The Carbon Footprint of AI Models
Let’s Play “The Price Is Right”: How Much CO₂ Does Your ChatGPT query emit? A gram? 100 grams? 10 tons? The answer is far from simple. According to Google’s data on its AI Gemini, the environmental impact seems negligible: a simple query would cost the planet just a few milligrams of carbon and a tiny amount of water. But in reality, the answer depends mainly on the calculation methodology, the complexity of the query, and the transparency level of tech giants. In this Q&A, we explore whether AI and its data centers will leave us any water in the washbasin.
This summary is based on an excellent article in French about the environmental impact of AI, written by Lou Welgryn and Théo Alves Da Costa and published on the media outlet Bon Pote.
Where Do the CO₂ Emissions of an AI Model Come From?
Most of an AI model’s carbon footprint comes from the electricity consumption of GPUs in data centers — chips optimized for the calculations required by AI models. Their average power is 700 watts, equivalent to your office microwave!
Large Language Models (LLMs) like OpenAI’s ChatGPT or Google’s Gemini first need to be trained on colossal amounts of data. The training phase involves massive calculations performed by these “microwave computers.” The power required for AI model training has been doubling every six months since 2012.
However, this isn’t what makes data centers sweat the most. About 80% of emissions come from the queries sent to AI models: requests for text, image, or video generation. When you send a query, your prompt is routed to a data center, where GPUs perform calculations before sending the result back to your computer or phone.
So, How Much CO₂ Does a ChatGPT Query Actually Emit?
Today, it’s difficult to pinpoint an exact number for two reasons: the context of the query can vary enormously, and AI model developers disclose very little about their carbon impact.
As you’ve understood, the carbon footprint of an AI query largely depends on the electricity consumption of data center GPUs. If your data center is in France, you benefit from low-carbon electricity (52 grams of CO₂/kWh), as nuclear power has low emissions. However, if your data center is in the United States, you’ll consume electricity that is about 10 times more carbon-intensive (520 gCO₂/kWh). Currently, most AI model data centers are located in the U.S.
On your end, asking ChatGPT to generate a Valentine’s Day poem, an image, or an entire video game drastically changes the amount of computation — and thus the carbon impact. Is your query two sentences long, or does it require analyzing a 1,320-page PDF? Are you asking for a concise answer or the drafting of a novel? Are you using Gemini 3 Pro, Mistral Large 3, or Claude Opus 4.6? All these parameters significantly influence the carbon footprint of your query.
Finally, it’s hard to get a precise calculation of the carbon impact due to sector opacity: today, 84% of queries go through models with no environmental information. As for the few available data points, they are rarely reliable or verifiable. A notable exception is Mistral AI, which has transparently communicated the environmental impacts of its Large 2 model.
Are You Sure? Don’t We Have Any Idea?
There’s no precise calculation, but estimates have been made based on open-source AI models. The International Energy Agency (IEA) published a study in April 2025. The two graphs below show the amount of electricity consumed for query calculations (in green, by GPUs) and for device charging (in purple).
Warning: The scales differ in the two graphs. Text generation requires a few watt-hours (Wh) of electricity, while generating a 6-second video requires over 100 Wh — enough to charge your laptop twice!
Generating a video longer than one minute is equivalent to running a microwave for one hour.
Text generation varies greatly depending on the model size: “low” is obtained with a 9-billion-parameter model, while “high” is based on a 70-billion-parameter model. The latest models can reach hundreds of billions of parameters.
By converting electricity consumption into CO₂ emission equivalents, the carbon impact of different types of AI queries is shown below. Remember, most AI model data centers are in the U.S., where electricity is, on average, highly carbon-intensive.
Source: Author, based on IEA’s “Energy and AI 2025” data for electricity consumption and ADEME for emission factors.
Source: Author, based on IEA’s “Energy and AI 2025” data for electricity consumption and ADEME for emission factors.
So, If I Only Use AI to Generate Text, It’s Not That Bad?
Generating synthetic text from short prompts is indeed much less impactful than generating AI Slop videos for social media. However, a simple Google search will likely be less emissive than an AI query, as it requires fewer calculations. Even better: don’t forget that your brain is a very energy-efficient machine. Before asking AI for a solution, why not rely on our wonderful cognitive abilities?
What Does the Overall Carbon Impact of AI Look Like?
A single AI query may seem low-impact, but ChatGPT alone receives over 2.5 billion queries per day (as of October 2025, and the number has likely increased since).
The combined training of models and AI queries has led to a sharp increase in CO₂ emissions from model developers. Google doubled its emissions between 2019 and 2023, while Microsoft increased its emissions by 30%.
Source: IFP
To meet the exploding electricity demand of data centers, coal plant closures have been delayed, and new gas plant projects have been announced (coal and gas are fossil fuels that generate significant CO₂, especially coal). In the U.S., 75% of data centers under construction will be gas-powered. It’s not uncommon for data centers to also be powered on-site by diesel generators.
Projections for the coming years suggest this upward trend will continue. In France, data center electricity consumption could triple by 2035 if no ecological transition policies are implemented. Emmanuel Macron wants to strongly develop data centers in France. In early February 2026, €109 billion in private AI investments were secured, much of it for data center construction. According to the government, such a policy remains compatible with sustainable AI, but not all commentators agree.
Even if precise data on the carbon impact of an AI query is hard to come by, we can confidently say that AI models are responsible for a significant increase in the carbon footprint of developers like Google and Microsoft. This trend will likely continue in the coming years. Meanwhile, AI promoters put forward somewhat audacious arguments. Sam Altman, CEO of OpenAI, defends AI’s carbon footprint by explaining that humans also need a lot of energy before they become intelligent enough to solve problems. So, would it be better for the planet to design AI rather than humans?
Note: The environmental impacts of AI are not limited to CO₂ emissions. We can also mention the water consumption of data centers, which leads to local usage conflicts, or the resource consumption for manufacturing GPUs and data centers. The author focused on the carbon impact to simplify this article.
Benjamin Karas, student at 42 and Climate Consultant
Article written in French without AI assistance, translated in English with the help of Mistral AI.
