New research from Mistral AI adds to the growing database of the shocking environmental impact of AI models, from their intensive training to everyday use.

The French AI model creator published a peer-reviewed report on the environmental impact of its Mistral Large 2 LLM across three key figures: greenhouse gas emissions, water consumption, and materials use.

With less than 18 months of existence, the Large 2 LLM generated 20,4 ktCO2e of greenhouse gases, consumed 281,000 cubit metres of water, and used 660 kg Sb eq of materials. The water usage is on par with 112 Olympic swimming pools.

What mainly drove these figures as the training and running of the model, otherwise known as inference, which accounted for 85.5% of all greenhouse gas emissions, and 91% of water consumption in the 18 months.

Mistal also tracked the amount of resource use and emissions over smaller use cases. Generating one page of text using the model emitted 1.14g CO2e of greenhouse gases, the same as watching online streaming in the US for 10 seconds, or 55 second for a user in France, and 0.05 litres of water, the same amount required in growing a small pink radish. The materials consumed is the equivalent of those used in producing a 2 euro cent coin, or 0.2 mg Sb eq.

“These figures reflect the scale of computation involved in Gen AI, requiring numerous GPUs, often in regions with carbon-intensive electricity and sometimes water stress,” the research said. “They also include “upstream emissions” – impacts from manufacturing servers for instance, and not just energy use.”

Further, while these figures may seem inconsequential, just a small number of users required just a page of text can see these numbers skyrocket.

Mistral found a strong correlation between a model’s size and its environmental 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,” the study noted.

Recommended reading

• Data Centre Emissions Could Be 600% Worse Than Big Tech Claims
• Report: UK Public Unaware of GenAI’s Environmental Impact
• What is the Carbon Cost of AI?
• Placing Data Centres in Scotland Could Cut Carbon Significantly

This underlines the importance of selecting the right model for the right use cases – bigger does not always mean better, but it does tend to mean a higher environmental impact.

The report also highlighted the need for standardised environmental impact monitoring for AI models, from their training to their everyday use, the better track these metrics and understand its total impact on the greater environment.

While the metrics tracked in the study are important, they did not highlight the impact of the specific materials used that needed to be mined or manufactured, as well as the impact of data centres on water pollution and quality.

The research did highlight, however, that the geographic location of AI model training data centres is very important in mitigating their impact. Choosing cooler areas with more readily available water can significantly decrease the environmental impact of data centres, which require more energy and water to cool themselves down.