Executive brief
In four rounds between June 2023 and September 2025, Mistral AI raised a disclosed €105m, €385m, €600m and €1.7bn — a capital stack assembled faster than any other European software company of the decade, and one whose published valuations do not agree with each other. The Series B is reported as $6bn by one set of publishers and €5.8bn by another, with no bridge between them. The Series C post-money is €11.7bn in the company's own release; the widely circulated "$14bn" is a currency relabelling of the same number, not a second data point, and we treat it as such. The lead investor of that round is not a fund but a machine builder: ASML, reported at €1.3bn for roughly 11% fully diluted. Underneath the equity sits a physical bet — a datacentre at Bruyères-le-Châtel, Essonne, whose capacity is published as 40 MW in one registry and 44 MW by another observer, and whose GPU count is 18,000 in a 2025 report and 13,800 in a 2026 one. We print every one of those splits with its publisher and date and reconcile none of them. What no source supports — an audited revenue figure, a company headcount, a state equity or grant amount specific to Mistral — is named and refused at the end.
I. The mechanism: a model company that had to buy a power station
The usual reading of Mistral AI is a talent story: researchers from DeepMind and Meta leave, incorporate in Paris in April 2023, and raise an unusually large seed. That reading is accurate and analytically useless, because it explains the first €105m and none of the €2.7bn that followed. The mechanism that explains the rest is a shift in what the company is buying.
Rounds one and two bought people and training runs on rented capacity. Round four bought siting: land, grid connection, cooling and a multi-year power contract in the Essonne département, south of Paris, developed with the datacentre operator Eclairion alongside MGX, Bpifrance and Nvidia. Once a model company owns megawatts rather than renting instance-hours, its cost base stops looking like software and starts looking like industry — depreciation schedules, power-purchase exposure, an interconnection queue. That is the single most important thing a candidate or an analyst can understand about the capital stack, and it is why the ASML investment is not a curiosity.
ASML does not make models; it makes the lithography machines without which no one makes the chips the models run on. A machine builder taking a reported 11% fully diluted position at €1.3bn in a model developer is a vertical claim on the whole stack — tools, silicon, capacity, weights — held on one continent. Whether that claim pays is unknown. That it was made, at that size, in September 2025, is documented.
II. The capital record, round by round, as published
We give each round its own date, its own lead and its own disclosed amount, and we do not sum them into a headline "total raised", because the published valuations attached to them are inconsistent and a total would inherit that inconsistency silently.
- Seed — 13 June 2023, €105m (reported as $113m). Led by Lightspeed Venture Partners at a reported $260m valuation, roughly four weeks after incorporation and before any model shipped. The euro and dollar figures are the same money at two exchange rates, and both are printed here because both are what publishers wrote.
- Series A — 11 December 2023, €385m. Co-led by Andreessen Horowitz and Lightspeed. Reuters reported the valuation as undisclosed. A "roughly $2bn" figure was attributed by Bloomberg to an unnamed source and repeated in trade coverage. We therefore hold that no Series A valuation is on the record: an unnamed-source number is evidence of reporting, not of a price.
- Series B — 11 June 2024, €600m. Led by General Catalyst. The post-money is published as $6bn (TechCrunch, CNBC) and as €5.8bn (Bloomberg). At mid-2024 rates those are not the same number, and no publisher states which currency the term sheet used. The split stands unreconciled.
- Series C — 9 September 2025, €1.7bn. Led by ASML at a reported €1.3bn for approximately 11% fully diluted, with a post-money of €11.7bn in the company's own release. The "$14bn valuation" in much of the coverage is that same €11.7bn converted; it is one figure in two currencies and is not treated as corroboration.
- Infrastructure financing — reported March 2026, approximately $830m. Tied specifically to the datacentre build rather than to the operating company, per European broadcast and trade reporting. Its instrument — equity, debt, vendor facility, or a mix — is not published, and we do not infer one.
The analytically interesting property of this sequence is not its speed but its investor composition drift: venture (Lightspeed, a16z), then growth (General Catalyst), then strategic industrial and sovereign-adjacent capital (ASML, MGX, Bpifrance, Nvidia). Each stage buys a different risk. The last stage buys the risk that Europe does not hold this layer at all.
III. The physical layer: Essonne, and two numbers for everything
The Bruyères-le-Châtel project is the point where the capital becomes concrete, and it is also where the public record is weakest. Capacity is published as 40 MW in a datacentre-industry registry and as 44 MW by a separate market observer. Accelerator count is published as 18,000 GPUs in March 2025 trade coverage and as 13,800 in 2026 coverage.
