The footprint arises in 2 phases: training, which happens once, and use (inference, meaning running a trained model to answer a query), which recurs with every query. The LCA published by Mistral puts Mistral Large 2 at 20,400 tCO2e, including training and the first 18 months of use. A short text query uses about 0.3 Wh, the same order of magnitude as a Google search: Google measured 0.24 Wh for a median Gemini query. Worldwide, data centres accounted for about 1.5% of electricity consumption in 2024, and the IEA projects a share of around 3% in 2030.
Read the source article: ChatGPT carbon footprint: what it means for your company