AI models are getting more efficient

8 August 2025

Mistral published Our contribution to a global environmental standard for AI, a “first-of-its-kind comprehensive study to quantify the environmental impacts of our LLMs”.

Mistral AI Infographie Acv V6(1).

What’s interesting:

• looking at the diagram, most of the the emissions are in model training and inference (inference = usage). The inference emissions cover 18 month of use. It’s a shame they’re not detailed separately from model training, but we can guess that most of the emissions are inference - and the the same will probably be true for all AI services.
• training and 18 months of usage footprint: 20,400 tCO2e, 281,000 m3 of water consumed, 660 kg Sb eq - this is a standard unit for measuring non-living resource depletion: kg antimony equivalents per year
• marginal impacts of inference, more precisely the use of our AI assistant Le Chat for a 400-token response - excluding users’ terminals: 1.14 gCO2e, 45 mL of water, and 0.16 mg of Sb eq. (Compare to Sam Altman’s recent comments on OpenAI: “the average query uses about 0.34 watt-hours [and] about 0.000085 gallons of water”.)
• “Our study also shows a strong correlation between a model’s size and its footprint. Benchmarks have shown impacts are roughly proportional to model size: a model 10 times bigger will generate impacts one order of magnitude larger than a smaller model for the same amount of generated tokens. This highlights the importance of choosing the right model for the right use case.”

Elsewhere in AI emissions:

• The GPT-OSS models are here… and they’re energy-efficient!: “the compute efficiency of the GPT-OSS models is encouraging. In practice, this efficiency can further be improved by optimizations like batching and quantization, which can further maximize the efficiency of specific deployment configurations. The impressive efficiency of the new GPT-OSS models is great news for the AI field as it demonstrates that models can be both incredibly efficient and high-performing.”
• You can explore the historic reporting (or lack of it) from the big AI platforms in this project from Huggingface and Salesforce. And there’s a paper on it, Misinformation by Omission: The Need for More Environmental Transparency in AI June 2025.
• Sustainable AI coalition: “use of artificial intelligence (AI) to progress towards the UN Agenda 2030 and the Sustainable Development Goals (UN SDGs), focusing in this Coalition on those related to climate action and protection of our environment”

Previously: Do the AIs know their carbon emissions? No. and Working notes: carbon emissions of AI