Mistral AI

IWhat it does

Mistral AI is a Paris-area company, founded in 2023, that builds large language models. These are the programs behind chatbots and coding assistants. It was started by Arthur Mensch, who worked at Google DeepMind, with Timothée Lacroix and Guillaume Lample, both former Meta researchers⁸. Mensch is chief executive, Lacroix is chief technology officer and Lample is chief scientist⁸.

It is not publicly traded. Its shares cannot be bought on any exchange. Anyone outside the company can own an interest only through a private round, a private secondary sale, or a holding in a listed shareholder such as ASML. The company has not filed to list, and its chief executive said in February 2026 that no listing was needed that year⁹.

The product range is wide. The chat and agent product was called Le Chat and is now called Vibe. A platform called Forge lets a company train a custom model on its own data. There are models for text, images, audio, document reading (OCR, which turns scans into text) and code. Mistral also makes a physics-modelling pipeline for engineering work, announced with Nvidia in October 2026⁴⁸. In October 2026 it put its newest flagship, Mistral Large 4, into public preview. It has about one trillion parameters in total, of which 49 billion are active at a time, and it was trained on 3,800 Nvidia Grace Blackwell chips in Mistral-run European data centres⁴.

Many of its models are open-weight. That means the trained model files are published, so a customer can run and change the model on its own machines. Mistral also hosts other makers’ open-weight models, starting with GLM-5.2 from the Chinese lab Z.ai³¹². That is a notable choice for a company whose pitch is European independence, and I take it as a sign that Mistral sees itself as an infrastructure provider as well as a model maker.

Since 2025 the company has moved down the stack into computing infrastructure. It bought the infrastructure start-up Koyeb in early 2026 and the Austrian physics-AI firm Emmi, reportedly valued at €330 million⁸¹¹. It runs a 40 megawatt GPU cluster near Paris, is building a 44 megawatt site south of Paris backed by an $830 million loan, and is building in Sweden¹²⁹. The target is up to 1 gigawatt of capacity in Europe by 2030³⁷.

IIHow it makes money

Mistral has not published audited revenue, profit or loss. Every revenue number I found is an annualised run rate, meaning a recent month or quarter multiplied up to a year, or a target. A run rate is not the same as revenue booked in an accounting year, and it can include contracts that have been signed but not yet delivered. Mistral says its figures are annual recurring revenue (ARR), which is the yearly value of subscriptions and usage contracts⁹.

Four things appear to bring in the money. First, usage-based access to its models through an application programming interface (API), where customers pay per unit of text processed¹⁸. Second, private or on-site deployments, where Mistral installs models on the customer’s own servers and sends engineers to work with the customer. TechCrunch compares this forward-deployed approach to Palantir’s⁸. Third, the chat and agent subscriptions. Fourth, a newer line of selling computing capacity³¹⁸.

The company’s chief revenue officer has said most customers are on very large, strategic, multi-year deals¹¹. That helps predictability, but it also means a few contracts may carry a large share of the total. Mistral has not given a customer breakdown, so I cannot test how concentrated revenue is.

The February 2026 figure came from the chief executive, speaking to the Financial Times: an annualised run rate above $400 million, up from about $20 million a year before⁹. Sacra, a research firm, estimates about $312 million in December 2025 and about $400 million in January 2026, and about $16 million at the end of 2024¹⁸. Fortune’s August 2026 piece says the run rate exceeded $400 million "in 2025", which does not match the February 2026 timing in the other sources; I treat the Financial Times account as the better dated one¹¹.

On the target: the chief financial officer, Johan Bergqvist, told Reuters in September 2026 that the company is on track to reach $1 billion of ARR by the end of 2026¹³. Euronews reports the same as €1 billion¹⁴. These are company statements, not results. I have not seen any figure for what Mistral earned in the months between February and September 2026.

Costs are not disclosed. What is visible is the commitment: roughly $830 million of debt for the Paris-area site, €1.2 billion for the Swedish data centres, and Mensch’s own estimate that a gigawatt of computing costs about $50 billion⁹¹². The Series D of €3 billion is small against that last number. This is the central tension in the business: the revenue is growing fast, but the cost to stay in the race is growing as well.

