Our contribution to a global environmental standard for AI

Mistral AI · 11 months ago

Mistral AI conducted the first comprehensive lifecycle analysis of an AI model, measuring environmental impacts across greenhouse gas emissions, water consumption, and resource depletion in collaboration with external auditors. Training Mistral Large 2 generated 20.4 kilotonnes of CO₂ equivalent, 281,000 cubic metres of water, and 660 kilograms of antimony equivalents over 18 months, while a single 400-token inference response consumed 1.14 grams of CO₂ equivalent and 45 millilitres of water. The company proposes that AI developers publish standardized environmental metrics and that procurement policies incorporate model efficiency criteria to enable informed purchasing decisions and reduce sector-wide environmental impact.