Mistral's new "environmental audit" shows how much AI is hurting the planet
9 months ago
- #AI
- #Environment
- #Sustainability
- Mistral conducted a first-of-its-kind environmental audit to quantify the impacts of its Large 2 AI model.
- The audit measured CO2 emissions, water consumption, and materials depletion, finding most impact occurs during model training and inference.
- A single AI prompt (400 tokens) emits 1.14g CO2 and consumes 45ml water—minimal individually but significant in aggregate.
- Over 18 months, Mistral's model emitted 20.4 ktons CO2 (≈4,500 cars/year) and used 281,000m³ water (≈112 Olympic pools).
- AI queries' CO2 impact is comparable to 10 sec of US streaming (55 sec in France) or 4-27 sec on Zoom.
- Writing a 10-minute email read by 100 recipients emits as much CO2 as 22.8 Mistral prompts.