Mistral AI, the French model developer, announced a Series D round of 3 billion euros at a post-money valuation above 21 billion euros, led by Samsung Electronics. It plans 1 gigawatt of European compute capacity by 2030, and its new Mistral Large 4 was trained on 3,800 NVIDIA GPUs in its European data centers. Most figures here are company statements.
This post reads Mistral's own announcements in order, explains the terms, and marks what is company reported and what is missing. Mistral is a private company, so its figures are unaudited. Nothing here is investment advice.
Key numbers at a glance
Every figure below is a company statement from Mistral's own news pages, except the revenue projection, which Euronews reported. None is audited.
- Series D size: 3 billion euros. The news page lists the announcement as 8 September 2026.
- Valuation: "more than 21 billion euros," post-money, meaning after the new money is counted.
- Lead investor: Samsung Electronics. Co-leads: Scaleup Europe Fund (managed by EQT) and PSG Equity.
- Previous raise: 1.7 billion euros, announced 9 September 2025, per the news listing. The Series D page says that round was led by ASML.
- Footprint: 20 countries and "125+ global enterprises," per the Series D page.
- Compute plan: up to 1 gigawatt of capacity by 2030, first stated in the 11 August 2026 post and restated as one gigawatt of European capacity in the 28 September 2026 Munich post.
- Revenue: not published by Mistral. Euronews reported on 8 September 2026 that Mistral projects passing 1 billion in annual recurring revenue by the end of 2026 (company projection; the report names no currency).
- Mistral Large 4: about 1 trillion parameters in total, 52 billion active per token. Announced 6 October 2026.
- Large 4 training hardware: 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European data centers.
- Large 4 preview price: $1.36 per million input tokens and $4.18 per million output tokens.
- Large 4 languages: more than 160, including all official EU languages.
- Mistral Compute announced: 11 June 2025, with "tens of thousands of GPUs" available and "rapid expansion" planned.
- Regional Endpoints: generally available from 11 August 2026, with processing in Europe or the US at the customer's choice.
- Emmi AI acquisition: May 2026, adding more than 30 physicists and engineers in Munich.
The funding round
Mistral's Series D page says the company raised 3 billion euros at a post-money valuation of more than 21 billion euros. It calls this "the largest equity fundraising round ever completed by a European technology company." That is Mistral's own claim; no independent ranking is cited for it here. The page says the announcement came "three years after the company's launch" and gives no calendar date, but the company's news listing places it on 8 September 2026.
Equity, in this context, means investors buy shares in the company. Series D is the label for the fourth major round. Post-money valuation is the company's value after the new cash is added.
Who invested matters because it signals who Mistral wants as partners. The page names:
- Lead: Samsung Electronics.
- Co-leads: the Scaleup Europe Fund, managed by EQT, and existing investor PSG Equity.
- New investors: Advent, funds and accounts managed by BlackRock, and the Grand Duchy of Luxembourg.
- Existing investors who took part: a16z, ASML, Belfius, BNP Paribas CIB, Bpifrance, Carmignac, DST Global, Eurazeo, General Catalyst, Headline, Hillspire, Index Ventures, Korelya Capital, Lightspeed, NVIDIA, a solar fund associated with Phoenix Court, and Salesforce Ventures.
The mix has a hardware maker, a chip equipment maker, banks, a state fund, and a government. NVIDIA is both an investor and a supplier. ASML, which builds the machines that make chips, led the previous round and participated again.
What does Mistral say it will spend the money on? The page lists expanding frontier research, scaling compute capacity for training powerful models, and expanding infrastructure, commercial growth and international footprint. It does not disclose revenue, spending amounts, or the split between those uses.
The only public revenue figure is a projection. Euronews, in its 8 September 2026 report on the round, wrote that "Mistral projects it will pass a billion in annual recurring revenue by the end of 2026." The report attributes this to the company, names no executive, and gives no currency. Annual recurring revenue, or ARR, is the yearly value of subscription and contract revenue running at a point in time; it is a rate, not a year's booked sales. A projection of a future rate is not a reported result, so the valuation still cannot be checked against actual sales, and this post does not estimate a valuation multiple.
What "sovereign AI" means here
The page defines sovereign AI along four lines: data kept within an organization's boundaries, models that can be controlled and customized, private and predictable compute, and production systems that can be audited. In plain terms, the pitch is that a bank, a manufacturer or a public agency in Europe can run advanced AI without handing its data or its dependence to a single foreign provider.
Mistral also says customers are not locked into a single vendor's roadmap, pricing or availability. That is a sales claim, and like all such claims it is worth testing in a contract. Open-weight models can be moved between hosts, which supports the claim. The services built around them, such as a managed endpoint with an uptime guarantee, are still tied to Mistral.
