Mistral publishes the first peer-reviewed LCA of a frontier model

Mistral's life-cycle assessment follows the AFNOR Frugal AI methodology and covers data-centre construction, server manufacturing and transport, training, inference and user devices. Resilio and Hubblo peer-reviewed it.

Training dominates: about 85.5% of greenhouse-gas emissions and 91% of water. The marginal inference impact of a 400-token Le Chat response is 1.14 gCO₂e, 45 mL of water and 0.16 mg Sb-eq — figures that include water consumed to generate electricity, which is why they are far above Google's on-site-only 0.26 mL.

Mistral proposes three indicators every provider should report: absolute training impact, marginal inference impact, and the ratio of total inference to total life-cycle impact. The index adopts the same distinction between model-level and facility-level data.