Why Mistral ranks first, and what it still does not say

The case for first place

Mistral's score of 75 rests on two pillars. Transparency, at 88, is the highest in the index: in July 2025 the company published the first full life-cycle assessment of a large language model, covering training, inference, hardware manufacturing and data-centre construction, conducted with Carbone 4 and ADEME and peer-reviewed by Resilio and Hubblo under France's AFNOR Frugal AI methodology. Carbon, at 85, follows from geography: Mistral trains and serves mostly on the French grid, nuclear- and hydro-dominated, among the lowest-carbon electricity any company in the index uses.

What the assessment says

Mistral Large 2, a 123-billion-parameter dense model, emitted 20.4 ktCO₂e over its first 18 months, 85.5% of it from training, consumed 281,000 m³ of water and depleted 660 kg of antimony-equivalent in minerals. A 400-token Le Chat response costs 1.14 gCO₂e, 45 mL of water and 0.16 mg Sb-eq at the margin, counting the water used to generate the electricity, which is why the water figure is 170 times Google's on-site 0.26 mL. Mistral proposed three indicators every model should report and committed to publishing them for future models.

What it does not say

Mistral publishes no company-wide greenhouse-gas inventory, no fleet water-use effectiveness and no energy figure per response (the index estimates 0.6 Wh). Its hosting in France and Sweden was leased until 2026, when it financed 13,800 GB300 GPUs and 44 MW of its own at Bruyères-le-Châtel, with 200 MW across Europe targeted by end-2027. It has set no net-zero, 24/7 carbon-free or water-replenishment date. The commitments pillar scores 55 for that reason, and the emissions-trend figure on the profile is an estimate: compute roughly doubling as the new cluster comes online.

The uncomfortable comparison

Mistral is first partly because it is small. Its whole life-cycle figure for its flagship model, 20.4 kt, is a rounding error against Google's 11.5 Mt corporate total, and a 44 MW cluster is about 4% of a Stargate. The index scores practice, not scale; it says so in the methodology, and the hardware block on each profile shows the scale separately. A reader who wants to know who emits most should sort the rankings by annual emissions, where Mistral is near the bottom.

What would keep it first

Publishing the three indicators for the next model family, a fleet WUE for the new cluster, and a dated target. A single assessment set the standard. The index will score the second one on whether the standard holds.