A careless dossier would pick the larger of each pair, or average them. Both moves are wrong, and for a reason worth stating plainly: the two GPU counts are separated by roughly a year and may describe different things — a planned fleet versus an installed one, a whole-site figure versus a first-phase figure, or two different accelerator generations counted per package rather than per die. No publisher states which. When the scope of a number is unknown, the number is not comparable, and pretending otherwise manufactures a trend out of a definitional difference. So: two capacity figures, two GPU figures, four publishers, no reconciliation, and no derived metric — no MW-per-GPU, no implied FLOPs, no cost-per-accelerator — built on top of them.
IV. The distribution problem, and who solved it for whom
A model with no distribution is a research artefact. Mistral's answer has been a partnership ladder, and the ladder is dated:
- Microsoft / Azure — 26 February 2024. Mistral Large launched first on Azure alongside a small equity investment; Reuters reported that the arrangement drew EU antitrust attention. This is the round's most instructive item, because it is where "European champion" and "hyperscaler distribution" touch: the fastest path to enterprise reach ran through a US cloud.
- IBM watsonx — 21 May 2024, with Mistral Large 2 subsequently available on the platform.
- Nvidia — 11 June 2025, announced around sovereign-AI infrastructure, i.e. the supply side of the Essonne bet rather than the demand side.
- Stellantis — February 2025 and CMA CGM — 6 April 2025, €100m over five years. These matter more than their size: they are named industrial customers with a published amount, in a market where most "enterprise adoption" is unnamed and unquantified.
- Helsing — company page only. A defence-adjacent collaboration appears on a company page with no independent date and no published terms. We record its existence and refuse to characterise it.
V. The model line, because the capital only makes sense against it
Licensing, not benchmarks, is the signal here. Mistral 7B (27 September 2023) and Mixtral (11 December 2023) shipped under Apache 2.0. Mistral Large (26 February 2024) did not. Medium 3 (7 May 2025) is proprietary. Magistral (10 Jun 2025) split the difference, with the small variant under Apache 2.0. Then Mistral 3 / Large 3 (2 December 2025) returned to Apache 2.0 at 41B active parameters of 675B total — a sparse architecture released openly at frontier scale. Le Chat's relaunch and mobile apps (6–7 February 2025) added a consumer surface.
Read as a sequence, that is not drift; it is a barbell. Open weights buy developer gravity and sovereignty credibility; closed weights and hosted products buy the revenue that services €1.7bn of equity. The strategic question the record does not answer is which end of the barbell carries the margin — and it does not answer it because no audited revenue exists.
VI. What we refuse to publish
These are the figures a reader would most like, and precisely the ones we will not invent. Each refusal is a statement about the public record, not about the company.
- Revenue or ARR. No audited figure exists. The estimates in circulation are mutually contradictory: about €300m ARR (September 2025), about $400m ARR, and a $1–5bn range from a data platform, while one tracker's $100m 2025 baseline contradicts another's. A range spanning an order of magnitude is not a measurement, and we publish none of them as fact.
- Headcount. Five third-party values sit in the same window: 350, about 1,000, about 1,258, 1,424, and a 1,000–2,000 range. No company-published figure exists. We print the spread and treat the number as unknown.
- Any state-to-Mistral equity or grant amount. Bpifrance's €10bn ecosystem deployment (27 March 2025) and the France 2030 €392.5bn AI plan are ecosystem-wide instruments that reference the sector; neither is a Mistral line item, and converting one into a company subsidy figure would be a fabrication.
- A single Series A valuation, for the unnamed-source reason above.
- A final MW or GPU figure for Essonne, and any metric derived from one.
- An ASML collaboration scope beyond the equity stake, and any Helsing date or terms.
- One tracker's $23bn valuation, which is uncorroborated by any other publisher and is not carried here even as a range endpoint.
VII. The analytical conclusion
The defensible reading of Mistral's capital record is that Europe has bought an option, at a price it can afford, on the one layer of the AI stack it was closest to losing entirely — and that the option's cost has migrated from salaries to megawatts within thirty months. The unresolved risks are legible without any private data: a revenue base nobody can audit against a capital base everybody can, distribution that partly runs through the hyperscalers this investment was meant to counterbalance, and an infrastructure programme whose two published capacity figures differ by four megawatts because no one has had to be precise in public yet.
For a reader deciding whether to work at this layer, the useful test is not the valuation. It is which clock a given role sits on: the research clock, funded by equity and measured in releases, or the infrastructure clock, funded by a power contract and measured in commissioning dates. The companion dossier filed on this anchor treats the roles themselves — what is actually posted, in which cities, and what the evidence does and does not support about moving into them.