirds of the capital now sits with European organizations, which hold 75% of the voting rights on the board.^[6:3] That is close to what I asked for in June, public money actually disbursed rather than summits, and I note it before I start quibbling.

Still, the biggest check came from Suwon, where Samsung is headquartered, the previous one from Veldhoven, home of ASML, and I had kept in my back pocket the finance ministry's written promise to watch foreign investment in Mistral. Samsung is leading a round in which the founders keep control,^[2:4] without buying the company, and I don't know what to make of that, except that the biggest check for European sovereignty in AI came from Korea this year and from the Netherlands last year, and that France has never yet signed that one. I am no venture capital expert. I only note that after Hugging Face, whose sale to Nvidia was announced last week, one gem remains, that it has just found its billions, and that if Europe wants it to stay, Europe will have to use it, ministries and agencies first, before asking it to compete.

What I tell the colleague who asks

If a colleague asks me tomorrow whether to bet on Mistral, I start by asking what for. For a well-bounded step in an existing workflow, sorting requests, summarizing case files, extracting fields, why not, and the argument that the data stays in Europe carries real weight inside a public agency. For building things, it's dead. A good part of my job is putting together prototypes, the mockups that get the business side reacting before anyone writes a specification and hands it to the developers, and that routine needs a model that codes and is rarely wrong across a long document. Today that model is called Claude, and I don't know anyone who does this work with Mistral.

The day the gap narrows, I switch, and I have reasons beyond price to want it. Until then, the model I'll go get on Mistral's platform is Chinese, hosted south of Paris, and billed in euros. That is not the sovereignty we were promised in 2023, and it may be the only kind for sale.