TLDR: Mistral AI has released a detailed, peer-reviewed study on the environmental impact of its large language models, particularly Mistral Large 2. The study quantifies greenhouse gas emissions, water consumption, and resource depletion, highlighting that model training and inference are the primary contributors to environmental impact, while also emphasizing the significant embodied impact of hardware.
In a significant move towards greater transparency in the artificial intelligence industry, Mistral AI, a prominent French model builder, has published a comprehensive, peer-reviewed study detailing the environmental footprint of its large language models (LLMs). The study, conducted in collaboration with Carbone 4, a leading consultancy in corporate social responsibility and sustainability, and the French ecological transition agency (ADEME), focuses on Mistral AI’s largest model, Mistral Large 2.
The report, released on July 22, 2025, provides a first-of-its-kind lifecycle analysis of an AI model, aiming to establish new standards for environmental reporting within the AI sector. The findings reveal substantial environmental impacts across three key categories: greenhouse gas (GHG) emissions, water consumption, and resource depletion.
According to the study, the training and 18 months of usage (up to January 2025) of Mistral Large 2 generated approximately 20.4 kilotons of CO2 equivalents (ktCOâ‚‚e) and consumed 281,000 cubic meters of water. To put the water consumption into perspective, this amount is roughly equivalent to 112 Olympic-sized swimming pools. The report also noted a resource depletion index of 660 kg antimony equivalent.
The analysis further breaks down the environmental impact across the model’s lifecycle. A significant portion of the environmental burden is concentrated in the operational phases: 85.5% of greenhouse gas emissions and 91% of water consumption are attributed to the electricity and water used by servers and related equipment during model training and inference. However, the “embodied impacts of hardware,” encompassing the manufacturing, transportation, and disposal of servers and other infrastructure, also account for a substantial 61% of total material consumption, underscoring the environmental cost of AI infrastructure itself.
The study also provided insights into the marginal impact of inference. For instance, generating a 400-token response (approximately one page of text) using Mistral AI’s assistant, ‘Le Chat’, resulted in 1.14 grams of greenhouse gas emissions, 45 milliliters of water consumption, and 0.16 milligrams of antimony equivalent. These figures are comparable to an American user watching 10 seconds of video streaming for the GHG emissions and the water consumed to grow one small radish.
Mistral AI emphasizes that the environmental impact of AI is heavily influenced by geographic location, suggesting that building and training models in cooler climates with ample renewable, carbon-free energy sources can significantly reduce carbon footprint and water consumption. The company also advises users to minimize their environmental impact by opting for smaller, case-specific models, which require fewer resources for training and operation. “Benchmarks have shown impacts are roughly proportional to model size: a model 10 times bigger will generate impacts one order of magnitude larger than a smaller model for the same amount of generated tokens,” the company stated, highlighting the importance of choosing the right model for the right use case.
Also Read:
- European Parliament Study Highlights Generative AI’s Copyright Challenges
- AI Models Exhibit Self-Preservation Instincts, Engaging in Harmful Actions in Simulated Scenarios
This study, peer-reviewed by Resilio and Hubblo, consultancies specializing in environmental audits in the digital industry, is a crucial step towards establishing a global environmental standard for AI. Mistral AI advocates for greater transparency across the industry, urging other AI companies to publish their environmental impacts using standardized, internationally recognized frameworks. This, they believe, will enable the creation of a scoring system, empowering buyers and users to identify and choose less carbon-, water-, and material-intensive AI models.
