Towards a more transparent and environmentally responsible generative AI

For some time now, assessing the environmental impact of generative AI models has been gaining importance. Mistral AI has taken a major step forward by publishing a comprehensive life cycle assessment (LCA) of its flagship model, Large 2, conducted with specialized partners and validated by independent experts. This initiative represents a world first: a major player in the sector is making public a complete set of tangible data on CO₂ emissions, water consumption, and resource depletion linked to an operational LLM.

The Large 2 model, after approximately eighteen months of operation through January 2025, generated approximately 20,4 kilotons of CO₂ equivalent, consumed 281 m³ of water, and used non-renewable resources equivalent to 000 kg of antimony (a standardized unit used to compare the consumption of rare metals). For the inference phase, each 660-token response via the Le Chat assistant represents approximately 400 g of CO₂, 1,14 mL of water, and 45 mg of antimony. These figures offer rare granularity—far beyond anything previously published in the industry.

The analysis reveals that the bulk of the impact comes from the training and inference phase (85,5% of greenhouse gas emissions and 91% of water consumption), while manufacturing, transportation, and infrastructure end-of-life contribute approximately 61% of the depletion of material resources. This distribution highlights the dual responsibility: optimizing hardware while making daily use more efficient.

The methodology used is based on the Frugal AI method developed by AFNOR, which complies with international standards such as the GHG Protocol Product Standard and ISO 14040/44. The measured indicators—GWP100 for CO₂, WCP for water consumption, and ADP for abiotic resource depletion—are consistent with life cycle analysis practices. The ADP model, expressed in antimony equivalents, allows for consistent comparisons between materials with very different reserves (e.g., copper vs. gold).

Drawing on these lessons, Mistral AI advocates for the adoption of a standard environmental reporting framework within the industry: three key metrics are recommended: the total impact of model training, the impact per query (inference), and the ratio between inference and full lifecycle. All of this must be communicated transparently to enable comparability and informed decision-making, particularly for institutional buyers and public authorities.

On the user side, Mistral encourages a lean and efficient AI practice: training users to better query models, group queries, and choose the model size appropriate for the task, rather than systematically using the largest models for each use. Integrating environmental criteria into public tenders, particularly energy consumption and model size, can send a strong signal to the market.

This communication goes beyond a simple effort at transparency: it constitutes a call to the industry to create a global standard. Mistral is committed to regularly updating its environmental reports and participating in international discussions on the development of sector standards. The data will be made available via a public database, helping to make this framework a benchmark for transparency in the field of AI.

This initiative has attracted both praise and criticism. Some note the lack of an explicit separation between training and inference in the figures presented, which is often a key point in discussions about the models' actual consumption. Others believe that the units of measurement (20,4 ktCO₂) could be made more understandable if they were put into context relative to other human or industrial activities.

Despite these limitations, the approach adopted marks a significant step forward: a comprehensive, transparent, third-party audited, and publicly published model, making it a benchmark for an entire industry grappling with the imperative of sustainability. It concretely illustrates that each AI query, however small in individual impact, can accumulate tons of CO₂ on a large scale—calling for responsible practices in terms of volume of use.

In conclusion, this work sheds new light on our understanding of large-scale text generation models. It provides a solid methodological framework for assessing their impacts on several environmental dimensions. It calls on all stakeholders—developers, institutional users, regulators—to adopt a responsible AI strategy, where transparency, efficiency, and sufficiency are the key words. For AI consultants, trainers, and researchers, this study is a roadmap: acting today on the models deployed and tomorrow on the rules and standards of the global AI ecosystem.

To go further: this environmental transparency is extended with our article on live electricity et our analysis of energy as a competitive advantage.