Mistral AI has put a detailed environmental accounting frame around one of its major AI systems, publishing what it calls the first comprehensive life cycle assessment of a large language model. The report centers on Mistral Large 2 and examines both training and 18 months of operation.
The findings matter because generative AI is often discussed in terms of capability, cost and speed, while its physical footprint is harder to see. Mistral’s report makes that footprint more concrete by attaching numbers to carbon emissions, water consumption and the rare metals and minerals tied to hardware production.
What Mistral measured
The assessment looks at Mistral Large 2 across a broad life cycle, rather than treating the model as a purely digital service. By January, the model had generated 20.4 kilotons of CO₂ equivalents, consumed 281,000 cubic meters of water, and used resources equivalent to 660 kilograms of antimony.
Antimony equivalents are used to represent the consumption of rare metals and minerals required for hardware production. That matters because a large language model depends not only on electricity and data center operations, but also on the equipment that makes training and inference possible.
Mistral also translated the broader footprint into a single-use example. For one 400-token response from Mistral’s Le Chat assistant, the company calculated an impact of 1.14 grams of CO₂ equivalents, 45 milliliters of water, and 0.16 milligrams of antimony equivalents.
Those per-request figures help make the issue easier to understand. A single prompt may appear small, but the environmental significance grows when AI systems are used repeatedly and at scale. The report does not frame this as a reason to stop using generative AI. Instead, it points toward measuring usage more clearly and choosing tools more carefully.
Why comparison is still difficult
The report arrives in an AI market where environmental claims are not yet easy to compare. OpenAI CEO Sam Altman recently claimed that an average ChatGPT request uses only 0.32 milliliters of water. That figure is less than one-hundredth of Mistral’s water estimate for a 400-token response.
But the source article makes clear that the numbers should not be treated as a simple ranking. OpenAI has not published a transparent breakdown of its environmental impact, and Altman only mentioned the statistic in passing. Without a comparable method, scope and calculation model, the gap between the two figures remains hard to interpret.
This is the central transparency problem the Mistral report tries to address. If companies use different assumptions, disclose different categories or omit key details, users cannot reliably compare AI environmental impact across providers. A headline number may be technically accurate within one method while still being difficult to compare with another company’s figure.
Model size changes the footprint
Mistral’s study identifies a clear relationship between model size and environmental cost. The larger the model, the larger the footprint. According to the report, a model that is ten times bigger has an environmental cost roughly an order of magnitude higher, assuming the same number of generated tokens.
That has practical consequences for product teams, companies and public institutions using AI systems. If a smaller model can handle a task, using a larger model may add unnecessary environmental cost. The source article does not reduce this to a single rule for every use case, but it does make the selection of the right model an environmental decision as well as a technical one.
Mistral also points to the infrastructure behind generative AI. The massive computational requirements often run on GPU clusters in regions with carbon-heavy electricity and sometimes under water stress. The environmental impact of AI is therefore tied not only to model architecture, but also to where and how the computation happens.
The reporting metrics Mistral wants
Based on the assessment, Mistral proposes three metrics for industry reporting. These are meant to give users, companies and institutions a clearer way to evaluate the environmental cost of large language models.
- Total impact of model training: the footprint created before the model is deployed for use.
- Per-inference or per-request impact: the environmental cost of using the model for an individual request.
- Ratio of inference to overall life cycle impact: a broader view of how usage compares with the total model life cycle.
Mistral argues that the first two data points should be mandatory so the public can better understand AI’s environmental effects. The third could be used internally, with optional public disclosure, to give a more complete life cycle view.
The company also outlines two broad ways to reduce the footprint of AI. First, AI companies should publish environmental impact data using internationally recognized standards, making it easier to compare models and select greener options. Second, users can improve efficiency by choosing the right model for their needs, bundling requests and avoiding unnecessary computations.
Public institutions could also influence the market. Mistral suggests they could consider model efficiency and size in procurement decisions, which would make environmental performance part of how AI systems are selected.
Limits of the first assessment
Mistral does not present the report as a final answer. The company says the initial analysis is only a rough estimate. Accurate calculations are difficult because there are no established standards for LLM life cycle assessments and no widely available assessment factors.
Another gap is hardware data. The source article notes that there is still no reliable life cycle data for GPUs, which are central to large-scale AI systems. That makes a complete accounting harder, especially when rare metals, minerals and hardware production are included.
Mistral plans to update its environmental reports and says it wants to help shape international industry standards. The results are expected to be published in the Base Empreinte database, a French reference platform for environmental impact data on products and services.
Regulation is also moving in this direction. In the EU, the new AI Act already requires providers of general-purpose AI models to document energy usage in detail. Developers must produce technical documentation that breaks down model energy consumption, and energy use is one factor regulators consider when determining whether a model poses a "systemic risk."
The broader message is straightforward: AI’s environmental impact cannot be managed well if it is not measured consistently. Mistral’s life cycle assessment gives the industry a concrete example of what more transparent reporting could look like, while also showing how much work remains before AI footprint comparisons become routine and reliable.