How much pollution does AI create? Mistral breaks it down

[Alfonso Maruccia]

What just happened? Mistral AI, a Paris-based venture focused on developing open-weight, open-source large language models, is now contributing to the AI discourse with a comprehensive lifecycle analysis of one of its models – guided by the same principles of openness that define its work.

Mistral recently published an analysis of the environmental impact of one of its large language models. As chatbots and other AI-powered technologies become increasingly embedded in the global economy, Mistral emphasized the importance of understanding – or at least estimating – the environmental footprint of these systems. The impact, the company warned, can be substantial.

To conduct the study, Mistral partnered with sustainability consultancy Carbone 4 and the French ecological transition agency. The findings were also peer-reviewed by environmental consulting firms Resilio and Hubblo. The analysis focused on the full lifecycle of the Mistral AI Large 2 model, evaluating its impact across three key areas: greenhouse gas emissions, water use and depletion, and material consumption.

Unsurprisingly, the study confirms that the most environmentally taxing phases of an AI model's lifecycle are training and inference. According to Mistral, 85.5 percent of the model's total GHG emissions and 91 percent of its water consumption occurred during model development and user interaction.

As of January 2025, after 18 months of operation, Mistral's Large 2 model had generated 20.4 kilotons of CO₂ emissions and consumed 281,000 cubic meters of water. The marginal impact of inference – measured by a user prompting the "Le Chat" chatbot for a 400-token response – was estimated at 1.14 grams of CO₂ and 45 milliliters of water.

According to Mistral, this is roughly equivalent to the greenhouse gas emissions generated by watching a streaming video for 10 seconds in the US. In simpler terms, a single chatbot query won't melt Antarctica, but millions or billions of users interacting with these systems over weeks and months will undoubtedly contribute to the planet's environmental challenges.

Mistral acknowledges several limitations in its study, particularly the difficulty of accurately quantifying hardware degradation caused by LLM workloads on GPUs and data center infrastructure. Even so, the reported figures align closely with environmental impact estimates from other research and institutions.

The French AI company plans to release updated versions of this environmental report in the future and is calling for greater transparency across the AI industry. Its stated goal is to help steer the sector toward alignment with global climate objectives, though it notes that the Trump administration is currently pursuing policies that run counter to that aim.

Permalink to story:

How much pollution does AI create? Mistral breaks it down

 

[Megalomaniac]

To live is to produce waste.
I find it hilarious these people who constantly want to talk about "pollutants" yet can't talk about the growing population since to even mention them being a problem would bring about accusations of eugenics or population control.

One -One billionth of the energy coming from the Sun reaches Earth. The energy is going to keep coming whether we use it or not. Long after the radiation within the Earth dissipates and it freezes inside out.

Would they be complaining if solar panels were set up to create the energy that goes to products and services they have a problem with? Of course they would!.

 

[stewi0001]

I was reading an article earlier today from MSN/Business Insider that residents could be paying more because of the data centers. https://www.msn.com/en-us/money/mar...bill-next-year-blame-data-centers/ar-AA1JqjY3

In the video they also point out other factors like sound pollution (if you live close enough) and demand for water.

Out of curiosity I even asked copilot about AI's impact on the environment. I was a little surprised that it admitted there is a problem.

The most energy-consuming part of AI is typically the training phase of large-scale models. Here's a breakdown of where the biggest environmental costs come from:

1. Model Training (Most Consuming)
Why it's costly: Training large models like GPT-4, DALL·E, or AlphaFold involves running billions to trillions of computations over weeks or months on thousands of GPUs or TPUs.
Energy use: This phase can consume hundreds of megawatt-hours of electricity.
Carbon footprint: If powered by fossil fuels, this can emit dozens to hundreds of tons of CO₂.
2. Inference (Running the Model)
Why it matters: Once trained, models are used millions or billions of times (e.g., answering questions, generating images).
Energy use: Each use is relatively small, but the scale of usage makes it significant.
Optimization: Companies often optimize inference for efficiency using quantization, distillation, or edge computing.
3. Hyperparameter Tuning & Experimentation
Why it's costly: Before final training, researchers run many trial runs to find the best model architecture and settings.
Energy use: This can multiply the total training cost by 2x to 10x, depending on how many experiments are run.
4. Data Storage and Movement
Why it matters: Massive datasets (e.g., Common Crawl, image/video corpora) are stored and moved across data centers.
Energy use: Less than training, but still non-trivial due to networking and storage infrastructure.

TLDR: Training, Tuning, and Experimentation are the highest impacts followed by a ton of people using it.

 

[opckieran]

With the explosive demand for huge, specialized, enterprise-grade GPUs (and their nodes as a whole), I’m curious as to how they will be decommissioned. Will they be sold in whole like what happened with Cheyenne, piecemeal on eBay, or just reclaimed through recycling?

 

[NotYourITguy]

"though it notes that the Trump administration is currently pursuing policies that run counter to that aim."

The linked article has zero mention of any policies specific to power consumption or adding taxes on energy. Can we please keep this as techspot not trumpspot?

 

[p51d007]

LOL, I wonder how much "Co2 gas" was put up in the atmosphere during Bezo's wedding with all those corporate jets flying around.

 

[AnilD]

Props to Mistral for actually pulling back the curtain on AI’s environmental cost. Everyone else is out here dropping billion-parameter models like they grow on carbon-neutral trees.

 

[monkeylove]

To live is to produce waste.
I find it hilarious these people who constantly want to talk about "pollutants" yet can't talk about the growing population since to even mention them being a problem would bring about accusations of eugenics or population control.

One -One billionth of the energy coming from the Sun reaches Earth. The energy is going to keep coming whether we use it or not. Long after the radiation within the Earth dissipates and it freezes inside out.

Would they be complaining if solar panels were set up to create the energy that goes to products and services they have a problem with? Of course they would!.

Capitalist systems require increasing numbers of people because they need more workers and consumers each time. That's because for-profit, competing businesses want increasing profits, which means increasing production and thus more workers, and increasing sales and thus more consumers (who also happen to be the same workers). Otherwise, they end up with lower sales and population ageing.