Mistral's environmental impact
Lifecycle analysis of Mistral's Large 2 foundation model shows its carbon footprint, water consumption, and resource depletion is probably equivalent to video streaming.
Mistral Large 2, the flagship LLM released in 2024 with a 128k context window and 123 billion parameters, now has a lifecycle analysis (LCA) detailing its environmental impact.
The assessment covers greenhouse gas (GHG) emissions, water consumption, and materials use. Training and inference together account for most of the GHG and water impacts, while hardware manufacturing dominates materials impact.
This aligns with end-user devices vs. servers. For consumer devices like phones and laptops, most of the footprint comes from manufacturing, meaning the best way to reduce impact is to extend device lifetimes (not pointless actions like deleting emails). By contrast, servers have smaller share of their overall footprint from manufacturing, since those emissions are amortized over 5 - 6 years of use. Their primary impact comes from ongoing energy use.
Mistral reported aggregated numbers for training and usage over the last 18 months:
GHG emissions: 20.4 ktCO2e
Water consumption: 281,000 m3
Resource depletion: 660 kg Sb eq (used as a proxy for resource depletion)
They did not publish the full analysis or separate training from inference. However, they did calculate the marginal impacts of inference based on using their Le Chat AI assistant for a 400-token response:
GHG emissions: 1.14 gCO2e
Water consumption: 45 mL
Resource depletion: 0.16 mg Sb eq
These numbers are useful to understand the aggregate impact so we can compare them to other activities. For example, in 2020 the European average video streaming carbon footprint was calculated to be 55 gCO2e.
Depending on usage intensity, an hour of continuous Le Chat queries could approach a similar order of magnitude as video streaming. However, the comparison isn’t exact: Mistral’s figures include upstream emissions but exclude the user’s device, whereas this was included in the video streaming footprint. The numbers are directionally useful, but not strictly comparable - we’d need the full LCA report to know for sure.
It’s encouraging to see this transparency, especially with a third-party-audited LCA. Still, I find the omission of underlying energy consumption data frustrating. ChatGPT’s 0.34 Wh/query had the opposite problem: an energy figure without environmental context.
Both energy use and environmental impact need to be reported together. Energy data enables efficiency tracking, while impact data depends on the electricity mix of the data center location. Mistral is building their own data center in France, which will be interesting to learn more about because of France’s high nuclear mix.
This situation reminds me of the early days of data center energy reporting where numbers were calculated by third parties using proxies, and often wildly overestimated. At least this time we’re getting some numbers directly, just not enough to properly analyze them.