Our contribution to a global environmental standard for AI

Mistral AI just did something no major AI lab has bothered to do properly: it measured, end to end, what one of its models actually costs the planet. Working with French consultancy Carbone 4 and the government's ADEME agency, plus outside audits from Resilio and Hubblo, Mistral tracked the full lifecycle of Large 2 — training, hardware manufacturing, and inference — and put real numbers on it.

The figures are specific enough to argue with. Training Large 2 through January 2025, after 18 months of use, generated 20,400 tonnes of CO2 equivalent, consumed 281,000 cubic meters of water, and depleted 660 kg of antimony-equivalent resources. A single 400-token response from Le Chat costs about 1.14 grams of CO2, 45 milliliters of water, and 0.16 milligrams of resource depletion — tiny per query, but multiply that by billions of chats and it stops looking trivial.

What's more useful than the raw totals is what Mistral learned about how to think about them. The company argues the industry needs three standard metrics: total training impact, marginal inference impact, and the ratio between total inference and full lifecycle cost — that last one tells you whether a model's training footprint actually gets amortized through heavy use or just sits there as wasted carbon. Mistral also found a roughly linear relationship between model size and impact: a model ten times larger produces impacts an order of magnitude bigger per generated token, which is a blunt argument for not defaulting to the biggest model available for every task.

The company is upfront that this is a first approximation, not gospel. There's no reliable public lifecycle inventory for GPU manufacturing yet, so Mistral had to estimate embodied hardware impacts, and those turned out to be a substantial chunk of the total. No agreed standard exists for AI environmental accounting either, though Mistral says it followed the GHG Protocol Product Standard and ISO 14040/44, plus a French AFNOR methodology called Frugal AI.

Mistral wants this to become a template: standardized public disclosure, a possible scoring system so buyers can compare models the way they might compare appliances, and procurement rules that reward efficient model choices, especially for governments buying AI at scale. The results will eventually land in ADEME's public Base Empreinte database. Whether competitors follow with their own numbers, rather than vague sustainability pledges, is the real test of whether this becomes a norm or stays a one-off PR move dressed up as science.