Mistral has assembled ASML, Amadeus, Capgemini, Caisse des Dépôts, and CMA CGM to pre-fund a 1-gigawatt European compute buildout targeting 2030. The funding mechanism—European Compute Units (ECUs)—converts multi-year financial commitments into capacity credits for inference, fine-tuning, or model adaptation. Named industrial balance sheets are locking in 5-year commitments to underwrite infrastructure Mistral has not yet built.
The first milestone is a data center at Bruyères-le-Châtel, south of Paris: 13,800 Nvidia GB300 GPUs, 44 megawatts of compute, financed by $830M in debt closed this year. Intermediate targets: 200 MW across Europe by end of 2027, then 1 GW by 2030. CEO Arthur Mensch estimates 1 GW deployed at $50B capital. Europe's AI compute base stands at 3.6 GW versus 31 GW in the US, per Alice Labs. The gap Mistral targets is structural.
| Phase | Capacity | Target Date | Financing / Capital |
|---|---|---|---|
| Phase 1 — Bruyères-le-Châtel | 44 MW / 13,800 Nvidia GB300 GPUs | 2025 (debt closed) | $830M debt round |
| Phase 2 — European expansion | 200 MW across Europe | End of 2027 | ECU pre-funding (multi-year industrial commitments) |
| Phase 3 — Full buildout | 1 GW across Europe | 2030 | ~$50B total capital (CEO estimate) |
ECUs operate as infrastructure commitments, not cloud contracts. CTO Timothée Lacroix stated partner commitments run five years with no early exit. ECU holders can deploy units across whatever Mistral Compute products exist at deployment. Mistral has not disclosed ECU compute value, unit cost, carry-over terms for unused allocation, or remedies for schedule misses. Teams evaluating the program must negotiate directly.
Mistral released two product updates. Regional endpoints let customers pin inference to Europe or the US at the API level, addressing data-residency and compliance. A Priority Tier offers 99.5% uptime SLA at 1.75× standard pricing, now in public preview. The platform opens to third-party open-weight models; Z.ai's GLM-5.2 launches first with a 1M-token context window priced at $1.40/M input tokens and $4.40/M output tokens. Third-party models run on shared infrastructure under the same regional controls.
| Feature | Details | Pricing / SLA |
|---|---|---|
| Regional Endpoints | Pin inference to EU or US at API level; addresses data-residency and compliance | Standard pricing |
| Priority Tier (public preview) | 99.5% uptime SLA; first-party models | 1.75× standard pricing |
| Third-party open-weight models | Z.ai GLM-5.2 — first on platform; 1M-token context window | $1.40/M input tokens · $4.40/M output tokens |
Anchor customers signal real demand. ASML led Mistral's €11.7B ($13.4B) Series C. CMA CGM's chair confirmed Mistral deployed across thousands of employees. Amadeus CEO Luis Maroto cited "capacity, deployment control, and operating continuity" as drivers. These are not pilots—the $830M debt round signals demand repeatable at scale. Microsoft has separately committed to rent capacity from the same European facilities, making an American hyperscaler an effective anchor tenant of Mistral's branded sovereign EU infrastructure.
Sovereignty claims carry limits. Microsoft consuming capacity and Z.ai's GLM-5.2 running on-platform blur a clean EU-control narrative. Mistral's regional inference toggle addresses data-residency, but physical compute control differs from API routing.
Mistral's independent hosting path now has concrete capacity roadmap and named industrial backers. The 44 MW facility is financeable and GPU-sourced. The 1 GW end-state requires five years of execution against $50B capital. Back near-term milestones; pressure-test long-term ones before converting ECU commitments into strategic dependencies.