Our Contribution to a Global Environmental Standard for AI

In this company article, Mistral AI focuses on the environmental transparency and industry standardization of AI models. The article points out that as AI is integrated into every layer of the economy and public services, developers, policymakers, enterprises, governments, and ordinary users all need a clearer understanding of generative AI's environmental footprint; otherwise it is hard for outsiders to compare different models, make procurement decisions, fulfill corporate non-financial disclosure obligations, or reduce the environmental impact of AI use. Mistral AI says it collaborated with Carbone 4 and France's ecological transition agency ADEME, with peer review by Resilio and Hubblo, to complete a full lifecycle analysis of AI models. This analysis covers three categories of environmental indicators: greenhouse gas emissions, water use, and resource depletion, and follows standards such as the Frugal AI methodology, the GHG Protocol Product Standard, and ISO 14040/44. The article discloses two sets of concrete figures: as of January 2025, Mistral Large 2, over its training plus 18 months of usage, is estimated to have produced 20.4 ktCO2e, consumed 281,000 cubic meters of water, and caused a resource depletion impact of 660 kg Sb eq; while the marginal inference impact of Le Chat generating a 400-token response is about 1.14 gCO2e, 45 mL of water, and 0.16 mg Sb eq, excluding the user's end device. Mistral AI emphasizes that these figures reflect not only GPU electricity use but also upstream impacts such as server manufacturing. The article also proposes three key indicators that should be tracked: the absolute impact of model training, the marginal impact of inference, and the share of total inference volume in the full lifecycle impact. Mistral AI believes the first two can become required items for public disclosure, while the third can serve as an indicator for internal management and selective disclosure. Its policy proposal is to establish an internationally recognizable, comparable framework for AI environmental disclosure, and to encourage users to choose appropriately sized models based on actual needs, improve AI literacy, and consolidate unnecessary queries, so that public institutions can also incorporate model size and efficiency into procurement criteria.