Revenue and cost signals (reported by the company unless stated)

IIIThe market and the competition

Mistral sells in three overlapping markets: AI models, AI services for companies, and computing capacity. In the first it faces OpenAI, Anthropic and Google DeepMind. Fortune reported in August 2026 that Anthropic had last been valued at $965 billion and OpenAI at $852 billion, both dozens of times Mistral’s size, and that Anthropic’s annual run rate was estimated at $47 billion¹¹. Those labs also report much larger revenue. Their numbers use different methods and I have not tried to make them comparable.

Fortune’s assessment, which I have read and found fair, is that Mistral’s models are not at present competitive with the leading US systems on the hardest tasks. On some benchmarks its flagship Large 3, released in December 2025, performed near the level OpenAI reached about eighteen months earlier¹¹. The Decoder, relaying the February report, notes Mistral is not prominent in the Arena.ai or Artificial Analysis rankings¹⁰. Mistral’s reply is that most business work does not need the frontier¹¹. That claim is plausible, but I cannot test it from public data.

In the second market it faces the large consulting and cloud companies, open-weight models from China, and a short list of European rivals. Euronews names Germany’s Aleph Alpha, the defence firm Helsing and Switzerland’s Apertus open-model project; they are smaller¹⁴. Microsoft is both a partner and a possible rival. It expanded its partnership with Mistral in July 2026 and, according to Channel Insider, made a multibillion-dollar commitment to use part of Mistral’s expanding European GPU capacity¹³. Microsoft also agreed to fund Nvidia Vera Rubin chips for Mistral’s European capacity without taking equity, per The Next Web¹².

Demand for European control over AI is the reason Mistral exists as a business. The Decoder reports that the EU sources more than 80% of its digital services from foreign, mostly American, providers, and quotes Mensch saying Europe has realised that its dependence on US digital services was excessive⁹. In the same report Mensch said customers ask for AI that does not run on anything owned by the hyperscalers (the largest US cloud companies)⁹. Governments of France, Germany, Luxembourg, Greece and Estonia are named as customers, and the Luxembourg state became a Series D investor⁹¹.

Events in 2026 strengthened that case. Fortune reports that in June 2026 the US Commerce Department temporarily cut foreign access to Anthropic’s Mythos model, and that European buyers who had been seeking access turned toward Mistral¹¹. TechCrunch describes the same period in different terms, as a US directive that led Anthropic to pull its latest models offline⁸. The sources differ on the mechanism, but both agree that US policy hit access to a US model. In May 2026 Mensch told the French National Assembly that Europe had about two years to build its own AI infrastructure or risk becoming a "vassal state"¹¹.

Policy money points the same way. Euronews reports the EU’s InvestAI initiative at a headline €200 billion, and that the Commission opened tenders in July 2026 for up to seven AI gigafactories, which are not expected to operate before 2027 or 2028¹⁴. The same piece says enterprise AI adoption across the EU is about 13.5%, a fraction of the US level¹⁴. A low adoption rate means a lot of room to grow, but it also means Europe’s buyers have been slow to pay for AI at all.

There is a limit to the sovereignty argument. Fortune notes that about 40% of Mistral’s revenue comes from the US and other non-European clients¹¹, and the February report says some Mistral offerings still run on servers from American hyperscalers⁹. Mistral relies on Nvidia chips and says it does not design its own⁸. So the independence it sells is partial: it is independence in who owns the contract, the model and the data location, not in the chips.

IVThe moat and how it could be disrupted

A moat is whatever stops a customer or a rival from taking the business away. For Mistral I see three possible moats and none is deep yet.

The first is political and legal: being European. Governments and regulated companies in Europe have reasons to prefer a European supplier, and some are required to keep data inside the EU. That is a real barrier for US rivals, but a modest one. US labs can sign local-hosting deals and Mistral’s own regional-endpoint feature shows how: customers can choose whether processing runs in Europe or the US³. If a US lab offers a European cloud with a sovereignty guarantee that satisfies procurement, Mistral’s edge narrows to preference rather than necessity.

