Voxtral creator leads call for environmental transparency in AI

Mistral AI, known for its open-source approach to generative models, has published the first full lifecycle analysis (LCA) of a large language model (LLM). In partnership with Carbone 4, ADEME, Resilio, and Hubblo, the study offers a detailed account of the environmental impact of developing and deploying its model, Mistral Large 2.

This move comes amid rising global interest in setting sustainability benchmarks for artificial intelligence. While initiatives like the Coalition for Sustainable AI signal progress, Mistral’s report pushes the conversation further by quantifying environmental impacts across three categories: greenhouse gas emissions, water consumption, and resource depletion.

Groundbreaking Metrics for Training and Inference

The study reveals that training Mistral Large 2 generated:

• 20.4 kilotons of CO₂ equivalent (ktCO₂e)

• 281,000 cubic meters of water

• 660 kg of antimony equivalent (Sb eq)

In addition to these total training impacts, Mistral also disclosed the marginal costs of using its assistant Le Chat to generate a typical 400-token response:

• 1.14 grams of CO₂

• 45 mL of water

• 0.16 milligrams of Sb eq

These figures, which include upstream impacts like hardware manufacturing, go beyond typical carbon-only disclosures and help contextualize the true cost of running AI at scale.

Towards a Standard for AI Accountability

According to Mistral, three key indicators should be considered industry standard:

1. Absolute training impact

2. Marginal inference impact

3. Ratio of total inference to lifecycle impact

The company emphasizes that a model’s environmental footprint is strongly correlated with its size—raising the stakes for selecting the right model for the right task. Public institutions, the report suggests, could lead the way by incorporating model size and efficiency into their procurement criteria.

A Call for Transparency

Mistral’s LCA aligns with international standards such as ISO 14040/44 and the GHG Protocol Product Standard, and follows the Frugal AI methodology developed by AFNOR. While the company acknowledges current limitations—such as the lack of full lifecycle inventories for GPUs—it sees this as a starting point for building standardized, comparable benchmarks across the industry.

The ultimate goal: to enable governments, enterprises, and users to make informed, sustainability-conscious decisions when adopting AI tools.

What Comes Next

Mistral plans to update its environmental impact reports regularly and contribute its findings to ADEME’s Base Empreinte database. It also advocates for creating internationally recognized frameworks for model comparison, potentially leading to a scoring system for environmental performance in AI.

By making this report public, Mistral positions itself at the forefront of open-source sustainability in AI—setting a precedent for environmental responsibility in one of the most rapidly growing sectors of the tech industry.