Research details the environmental footprint associated with Large Language Models
In an era where countries are setting more ambitious climate targets and the AI industry is striving to address its environmental impact, Mistral AI has published one of the most comprehensive environmental impact assessments of an AI model to date. The report, which focuses on the company's Mistral Large 2 model, provides a clear picture of the environmental costs associated with training and inference in large language models.
According to the study, conducted with expert sustainability partners Carbone 4 and the French ecological transition agency (ADEME), and peer-reviewed by specialist consultancies in digital environmental audits Resilio and Hubblo, the model's total greenhouse gas emissions and water use are significant. Over 85.5% of the model's emissions and 91% of its water use stem from the compute processes that power both training and inference.
The 18-month lifecycle analysis (LCA) reveals that the training process, which took place up until January 2025, had a carbon footprint of 20.4 kilotonnes CO₂ equivalent (ktCO₂e), consumed 281,000 cubic meters of water, and resulted in 660 kilograms of antimony-equivalent resource depletion. On the other hand, each user interaction using their AI assistant Le Chat, excluding user devices, results in a carbon footprint of 1.14 grams CO₂e, uses 45 milliliters of water, and depletes 0.16 milligrams of rare earth resources.
The report also sheds light on factors influencing the environmental impact, such as geographic location and energy source, model size, and efficient inference techniques. Training in cool climates with access to renewable, carbon-free energy minimizes environmental costs, while choosing appropriately sized models tailored to specific use cases and employing efficient inference techniques are vital to reducing footprint.
Mistral's recommendations for reducing AI's impact include selecting smaller models, batching requests, choosing data centers powered by renewable energy, and including environmental factors in vendor selection. The future of AI could be one that is not just powerful and safe, but sustainable as well, if the field adopts these recommendations and best practices.
As enterprises and governments increasingly adopt technologies that align with sustainability goals, Mistral's report may set a benchmark for AI systems' environmental footprint to become a key performance indicator. The industry needs to move towards greater transparency and common environmental reporting standards to address the environmental impact of AI effectively.
In conclusion, the study by Mistral AI is a significant step towards quantifying and understanding the environmental costs of AI, and it calls for transparency, best practices, and collective accountability in the industry's approach to sustainability. The cost of each token in AI should matter not just for technological advancement, but also for the planet's sake.
The report published by Mistral AI, in collaboration with sustainability partners, has revealed that a significant portion of the emissions and water usage in their large language model, the Mistral Large 2, originates from the compute processes powering both training and inference. Moreover, the 18-month lifecycle analysis (LCA) showed that these activities contribute to substantial greenhouse gas emissions (20.4 ktCO₂e), water consumption (281,000 cubic meters), and resource depletion (660 kilograms of antimony-equivalent). As AI systems become increasingly influential, the environmental impact of technology, such as AI, should be considered a key performance indicator to drive greater transparency and the adoption of common environmental reporting standards within the industry.