Where the compute comes from
Compute is the scarce input, so this is the part of the story to read slowly.
Mistral Compute (June 2025). Mistral announced this offering on 11 June 2025, with launch partners including BNP Paribas, Orange, Schneider Electric, SNCF, Thales and Veolia, among others. It promised NVIDIA reference architectures "with availability of tens of thousands of GPUs, and rapid expansion in the coming years." It described a stack running "from bare-metal servers to fully-managed PaaS," meaning customers can rent raw machines or a fully managed platform. The page names no data center site and gives no exact GPU count. It says the infrastructure is built to meet European regulations and aims to use decarbonized energy.
The 1 gigawatt goal. The target first appears in the summary block of Mistral's 11 August 2026 post, which mentions "plans to build up to 1 GW of capacity by 2030." The Munich announcement of 28 September 2026 restates it more firmly: "to secure Europe's AI sovereignty, Mistral will build one gigawatt of European compute capacity by 2030." Neither page names sites, power sources or a cost. A gigawatt is a billion watts of electrical load, about what a large power station produces. For scale, SpaceX's recent filing reports 1.4 gigawatts of nameplate AI compute (see our post on xAI's filings), so Mistral's 2030 target is in the same range as what one US lab already reports today. That comparison is this post's own, and the two figures are measured in different ways, so treat it as context, not a ranking. For cost per gigawatt, see our explainer on data center costs.
How customers fund it. The 11 August post says Mistral is assembling an anchor group of enterprises making multi-year commitments to fund European infrastructure. It introduces European Compute Units, or ECUs, which convert those commitments into multi-year access to Mistral-built infrastructure usable across Mistral Compute products. Named supporters with quoted executives include Factory, Amadeus, ASML, Capgemini, Caisse des Dépôts and CMA CGM. In other words, part of the plan is that customers pay in advance for capacity, which is a way to finance construction without borrowing the full amount. Compare that with the structure of Anthropic's compute commitments, where the lab itself signs the long contracts.
What Large 4 tells us. Mistral's 6 October 2026 post says Mistral Large 4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European data centers. That is the most concrete hardware number Mistral has published. It is a count of chips for one training run, not of everything the company owns.
The models and the open-weight bet
Open-weight means the trained model files are published for others to download and run. It is not the same as open source, where the code, data and training details are also available. See our explainer on that difference.
Mistral Large 4, per its 6 October 2026 post, is a natively multimodal model, meaning it accepts images as well as text, with 1 trillion parameters in total and 52 billion active. It is a mixture of experts design: the model contains many specialist sub-networks and activates only a few for each token, so it costs less to run than a dense model of the same total size. The model card describes it as a hybrid instruct and reasoning model.
Three details need care.
- Weights are not out yet. The post says weights will be released by the end of October. A public preview API is available on Mistral Studio now. The license is not stated in the post. Check the model page for the release date and license terms before relying on either.
- Price. The preview lists $1.36 per million input tokens and $4.18 per million output tokens.
- Benchmarks are vendor reported. Mistral lists 61.7 percent on DeepSWE v1.1, 59.4 percent on SWE-Atlas-QnA, 28.3 percent on Terminal-Bench 4, 93 percent on Cybench, and a 93.3 percent resistance rate on a prompt injection benchmark from Lakera. In a human coding evaluation it cites, it ranks second of five. These are Mistral's chosen results, and the right test is your own workload.
Mistral has also been building around the model. On 16 March 2026 it announced it had joined the NVIDIA Nemotron Coalition as a founding member, a group aiming to advance open frontier models, and released Mistral Small 4 the same day. On 11 August 2026 it added regional inference, third-party open models on its platform starting with Z.ai's GLM-5.2, and a Priority Tier with an uptime guarantee in public preview. On 28 September 2026 it opened a hub in Munich after acquiring Emmi AI in May 2026, with partners BMW, Siemens Energy and the Technical University of Munich. The Munich post says the focus is physics and industrial AI, such as crash simulation and aerodynamics.
The pattern is a company trying to be a full stack European provider: models, compute, and applications for industry. It is a different strategy from a lab that sells one flagship chat product.
Regional endpoints: a feature that matters for compliance
For a security or compliance reader, the 11 August 2026 post has the most practical item. Mistral Regional Endpoints are generally available. Customers choose whether inference runs in Europe or the US. Processing stays in the chosen region except for "limited, safeguarded transfers to sub-processors," as described in Mistral's Trust Center. A sub-processor is another company that handles data on the provider's behalf.
