Mistral’s new “environmental audit” shows how much AI is hurting the planet

In an effort to address the growing concerns about the environmental impact of artificial intelligence, Mistral, a French AI development firm, has released what it describes as a pioneering environmental audit. This initiative aims to provide clear, quantifiable data on the CO2 emissions and water usage associated with its large language models (LLMs). The findings from this audit align closely with previous academic estimates, indicating that while the environmental impact of a single AI query may appear minor in isolation, the cumulative effect of billions of queries each year can result in substantial environmental consequences. To conduct this lifecycle analysis of its 'Large 2' model, which has been operational for nearly 18 months, Mistral collaborated with sustainability experts at Carbone 4 and the French Agency for Ecological Transition. Following the French government's Frugal AI guidelines, the audit assessed three primary categories: greenhouse gas emissions, water consumption, and material usage, primarily focusing on the depletion of non-renewable resources stemming from wear and tear on AI server GPUs. The results revealed that an overwhelming majority of CO2 emissions (85.5%) and water consumption (91%) occurred during the model’s training and inference phases, rather than during data center construction or the energy used by end-user devices. Specifically, the audit found that the environmental cost of generating an average prompt—equivalent to about a page of text—resulted in emissions of just 1.14 grams of CO2 and a water consumption of 45 milliliters. However, when looking at the broader picture, Mistral identified a significant total impact over its first 18 months: approximately 20.4 kilotons of CO2 emissions, comparable to the output of 4,500 internal combustion engine vehicles operating for a full year, as reported by the Environmental Protection Agency. Additionally, the process resulted in the evaporation of 281,000 cubic meters of water, which is enough to fill 112 Olympic-sized swimming pools. To contextualize these findings, Mistral compared its environmental impact to other common online activities. For example, the CO2 emissions from a single average LLM query are roughly equivalent to watching 10 seconds of a streaming video in the United States or 55 seconds in France, where the energy sources are cleaner. Furthermore, engaging in a Zoom call for just four to 27 seconds emits a similar amount of CO2, based on data from the Mozilla Foundation, and spending 10 minutes composing an email that is read by one of its 100 recipients generates as much CO2 as 22.8 Mistral prompts, according to Carbon Literacy data.