The second is switching cost. Where Mistral models are installed on a customer’s own servers, trained on its documents through Forge, and wired into its workflows, moving is hard. That is a real source of stickiness. The long compute commitments from ASML, Amadeus, Capgemini, Caisse des Dépôts and CMA CGM add to it: the partners commit for about five years with no early exit, per The Next Web¹². I note that the largest of these anchor customers is also a shareholder, which means the demand is partly circular and does not show how an independent buyer would behave.

The third is scale in compute, which is still being built and is far smaller than the US labs’. It cannot yet be called a moat. If the company completes its plan, owning infrastructure could lower its cost per unit of AI and secure supply. But the US labs are raising far more money, so Mistral is more likely to stay behind on scale than to catch up.

What is not a moat: the models themselves. Open-weight models can be copied, fine-tuned and hosted by anyone, including Mistral’s competitors, and Mistral itself now hosts Chinese open-weight models¹². The price of a unit of AI has been falling and the quality of open models rises quickly. If a free or cheap model is good enough for most business tasks, the premium for a European brand has to carry the whole business.

I rate the moat as contestable. It is real, because European buyers have reasons to want a local supplier. It is not durable, because a credible challenger already exists (a US lab with European hosting) and a policy change could shrink the advantage as quickly as the 2026 events grew it.

VThe numbers we can see

Funding is the best-documented part of the story. The seed round in June 2023 raised $113 million at a reported $260 million valuation. Series A in December 2023 raised €385 million at a reported $2 billion, led by Andreessen Horowitz. A June 2024 round raised €600 million of equity and debt at a reported $6 billion. In September 2025 the Series C raised €1.7 billion at €11.7 billion post-money, led by ASML. In September 2026 the Series D raised €3 billion at more than €21 billion post-money, led by Samsung Electronics, with co-leads Scaleup Europe Fund (managed by EQT) and PSG Equity⁸²¹.

On the arithmetic: Techzine reports that the Series C valued Mistral at €10 billion before ASML’s money¹⁶. By my calculation, the Series D implies a pre-money value of about €18 billion (€21 billion less €3 billion), which is a guess because Mistral gives only the post-money figure. In dollars, TechCrunch converts €3 billion to about $3.58 billion and €21 billion to about $24.39 billion, a rate of roughly 1.16 dollars per euro, which I use throughout⁷. Some outlets quote slightly different dollar totals, such as $24 billion, which is a currency conversion and not a different deal¹³.

Where sources disagree: Sacra’s page gives a headline valuation of $24 billion but also describes a June 2026 raise of $3.5 billion at $20 billion; it does not reconcile the two¹⁸. TechCrunch in July described a rumoured $3.5 billion round at about $23 billion that was unconfirmed⁸. The company’s own September announcement is the one I rely on¹, and Orrick, the law firm that advised Mistral, gives the same €21 billion post-money figure⁶. I could not find a figure for how much of the Series D was debt, if any.

What is not visible: profit or loss, cash balance, gross margin, headcount, the cap table (who owns what) and the terms of the preferred shares. Preferred shares often carry rights to be repaid before ordinary shareholders, and those terms can change who gets what in a sale for less than the headline value. Without them, a headline valuation overstates what an ordinary holder could expect.

Funding history as reported

VIWho runs it and who owns it

The three founders still run the technical side. Others have been added: Johan Bergqvist as chief financial officer, Brian Hall as chief marketing officer, Kamal Brar in partnerships and Marjorie Janiewicz as chief revenue officer⁸¹¹. Jean-Charles Samuelian-Werve, co-founder of the health insurer Alan, sits on the board⁸.

ASML is the shareholder that gets the most attention. It is the Dutch maker of the machines that print the most advanced computer chips. In September 2025 it invested €1.3 billion in the Series C as lead investor. ASML’s annual report on Form 20-F, filed 25 February 2026, states that it holds about 11% on a fully diluted basis, and puts the figure at 11.1% at 31 December 2025⁵. The cash outflow was €1,302.2 million, and the carrying value on its balance sheet at year-end was €1,320.7 million⁵. ASML’s 2025 net sales were €32.7 billion and net income €9.6 billion, so the stake is about 4% of one year’s profit⁵.

ASML’s accounting matters for how to read the stake. It records the holding at cost, less impairment, plus or minus changes from observable price changes in orderly transactions⁵. That is not the equity method that it uses for associates such as Carl Zeiss SMT. A later priced round such as the Series D can count as an observable price change, but I have not seen ASML’s 2026 filings and cannot say whether it has written the holding up⁵.

