#1 of 30 · Open-weight lab
Mistral AI
The only frontier lab with a peer-reviewed life-cycle assessment.
In plain words
How scores workMistral scores 75 out of 100 — leader. Discloses real numbers and runs mostly on clean power. Still has a footprint, but you can see it. Confidence in this score is good confidence (B).
What helps
- Efficiency · Mistral's catalogue leans on small and mid-size dense and sparse models (7B–123B) rather than trillion-parameter systems.
- Carbon · Training and serving happen mostly on the French grid (nuclear and hydro dominated) — one of the lowest-carbon grids among the companies in the index.
- Transparency · LCA conducted with Carbone 4 and ADEME, peer-reviewed by Resilio and Hubblo under the AFNOR Frugal AI methodology.
What hurts
- Water · No fleet water-use-effectiveness (WUE) figure is published; the LCA is model-level, not facility-level.
- Commitments · No dated net-zero, 24/7 carbon-free energy or water-replenishment target has been published.
What nobody knows
- Location-based Scope 2 reported
- Water-use effectiveness reported
Mistral published the first full life-cycle assessment of a large language model in July 2025, covering training, inference, hardware manufacturing and data-centre construction, and proposed three standard indicators for the industry. Its models are hosted largely on French and Nordic grids, among the lowest-carbon electricity in the world, and in 2026 it financed its own 44 MW, 13,800-GPU cluster south of Paris.
Key figures
Per-prompt and fleet metrics. Hover a title for what it means; open “How we got this number” for the working.
ecogpt.co estimate; no company figure.
How we got this number
Mistral reports carbon and water per 400-token response but not energy. A 123B dense model on H100-class hardware sits between Epoch's 0.3 Wh for a typical GPT-4o query and the 1.6 Wh of a diffusion image; a 400-token answer is longer than a median prompt. Uncertainty roughly ±3×.
Full scope — includes water consumed generating the electricity, so not comparable with Google's on-site figure.
Model-level, not a company inventory.
How we got this number
Mistral has no company inventory. Its own 44 MW cluster comes online in 2026 on top of leased capacity, so compute — and the emissions behind it — roughly double, on a very low-carbon grid.
How we got this number
No fleet WUE published. The Essonne site (Eclairion) is liquid-cooled and European hyperscale fleets report 0.2–0.5 L/kWh; 0.4 is the midpoint.
How we got this number
France's grid is roughly 90% nuclear, hydro and renewables; Mistral has not published a procurement figure.
Hardware & materials
The chips, power, copper and future e-waste behind the fleet — the part sustainability reports leave out
GPUs, TPUs or custom accelerators the company runs. Analyst figures are in H100-equivalents, company figures in physical chips.
How we got this number
Own cluster at Bruyères-le-Châtel; leased capacity in France and Sweden on top.
Megawatts the AI fleet pulls. One megawatt runs roughly 750 US homes.
How we got this number
Target: 200 MW across Europe by end-2027.
Terawatt-hours per year. One TWh is a year of electricity for about 95,000 US homes.
How we got this number
44 MW × 8,760 h × 80% utilisation. Compute fleet only.
CO₂e emitted making the chips and servers before they answer a single prompt.
How we got this number
13,800 accelerators × 514 kg CO₂e each (chip plus share of the host server). Emitted in manufacturing, before the first prompt is served.
Tonnes of server hardware that will be retired in three to five years.
How we got this number
13,800 accelerators × 17 kg of server hardware each. Tomorrow's e-waste unless reused; generative AI could produce 2.5 Mt of it a year by 2030.
Tonnes of copper in cabling, busbars and transformers.
How we got this number
44 MW × 27 t of copper per MW (cabling, busbars, transformers).
Diesel backup units or gas turbines at the data centres, permitted or observed.
How we got this number
44 MW ÷ 2.6 MW per generator, the average size permitted at US data centres. Permit counts are not published fleet-wide; this is the number a fleet this size normally carries.
