Mistral's recent environmental audit quantifies the impacts of their Large 2 model, finding that 85.5% of CO2 emissions and 91% of water use occur during model training and inference. The audit reveals that generating a single average prompt emits 1.14 grams of CO2 and consumes 45 milliliters of water. While individual impacts are relatively small, the aggregate effects of billions of AI prompts contribute to substantial environmental consequences. Collaborating with Carbone 4 and following Frugal AI guidelines, Mistral aimed to provide clarity on the sustainability of AI models.
The environmental audit by Mistral reveals that the majority of CO2 emissions and water consumption arise during model training and inference, not from construction or end-user equipment.
Mistral's assessment shows that a single AI prompt generates 1.14 grams of CO2 and consumes 45 milliliters of water, indicating individual impacts are minimal.
Despite individual AI query impacts being low, the sheer volume of AI usage leads to significant cumulative environmental effects.
Collaborating with Carbone 4, Mistral conducted an environmental audit adhering to French guidelines, focusing on CO2 emissions, water use, and materials consumption.
Collection
[
|
...
]