Mistral publishes lifecycle footprint of Large 2

Mistral AI published what it calls the first comprehensive lifecycle analysis (LCA) of an AI model. It carried out the study with sustainability consultancy Carbone 4 and France's ecological transition agency, ADEME. Resilio and Hubblo, two consultancies that specialise in digital environmental audits, peer-reviewed the work.

The study measured greenhouse gas emissions, water use and resource depletion. As of January 2025, after 18 months of use, Mistral Large 2 had generated 20.4 ktCO₂e, consumed 281,000 m³ of water and accounted for 660 kg Sb eq of resource depletion. The marginal cost of one 400-token Le Chat response, excluding users' devices, was 1.14 gCO₂e, 45 mL of water and 0.16 mg Sb eq. The figures include upstream impacts such as server manufacturing, not just energy use.

Mistral found that impacts scale roughly in proportion to model size. For the same number of generated tokens, a model 10 times larger has about an order of magnitude more impact.

The company proposes that developers be required to report absolute training impacts and marginal inference impacts. A third metric, the ratio of inference to total lifecycle impacts, could be disclosed optionally. Mistral also calls for standardised frameworks that would support a scoring system for buyers, and for public procurement to take model size into account.

Mistral describes the results as a first approximation. No LLM accounting standard exists, and GPU embodied impacts, a significant share of the total, had to be estimated. The company says it will update its reports.