Share of retired servers and parts given a second life.
Kilograms of antimony-equivalent, the life-cycle unit for scarce minerals consumed.
How we got this number
Antimony-equivalent depletion of scarce minerals; 0.16 mg per 400-token response.
Score breakdown
Each pillar is 0–100; weights are fixed across the index
- A 400-token Le Chat response is reported at 1.14 gCO₂e and 45 mL of water as a marginal, full-scope inference figure (electricity-generation water included). Mistral AI
- Mistral's catalogue leans on small and mid-size dense and sparse models (7B–123B) rather than trillion-parameter systems. The Decoder
- Training and serving happen mostly on the French grid (nuclear and hydro dominated) — one of the lowest-carbon grids among the companies in the index. Mistral AI
- Mistral Large 2's full life cycle emitted 20.4 ktCO₂e over 18 months, 85.5% of it from training; this is disclosed, not estimated. Mistral AIThe Decoder
- 281,000 m³ of water over Mistral Large 2's life cycle; 45 mL per 400-token response because the figure includes water consumed to generate the electricity, not just on-site cooling. Mistral AI
- No fleet water-use-effectiveness (WUE) figure is published; the LCA is model-level, not facility-level. Mistral AI
- LCA conducted with Carbone 4 and ADEME, peer-reviewed by Resilio and Hubblo under the AFNOR Frugal AI methodology. Mistral AIThe Decoder
- Proposes three reportable indicators for every model: absolute training impact, marginal inference impact, and the inference-to-lifecycle ratio. Mistral AI
- Most models ship with open weights, so independent labs (AI Energy Score, ML.ENERGY) can measure them directly. Hugging Face, Salesforce, Cohere, Carnegie Mellon
- Committed to publishing environmental impacts for future models and to contributing to an international standard. Mistral AI
- No dated net-zero, 24/7 carbon-free energy or water-replenishment target has been published. Mistral AI
- No environmental enforcement actions, permit disputes or community complaints on record. Mistral AI
Models
What this company ships, and what each prompt costs
| Model | Weights | Energy / prompt | Water | Training CO₂e |
|---|---|---|---|---|
| Mistral Large 2 (Le Chat, 400-token response) 123B dense | Open | 0.60 Wh Estimateecogpt.co estimate. Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement. | 45.0 mL DisclosedCompany figure. Published by the company itself. We take it as given and record the scope next to it. | 20,400 t DisclosedCompany figure. Published by the company itself. We take it as given and record the scope next to it. |
Where the compute lives
Named sites from company disclosures and press reporting
- Mistral Compute (Essonne)Bruyères-le-Châtel, France · 13,800 GB300 GPUs, 44 MW; hosted by Eclairion.
- Paris region hostingÎle-de-France, France
- Nordic hostingSweden
Timeline
Incidents and developments involving this company
Mistral finances 13,800 GB300 GPUs and 44 MW for its own data centre south of Paris
Mistral raised $830 million in debt to buy 13,800 NVIDIA GB300 GPUs for a 44 MW site at Bruyères-le-Châtel run by Eclairion — a revised version of the 18,000-GPU, 40 MW plan announced with NVIDIA in June 2025. The company targets 200 MW across Europe by end-2027.
MistralMistral publishes the first peer-reviewed LCA of a frontier model
Mistral Large 2: 20.4 ktCO₂e, 281,000 m³ of water and 660 kg Sb-eq over 18 months; 1.14 gCO₂e and 45 mL per 400-token response. Conducted with Carbone 4 and ADEME.
MistralAI Energy Score launches star ratings for 166 models across 10 tasks
Hugging Face, Salesforce, Cohere and Carnegie Mellon publish an Energy Star-style leaderboard: GPU watt-hours per task, with five relative star bands. Open models are tested automatically; closed models through a sealed sandbox.
Salesforce, Hugging Face, Cohere, Meta, Mistral, Alibaba, DeepSeek, IBM