French AI startup Mistral is pulling back the curtain on what your chatbot sessions are actually costing the planet. Teaming up with sustainability consultants and backed by France’s official environmental agency, they published a rare full environmental audit of their “Large 2” language model. Turns out, most of the AI’s carbon footprint—86% of it—comes from training the models and running prompts, not from the actual buildings full of humming servers or the laptops we query them with. Same deal with water: 91% of it is drained during training and inference, not day-to-day device use.
Here’s the bite-size impact: generate a typical AI response and you’ve just emitted 1.14 grams of CO2 and evaporated 45 milliliters of water—about enough to make your houseplant blink. But stack up all the queries over Mistral’s first 18 months, and the emissions soared to 20,400 tons of CO2 (roughly the annual output of 4,500 gas-powered cars) and 281,000 cubic meters of water—enough to fill 112 Olympic pools. These aren’t just abstract numbers; they put an actual cost on convenience at scale.
To make it relatable, Mistral did some math: One prompt equals the carbon of watching 10 seconds of a streaming show in the US or sitting through a few blinks of a Zoom call. That email you spent 10 minutes crafting for your team of 100? It cost the same CO2 as nearly 23 AI prompts. Marketers and brands leaning into generative AI might want to consider not just ROI but ECOI—environmental cost of inference.
Full story at Ars Technica.