Back to Insights

Technology

Mistral: Europe's champion and the fine print in its licences

Mistral did what Aleph Alpha stopped doing: it kept training models and built its own compute. EUR 3 billion in September, a 1.4 GW campus on the way and three types of licence that are worth telling apart.

By Sebastián Rodríguez Sánchez· Chief Executive Officer, Bajío Ventures· 6 min
Compute infrastructure and automated manufacturing

Energy and infrastructure for compute and industrial loads in the Bajío.

Let's discuss your project

Bajío Ventures · October 2026 · Sovereign AI, 2 of 5

Mistral AI was founded in Paris in 2023. In three years it became the European reference for language models, and it took the opposite path to Aleph Alpha: it did not stop training models, and it also started to build its own compute. For an infrastructure investor, it is the case that shows how much energy sits behind the word “sovereignty”.

The capital

Round Date Amount Lead Valuation
Series C Sep 2025 EUR 1.7 billion ASML, with EUR 1.3 billion EUR 11.7 billion
Series D 8 Sep 2026 EUR 3 billion Samsung Electronics More than EUR 21 billion

The co-leads of the Series D were Scaleup Europe Fund and PSG Equity. ASML, Nvidia, Bpifrance, the Grand Duchy of Luxembourg, funds managed by BlackRock, a16z and General Catalyst also took part, among others. Mistral operates in 20 countries and says it serves more than 125 companies; it names Airbus, ASML and HSBC.

Read the shareholder list with the sovereignty question in mind. The lead of the most recent round is Korean. The chip maker, a US company, is also a partner. French and Luxembourg public capital sits next to Silicon Valley funds. The European champion is European by headquarters, by regulation and by customers, not by cap table.

The compute, measured in megawatts

  • Bruyères-le-Châtel, south of Paris. In March 2026 Mistral obtained USD 830 million of debt from seven banks for a center with 13,800 Nvidia chips, which was planned for the second quarter of 2026. The site has been reported at 44 MW.
  • 1.4 GW campus in the Paris region. Announced in May 2025 with MGX (Abu Dhabi), Bpifrance and Nvidia: EUR 8.5 billion, construction from the second half of 2026 and operation in 2028.

For reference: 1.4 GW is five times the installed capacity of all the data centers in Mexico, which the industry association puts at 279 MW.

France offers that power on the back of its nuclear fleet. From an energy point of view, the EUR 109 billion package that it announced in February 2025 reads as an offer of firm electricity.

The models and their three licences

Mistral publishes many of its models with open weights, but not all of them under the same conditions.

Licence What it permits Models
Apache 2.0 Commercial use without restriction Mistral Large 3 (675 billion parameters, 41 billion active; Dec 2025) · Ministral 3 (3, 8 and 14 billion) · Mistral Small 4 (119 billion, 6 billion active; Mar 2026) · Mistral Nemo · Devstral Small
“Modified MIT” Open, with a revenue cap Devstral 2 (Dec 2025) · Mistral Medium 3.5 (Apr 2026)
Research or non-production licence Non-commercial use only Mistral Large from 2024 · Pixtral Large · Codestral from 2024

The Devstral 2 licence is the one that shows the fine print: it does not authorize use by companies with global consolidated revenue above USD 20 million in the previous month. Those companies must request a commercial licence, which Mistral grants at its discretion, or use its hosted services. The Medium 3.5 licence file sets the same threshold. The model card does not mention it: read the licence file before you deploy.

For a mid-sized industrial group, USD 20 million a month is not a high threshold. The Tier 1 groups with plants in the Bajío exceed it.

Which one works inside a plant

  • Ministral 3 (3 to 14 billion parameters, Apache 2.0). Workstation size, or a single GPU. It is the entry point to classify reports, read manuals or assist maintenance.
  • Mistral Small 4 (Apache 2.0). Mistral states a minimum infrastructure of 4× NVIDIA HGX H100, 2× HGX H200 or 1× DGX B200. That is already data center equipment.
  • Mistral Large 3 (Apache 2.0). One node of 8 GPUs (A100 or H100) for inference, according to Mistral. Training used 3,000 H200 GPUs.

What the case leaves

  1. Training frontier models requires capital in the billions and electricity in the hundreds of megawatts. Mistral obtained both; Aleph Alpha did not.
  2. Openness is a commercial strategy, not a donation. The small models are given away to win adoption; some of the large ones carry conditions.
  3. The version published under Apache 2.0 stays open. The licence of a version that is already published cannot be withdrawn. What is not guaranteed is the next version.

In this series: Guide: what sovereign AI is · 1. Aleph Alpha · 2. Mistral · 3. United States · 4. China · 5. Spanish-language models and Mexico

Sources: Mistral AI, Series C, Sep 2025 · Tech.eu, Series D, 8 Sep 2026 · The Next Web, data center debt, Mar 2026 · Mission Open Data, Mistral data center · RCR Wireless, 1.4 GW campus, May 2025 · TechCrunch, French package, Feb 2025 · Mistral AI, Mistral 3, Dec 2025 · Mistral AI, Small 4, Mar 2026 · Mistral AI, Medium 3.5 model card · Hugging Face, Medium 3.5 licence · Hugging Face, Devstral 2 licence · Mistral AI, model cards for Large 2.1, Pixtral Large and Codestral 24.05 · ASML, investment in Mistral, Sep 2025 · DCD, CFE and MEXDC, Sep 2026

Bajío Ventures — a Mexican energy-infrastructure investment group based in Guanajuato: capital, a data platform on the region’s industrial base, and a German industrial channel, with its own advisory presence in Berlin.

Energy and infrastructure for compute and industrial loads in the Bajío.

Let's discuss your project
ShareLinkedInWhatsApp

This article is also available in Spanish. Read in Spanish

Get every Bajío Perspectives note in your inbox.

Keep reading

Newsletter

Bajío Perspectives

Short analysis on energy, infrastructure and industry in central Mexico. One email when we publish; nothing else.

Unsubscribe at any time. · Privacy notice