🚀 The Rise (2023–2024)
On Feb 26, 2024, Mistral launched Mistral Large and claimed it was "the world's second-ranked model available through an API, behind GPT-4" — a 14-month-old French startup compared directly to OpenAI. 0:00
The founding bet: Arthur Mensch (Google DeepMind) + Guillaume Lample & Timothée Lacroix (Meta, Llama team), April 2023. Their first model, Mistral 7B (Sept 2023), "beat models roughly twice its size." The claim: better data + better training beats a bigger budget. And they published the weights — through 2023 and most of 2024, "Mistral's models were the best you could get that way." 0:47
The EU AI Act leverage: France pushed for lighter rules on open models during AI Act negotiations — "and Mistral was the reason it pushed." The French government began treating Mistral as the company Europe would build its AI industry around. 1:25
The fall in numbers: today on Artificial Analysis, Mistral's flagship sits at 23rd place, scoring just 30 — less than half of leading models. By late 2025, Mistral was only 6.8% of HuggingFace downloads for 1B+ models, vs Meta's 23% and Alibaba's 20%. 0:19
🤝 The Microsoft Deal
Alongside Mistral Large, Mistral announced a Microsoft partnership: €15M investment + Azure distribution. "European lawmakers were angry." 1:50
The contradiction: Mistral spent months telling Europe it needed lighter regulation to survive as the independent alternative to American labs. It got the lighter regulation — then took Microsoft's money weeks later. Kai Zenner (EP digital policy adviser): "Parliament was extremely furious." Mistral's answer: "commercial deals pay for the research." 2:05
💾 The Chips Problem
"Mistral's first problem was that it did not have enough chips." In April 2024, Mensch said Mistral had ~1,500 H100s — "a few percent of what the leading American labs had." Efficiency was a philosophy, "but it was also the only option they had." 2:26
The math that broke: efficiency worked while a competitive model cost tens of millions to train; it stopped working when frontier training cost billions. By Dec 2025: Mistral raised $2.7B at $13.7B — vs OpenAI's $57B at $500B and Anthropic's $45B at $350B. "Mistral was fighting companies with roughly 20× its money." 2:58
The Chinese made the same bet with more chips: DeepSeek, Alibaba's Qwen, Moonshot, Z.AI built sparse Mixture-of-Experts models (a design Mistral helped popularize with Mixtral in 2023), released them under Apache 2.0/MIT, and "all the Chinese models beat Mistral significantly — every one downloadable, every one roughly doubles Mistral's score." 3:20
📉 The Three Releases That Showed the Fall
| Release | Date | What happened |
|---|---|---|
| Mistral Large 3 | Dec 2, 2025 | 675B total / 41B active, 256k context, Apache 2.0, ~3,000 H200s. "Most capable model ever built" — but independent tests put it below average for its size, priced above average. |
| Mistral Medium 3.5 | Apr 29, 2026 | Developers reacted badly: "Qwen 3.6 is 4.7× smaller and scores about the same on coding," "not the best at anything," "costs several times more." |
| The promised model | Jul 2026 | In the same two weeks as Kimi K3, GPT-5.6, Grok 4.5, Opus 5, and Inkling — Mensch promised an open-weight model in early access. "It still hasn't shipped. No weights, no benchmark scores." |
Mensch's own admission (LinkedIn): "Mistral does not yet own the best language models, but has constantly reduced the gap." 5:35
📈 The Revenue Paradox
"Mistral's revenue went the other way" — the models got worse relative to everyone else while revenue exploded. 6:13
| Metric | Value |
|---|---|
| ARR end of 2024 | ~$16M |
| ARR Jan 2026 | ~$400M (~20× in one year) |
| Mensch's 2026 target | €1B+ by end of 2026 |
| Series C (Sept 2025) | €1.7B at €11.7B, ASML led with €1.3B for 11% (largest shareholder) |
| Microsoft (Aug 2026) | Multi-billion commitment to rent compute from Mistral's European data centres |
| Debt (Mar 2026) | $830M from seven banks for a Paris data centre — 13,800 GB300 chips, 44MW |
🇪🇺 The Real Product: Sovereignty
"Mistral sells European companies the promise that their data never leaves Europe. And that is where Mistral is winning." European banks, hospitals, defence contractors, and government departments follow rules about where data is processed and who controls the computers — "for a lot of them, the best model is the one they are not allowed to use." 7:05
The infrastructure bet: $1.4B committed in Sweden, the Paris data centre, and "European hardware under European law — customers can download the weights and run everything inside their own buildings." 7:47
The turning point — US export controls: in June 2026 the US Commerce Department cut off access to Anthropic's Fable 5 outside the US (lifted July 1st). "European companies lost access for 3 weeks — but they had just watched a foreign government switch off their frontier model, then switch it back on." That gave Mistral enormous leverage. Weeks later, Samsung was in talks to invest up to €1B at a ~€20B valuation. 8:08
The verdict: "Mistral set out to prove a European company could compete with the best AI labs on far less money — it wasn't able to. But what it proved instead is that European banks and governments will buy a less capable model if it keeps their data inside Europe. That's why investors are discussing a €20B valuation for a company whose models keep sliding down the rankings." 8:45
✅ Key Takeaways
- Mistral lost the model race on money. $2.7B raised vs OpenAI's $57B — "fighting companies with 20× its money" — and the efficiency philosophy stopped working when frontier training hit billions.
- The Chinese out-executed its open-source bet. Sparse MoE + Apache 2.0/MIT at bigger scale: "all the Chinese models beat Mistral significantly, every one downloadable."
- Three releases tell the story — Large 3 (below-average, above-average price), Medium 3.5 (developer backlash), and a promised open model that still hasn't shipped.
- Revenue decoupled from model quality. ~$16M → ~$400M ARR in a year while the flagship slid to 23rd place.
- The real product is data sovereignty. European banks/governments buy "the model they're allowed to use," not the best one.
- US export controls supercharged it. A 3-week Fable 5 shutdown turned sovereignty from a nice-to-have into a strategic necessity — hence the ~€20B valuation talk.