After the Series D, Techzine, citing the Eindhoven newspaper Eindhovens Dagblad, reported that ASML put in a further €400 million to hold its stake at roughly 11%¹⁵. This is press-reported and ASML made no statement in that article. My check: holding 11.1% of a company worth €21 billion is about €2.3 billion, and ASML would have been diluted to roughly 9.5% without a new cheque, so the order of magnitude fits. Mistral’s list of Series D investors names ASML as an existing investor¹. If those figures are right, ASML’s total cost is about €1.7 billion and the stake is worth about €2.3 billion at the Series D price, on paper.

ASML also gets a seat on Mistral’s strategic committee, held by its CFO Roger Dassen, and a long-term collaboration agreement to explore AI use across its products, research and operations¹⁶¹⁷. An early Reuters-sourced report that ASML would take a majority stake turned out to be wrong; the actual figure is 11%¹⁶. ASML is also now one of the five anchor customers for Mistral’s compute units¹², and it has said Mistral’s coding-capable model plays a role in developing its systems⁵. So ASML is shareholder, customer and partner. This ties the two companies together, and it makes ASML’s endorsement less independent than it might look.

The Series D investors include Samsung Electronics, the EU-backed Scaleup Europe Fund managed by EQT, PSG Equity, Advent, BlackRock-managed funds and the Grand Duchy of Luxembourg as new investors, and existing holders such as a16z, Nvidia, Salesforce Ventures, Bpifrance, General Catalyst, Lightspeed, Index Ventures and DST Global¹. Euronews also lists Microsoft as an existing investor, although Mistral’s own list does not¹⁴¹. Microsoft made a €15 million investment in 2024⁸. The mix of state-linked, industrial and financial investors supports the sovereignty story, and it also means several shareholders are customers or suppliers.

VIIValuation: is the last price supported?

Equity value in US$ billions. Not a share price: a private company has none. Latest mark: primary round, post-money (Series D, above €21B, converted at about 1.16 USD per EUR).

What the last valuation assumes

At about $24.4 billion the price is roughly 24 times the run rate the company says it is on track to reach by the end of 2026, and 61 times the last number it has stated outright. For that to hold, revenue has to keep growing very fast for several years, ending near my base case of $6 billion in 2030 or higher, and investors in 2030 have to pay 7 to 10 times revenue for it. It also assumes Mistral can fund its compute build without giving away a large share of the company again.

Bear, base and bull

• Bear. Run rate reaches only about $0.7B at end 2026 and revenue is $2.5B in 2030. Terminal multiple 4x revenue, discount rate 18%, 20% dilution to 2030. Enterprise value $10B in 2030, $5.0B today, $4.0B to equity. The models fall behind and sovereignty demand cools.
• Base. Run rate about $0.9B at end 2026, just under the company target, and revenue $6.0B in 2030. Terminal multiple 7x revenue, discount rate 15%, 12% dilution. Enterprise value $42B in 2030, $23.5B today, $20.7B to equity (rounded to $21B). Net cash is assumed to be about zero, because round proceeds are spent on compute.
• Bull. Run rate $1.0B at end 2026 and revenue $12B in 2030. Terminal multiple 10x revenue, discount rate 13%, 10% dilution. Enterprise value $120B in 2030, $72.3B today, $65.0B to equity. Europe builds around Mistral as its default AI supplier.

Comparables used

Where I come down

The latest valuation is modestly above my base of $21 billion, about 16% higher, and well above my bear case. My margin-of-safety figure, 25% under base, is $16 billion. To justify $24.4 billion on my model you have to believe the revenue in 2030 is closer to $7 billion, or that the exit multiple is nearer 8 to 9 times, or that dilution is lower. Those are not outlandish. The market may be right even if the price looks high for three reasons. First, strategic investors such as Samsung, ASML and Nvidia may be paying for access and supply, which is worth more to them than a financial return. Second, there are few European alternatives, so scarcity can lift the price. Third, the company has grown from about $20 million to more than $400 million in a year, and a business growing like that can outrun any fixed model. The three listed comparables range from 7x to 57x, which shows how wide the range of defensible multiples is. Mistral’s price sits inside that range. This is research, not a recommendation.