Mistral says it is "the only European AI lab to offer both" regional processing choice and an SLA-backed service level. That is Mistral's statement, not an independent comparison. If data residency is a requirement, ask for the sub-processor list and read the exceptions, since the wording leaves room for limited transfers.
Why a private company's pages need extra care
Mistral is not a listed company, so it does not file quarterly accounts. Everything in this post comes from announcements Mistral chose to publish. That is normal for a private firm, but it changes how you should read it.
A listed company must disclose revenue, losses and risks in a standard format that an auditor and a regulator can examine. A private company can pick which numbers to show. Mistral shows the funding amount, the valuation, the investor list, the country count and a customer count, and it does not show revenue, losses or total compute spending. None of that implies a problem. It means you cannot judge financial strength from these pages alone.
If you are a customer making a multi year commitment, for example through the European Compute Units scheme, the sensible step is to ask for what you need in the contract: financial statements under confidentiality, service level terms, exit and data return provisions, and clarity on which sub-processors touch your data. Those questions are standard for any supplier of critical services, and they matter more as the commitment grows longer. The same advice applies to any young AI vendor, in any country.
What is firm and what is uncertain
Firm, because it is Mistral's own published statement:
- The size, valuation, lead and named investors of the Series D.
- The existence of the 1 gigawatt by 2030 goal (first stated 11 August 2026, restated 28 September 2026), the ECU commitment scheme and Regional Endpoints.
- The Large 4 training chip count, parameter figures and preview price.
Uncertain or missing:
- Revenue and spending. Not published by Mistral. The only figure is the company's projection, reported by Euronews, of passing 1 billion in annual recurring revenue by the end of 2026.
- The "largest" claim. Mistral's own statement.
- Where the gigawatt will be built, and how it is financed. The pages give no sites, no cost, and no power sources.
- Large 4 weights and license. Promised for the end of October, license not stated.
- Current GPU totals. Only the 3,800 figure for one run and "tens of thousands" from 2025.
What to watch next
- The Large 4 weights release. Check the date, the license and whether the multimodal parts are included.
- A named data center site and power contract. The 1 gigawatt goal needs both.
- How many enterprises sign ECU commitments, and for how long.
- Independent evaluations of Large 4, especially on coding and security tasks.
- The Series D investors' own announcements. Samsung's and EQT's statements may add detail the Mistral page lacks.
- European regulation. Data residency and AI rules will shape how the regional endpoints are used.
Quick check
How much did Mistral raise and at what valuation?
Mistral says it raised 3 billion euros in a Series D at a post-money valuation of more than 21 billion euros, led by Samsung Electronics. These are company figures for a private company and are unaudited.
Is Mistral Large 4 open-weight?
Mistral describes it as an open-weight model, but its 6 October 2026 post says the weights will be released by the end of that month. A preview API is available now. The license is not stated in the post, so check the model page when weights appear.
How much compute does Mistral have?
Mistral has not published a total. It says Large 4 was trained on 3,800 NVIDIA Grace Blackwell GPUs, announced Mistral Compute in 2025 with tens of thousands of GPUs available, and plans 1 gigawatt of European capacity by 2030.
If you use open-weight models in your own products, it helps to keep a record of what you ship. Safeguard offers SBOM and dependency scanning, and a free scan is available at https://gold.safeguard.sh/scan.
Sources
- Mistral AI news listing, checked 19 October 2026: https://mistral.ai/news
- Mistral AI, Series D announcement: https://mistral.ai/news/mistral-makes-sovereign-open-weight-ai-to-frontier/
- Mistral AI, "In-region inference, open models, and new European infrastructure for sovereign AI," 11 August 2026: https://mistral.ai/news/regional-inference-open-models-new-compute/
- Mistral AI, "Introducing Mistral Large 4," 6 October 2026: https://mistral.ai/news/mistral-large-4/
- Mistral AI, "Hallo, Deutschland!," 28 September 2026: https://mistral.ai/news/hallo-deutschland/
- Euronews, report on Mistral's Series D, 8 September 2026: https://www.euronews.com/2026/09/08/mistral-ai-raises-record-3-billion-in-samsung-led-funding-round
- Mistral AI, "Mistral Compute," 11 June 2025: https://mistral.ai/news/mistral-compute/
- Mistral AI, partnership with NVIDIA on open frontier models, 16 March 2026: https://mistral.ai/news/mistral-ai-and-nvidia-partner-to-accelerate-open-frontier-models/
- SpaceX, second quarter 2026 earnings release (for the 1.4 gigawatt comparison): https://www.sec.gov/Archives/edgar/data/1181412/000162828026052515/earningsreleaseq22608042.htm
Last reviewed: 19 October 2026.
