Mistral AI sustainability score: 35/100

No formal corporate GHG inventory (Scope 1/2) has been published to GHG Protocol or equivalent standard. Mistral's July 2025 lifecycle assessment (LCA) of Mistral Large 2, conducted with Carbone 4 and ADEME and peer-reviewed by Resilio and Hubblo, discloses 20.4 ktCO₂e across training and 18 months of inference for one model — covering data center energy, server manufacturing, and upstream electricity supply chain. This is a model-level LCA, not a corporate emissions inventory. No BEGES filing was found on ADEME's registry (Mistral was below the 500-employee threshold for most of its history). Estimated corporate total emissions: likely in the 50,000–150,000 tCO₂e/year range given growing compute capacity (under 200 MW operational as of mid-2026), placing this near the boundary of the 100K-1M tonne scale band. Score of 4 reflects meaningful model-level data estimated from operational activity, peer-reviewed by specialists, but absent a corporate Scope 1+2 inventory.

The July 2025 LCA, aligned with ISO 14040/44 and the Frugal AI methodology (AFNOR), covers upstream hardware manufacturing (servers, data center construction), electricity supply chain (including extraction of fuels for power plants, manufacturing of wind turbines and solar panels), training compute, and inference — effectively mapping the most material Scope 3 categories for an AI model company. Hardware manufacturing accounts for 61% of material resource depletion and 11% of GHG emissions. However, this is a product-level LCA for one model, not a comprehensive corporate Scope 3 inventory across all categories (business travel, offices, other models). No GHG Protocol-aligned Scope 3 corporate disclosure exists. Score of 5 reflects coverage of the most material supply chain emission sources for the core product, peer-reviewed, but falling short of a full corporate value chain disclosure.

No multi-year corporate emissions trajectory exists — the July 2025 LCA is the first and only substantive environmental disclosure. Mistral is in a phase of rapid infrastructure expansion: it operates less than 200 MW of capacity as of mid-2026, is targeting 1 GW by 2030, secured €830M in debt to buy 13,800 Nvidia chips for a new Paris data center, and announced a €1.2B partnership with EcoDataCenter in Sweden. Absolute emissions are almost certainly rising steeply as new data centers come online and more models are trained. There is no evidence of absolute reductions, no baseline year, and no year-on-year data. Score of 2 reflects a small-but-rapidly-scaling company with no trajectory evidence and strong structural indicators of rising absolute emissions.

Mistral's new own-operated data centers are being located in low-carbon regions: a 44MW facility near Paris benefits from France's predominantly nuclear electricity grid, a 23MW facility in Sweden is operated by EcoDataCenter using renewable energy and advanced cooling, and a 10MW site in Les Ulis, France. Mistral also states it is building its own data center in France to exploit low-carbon nuclear power. However, historically training was performed on third-party GPU clouds in regions with carbon-intensive electricity, as Mistral's own LCA acknowledges. No RE100 membership, no published renewable energy percentage for the overall compute portfolio, no PPAs documented. Score of 4 reflects deliberate geographic siting for low-carbon energy in own facilities, not yet a documented 50%+ renewable commitment or verified renewable tariff across all compute.

No TNFD disclosure, no formal biodiversity policy, and no nature risk assessment was found. Direct land-use and ecosystem exposure for an AI software company is low — Mistral owns no mines, agricultural land, or forestry operations. The LCA quantifies abiotic resource depletion (660 kg Sb eq for hardware materials over 18 months), providing some awareness of mineral extraction impacts embedded in supply chain. No deforestation commitments or habitat assessments documented. Score of 3 reflects minimal direct nature exposure with some proxy awareness via resource depletion metrics in the LCA, but absence of any formal nature policy or TNFD-aligned work.

The LCA covers material resource consumption including server manufacturing, GPU production, and end-of-life hardware — quantified at 660 kg Sb eq for one model over 18 months, with hardware manufacturing and transport accounting for 61% of material consumption. Mistral acknowledges the absence of reliable GPU lifecycle data as a limitation and commits to improving this. No formal circular economy policy, e-waste management programme, or supplier sustainability requirements documented publicly. The LCA's commitment to contribute findings to ADEME's Base Empreinte database is a form of public reporting. Score of 4 reflects quantified resource depletion in the product LCA and acknowledged supply chain material risks, but no formal circular or waste reduction programme.

Mistral's LCA discloses 281,000 m³ of water consumed across training and 18 months of inference for Mistral Large 2 — 91% from training and inference phases. Per-query water use of 45 mL per 400-token Le Chat response is quantified. Mistral explicitly notes that geographic location affects water consumption and is building new data centers in cooler climates (Sweden, northern France) to reduce evaporative cooling demand. No formal water reduction targets or water stewardship programme documented. No water pollution incidents found. Score of 4 reflects quantified water footprint for core product and strategic siting to reduce future water stress, but absence of formal reduction targets or independent verification of water data.

No SBTi commitment found on the SBTi target dashboard. No absolute emissions reduction targets with specific baselines, timelines, or percentages. No net-zero pathway published. Mistral has stated a commitment to update environmental impact reports regularly and participate in development of international industry standards — this is a transparency and process commitment, not an emissions reduction target. No offsets are claimed (appropriate given no formal inventory). The decision to locate own data centers in low-carbon regions (France, Sweden) represents a design-level mitigation approach, but without quantified targets this cannot score above 2. Score of 2 reflects informal goals around transparency without specific, measurable reduction commitments.

Mistral's July 2025 LCA — the first for any LLM, conducted with Carbone 4 and ADEME, peer-reviewed by Resilio and Hubblo, aligned with ISO 14040/44 and the Frugal AI (AFNOR) methodology — is a sector-leading act of transparency that goes beyond what any comparable AI company has published. Mistral also published a public sustainability tracker tool. However, there is no formal sustainability governance structure documented: no dedicated sustainability team found, no board sustainability oversight, no executive compensation linked to environmental targets, no TCFD or CSRD disclosure, and no annual corporate sustainability report covering Scope 1/2/3. Score of 5 reflects world-leading product-level transparency in the AI sector, partially offset by absence of corporate governance structures expected at Mistral's scale and growth trajectory.

No environmental regulatory actions, greenwashing rulings, or fines found. No upheld complaints from advertising standards bodies or environmental regulators. InfluenceMap does not profile Mistral; no evidence of lobbying against climate policy. The controversy over Mistral's lobbying related to the EU AI Act pertains to open-source AI regulation, not climate or environmental policy — no deduction applies. Mistral does not operate in a fossil-sector-enabling structural role (it is not a payment processor, bank, insurer, or advertising holding company). No biodiversity incidents or water pollution found. Starting from 8, no deductions are warranted. Score of 8.