What I cannot judge

I cannot judge profit, cash burn, gross margin or the true size of recognised revenue, because none are published. I cannot see the cap table or share terms. I cannot test the quality of the models against customers’ needs, or see how much of the revenue is from related parties. I have not seen ASML’s 2026 filings, so I cannot say how it has treated the Series D. My range is wide, from $4 billion to $65 billion, and the base case rests on a single set of assumptions that I would revise on new data.

VIIIThe IPO

No IPO plan reported. Mensch said at Davos in January 2025 that Mistral is not for sale and aims for an IPO⁸. In February 2026 he said no IPO was needed that year because debt financing was available, though one remains on the long-term radar⁹. I found no report of a confidential filing, named banks, an exchange, a ticker or a size. The Series D in September 2026 and the compute-unit forward sales point to private funding for now¹¹².

What to watch

• A report of a confidential filing or bank mandate (Paris or Amsterdam listing versus a US listing)
• Audited accounts or a stated revenue figure for 2026
• Whether ARR reaches the $1B target by end 2026
• Any further round, debt raise or secondary sale and its price
• ASML’s 2026 filings for how it carries the stake after the Series D
• Terms and delivery dates for the compute units sold to anchor customers
• Further US export or access rules affecting European buyers of US models

IXWhat could go wrong

What most likely hurts this business is a mismatch between spending and revenue. A gigawatt of computing costs about $50 billion by Mensch’s estimate, and the Series D raised about $3.6 billion¹²⁷. The company plans to own about twice as much compute in five years as it does now¹⁴. That cannot be paid from the latest round alone. It needs more debt, more equity, customer prepayments (the compute-unit forward sales are a step in that direction), or partners who fund capacity, as Microsoft is doing¹². If revenue grows more slowly than spending, the company has to keep raising money, each time on terms set by investors who can see its results.

Second, the models could fall further behind. If European buyers conclude that a frontier US model is worth the sovereignty trade-off, Mistral’s pitch weakens. If they conclude that cheap open models, including Chinese ones, are good enough, its pricing weakens. Mistral sits between the two.

Third, sovereignty demand may be driven by events that fade. A change in US trade or technology policy, or a US lab offering contractual European guarantees, could remove the urgency. Mensch’s two-year warning to the French legislature is also a statement about urgency that cuts both ways¹¹.

Fourth, concentration and circular demand. Several of the largest backers are also customers: ASML is a shareholder and an anchor compute buyer, and Samsung, the Series D lead, is a prospective user, with Mensch expecting similar work as with ASML¹³. Where a shareholder buys from the company, revenue growth looks better than it may be on an arm’s-length basis.

Fifth, reputation and politics. Reuters reported in May 2026 that Mistral defended AI use in warfare after criticism from the Pope⁸. That is not a financial risk by itself, but it shows how a company in the sovereignty business can be pulled into public argument.

What an owner can really take out: nothing, today. There are no dividends, and a private minority holder can only exit through a listing, a sale or a secondary trade. Mensch said in January 2025 that the company is not for sale and plans an IPO, and said in February 2026 that none was needed that year⁸⁹. A sale to a US buyer would cut against the sovereignty case.

What has to go right at the same time. I list four things and give my own rough odds, which are my judgment and not data: the revenue target is reached or nearly so (about 65%); the product keeps up well enough for business buyers (about 70%); sovereignty demand stays strong for five years (about 70%); and the compute build is financed without heavily diluting owners (about 60%). Multiplied, 0.65 x 0.70 x 0.70 x 0.60 comes to about 19%. The events are not independent, so the true figure could be higher or lower, but it shows why a good story and a good outcome are different things.

The next best alternative for the same dollar, in my view, is a listed company that already earns money from AI demand. ServiceNow trades at about 7 times trailing revenue on $15.4 billion of revenue²¹. ASML itself gives indirect exposure to Mistral through its stake, but the stake is small against ASML’s own business⁵. I make no recommendation. I note only that Mistral’s price asks a buyer to accept a long wait and uncertain exit terms, with no cash returns, in exchange for the chance of a much larger business.