Artificial intelligence has become a reflex. We ask it for a summary, a sales follow-up, a summary table, without even thinking about it. The gesture seems free, instant and clean. Behind the screen, it isn’t.
Let’s be frank: AI pollutes. It isn’t a technical detail that a better algorithm will erase tomorrow. It is a very real environmental cost, in electricity, water and metals, rising as fast as usage. Denying it would be dishonest.
Why does artificial intelligence pollute so much?
AI’s energy consumption: the weight of electricity
Taken on its own, a single query looks negligible. Google cites a median of 0.24 Wh per text query on Gemini, and OpenAI about 0.34 Wh per query on ChatGPT (Google, 2025, and S. Altman’s blog, OpenAI, 2025). But these figures are self-reported and unaudited, and above all, they are multiplied by billions of queries every day. At that scale, the total becomes staggering.
A growing need for data centers
AI doesn’t run “in the cloud” in some abstract sense: it runs in data centers, huge warehouses full of servers. These don’t serve AI alone; they also host cloud services, streaming and storage. But AI has become their fastest-growing driver: according to the International Energy Agency, the consumption of servers dedicated to AI is growing by about 30% a year and on its own accounts for nearly half of the increase in data center consumption expected by 2030 (IEA, 2026).
Locally, the effect is already spectacular: in Ireland, data centers absorbed 22% of the country’s electricity in 2024 (CSO, Ireland, 2025). In France, their consumption, about 10 TWh today, could triple to 23 to 28 TWh by 2035 (RTE, 2025).
The problem is that demand is moving faster than the grids. In the United States, data center construction is far outpacing the modernization of the power grid: rather than waiting for the grid to catch up, some operators draw their power directly from gas. The International Energy Agency estimates that between 15 and 27 GW of on-site gas could power US data centers by 2030 (IEA, 2026). Carbon-based energy, sometimes from old power plants brought back into service.
AI’s water consumption: billions of cubic meters to cool the servers
The servers that run the models heat up, and they are cooled with massive amounts of water. The few drops announced per query only tell part of the story: they leave out the water used to produce the electricity, often far more. By 2027, according to a study by researchers at the University of California, Riverside, cited by the United Nations Regional Information Centre for Western Europe (UNRIC), global demand linked to AI could require between 4.2 and 6.6 billion cubic meters of water a year, or 4 to 6 times the annual consumption of Denmark (UNRIC, 2025).
And water is only one side of the problem: on a global scale, AI systems may already emit as much CO₂ as New York City (de Vries-Gao, Patterns, 2025).
Servers and rare metals: AI’s invisible material footprint
Finally, there is the hardware. The graphics processors (GPUs) and semiconductors that train and run the models require strategic metals such as copper, used heavily in the servers and cabling of data centers. The Banque de France has raised the alarm on this point: the boom in AI and data centers is swelling a demand for copper already driven by the energy transition (batteries, wind turbines, solar panels), at the risk of worsening the imbalances of this market, with issues of dependency, shortages and even financial stability (Banque de France, Eco Notepad no. 402, 2025).
The environment isn’t the only cost of AI: rolled out badly, it also weighs on employees. Read: how AI affects employees’ mental health
How do you make AI greener and reduce its carbon footprint?
The picture is grim, but nothing is set in stone. Some levers can reduce this footprint, as long as they don’t become an alibi for consuming ever more.
Make the electricity that powers it greener
This is lever number one, because everything depends on the electricity mix. In China, nearly 70% of the electricity used by data centers comes from coal (IEA, 2026), whereas in France, with its largely nuclear mix, the same query emits far less, an asset that RTE (the French electricity transmission operator) sees as a real card to play. The good news: by 2030, renewables should cover nearly half of the increase in data center consumption (IEA, 2026). Building data centers where electricity is clean, signing green power contracts, waiting for the grid to catch up rather than burning gas: all choices that radically change the footprint.
Use AI wisely: choose and measure
Not every task deserves AI. Automating a repetitive process that saves hours makes sense; generating ten versions of a visual “just to see” makes much less. The first question isn’t “can we do it with AI?” but “is it worth the cost?”.
Frugal AI is measured AI. Hours saved, tasks actually automated, uses dropped: without these metrics, tools pile up without anyone knowing whether they are useful, and the environmental cost becomes one more blind spot.
Frugal AI and the new European regulations
Frugality doesn’t have to be reinvented: frameworks already exist to put it into practice. In France, it even has a name: frugal AI, set out in a reference framework from the French government (Ecolab) and AFNOR, the French standards body. Three principles sum it up: check that you really need AI rather than a lighter solution, apply good practices to limit its impact, and question your uses in light of planetary boundaries (Ecolab/AFNOR, AFNOR Spec 2314, 2024).
Regulation, for its part, is starting to set concrete obligations. In the European Union, data center operators must now report their performance indicators every year to a European database, under the new scheme adopted by the Commission to rate their sustainability. The stated goal: more transparency and, eventually, lower energy and water consumption, more renewables and better reuse of the heat from servers (European Commission, 2024). The scheme can still be improved: it relies mostly on self-reporting, and its usefulness will depend on how firmly it is enforced. But the principle is set, and what is measured can finally be managed.
Conclusion
The impact of AI on the environment is real, and it is growing. But it depends on the electricity mix that powers it, on technical progress, and above all on what we choose to do with it.
We can make AI greener. Above all, we can decide to use it only where it brings a real gain.
Useful AI is, first of all, AI we have chosen to use wisely.
FAQ: your questions about AI’s environmental impact
What is the impact of AI on the environment?
AI has a real environmental cost, in electricity, water and metals, which rises with usage. According to the International Energy Agency, the consumption of servers dedicated to AI is growing by about 30% a year (IEA, 2026). By 2027, global demand linked to AI could require between 4.2 and 6.6 billion cubic meters of water a year (UNRIC, 2025), and AI systems may already emit as much CO₂ as New York City (de Vries-Gao, Patterns, 2025).
How much energy does a ChatGPT query use?
OpenAI cites about 0.34 Wh per query on ChatGPT, and Google a median of 0.24 Wh per text query on Gemini (2025). These figures are self-reported and unaudited, and above all they are multiplied by billions of queries every day. Nor do they account for all the water consumed, in particular the water used to produce the electricity.
How do you reduce AI’s environmental footprint?
The first lever is the electricity that powers data centers: the same query emits far less in France, with its largely nuclear mix, than in China, where nearly 70% of that electricity comes from coal (IEA, 2026). The second is usage: use AI only where it brings a real gain, and measure hours saved, tasks automated and uses dropped. The frugal AI reference framework of the French government (Ecolab) and AFNOR provides the method.
