AI

How much energy and water does AI use? A plain guide

One AI question costs very little. The data centres behind millions of them cost a lot, and their water use has grown. What the figures say, in plain words.

By Brahim Eloidii, Co-Founder, ESG & CSR Lead

Published 22 October 2026 · Sources last checked 22 October 2026

First, what is a data centre, and why does it need water?

A data centre is a large building full of computers that run day and night. Computers give off heat, and the heat has to be removed or they fail. Many data centres do this with water. In the most common design, water absorbs the heat and then evaporates into the air. Water that evaporates is counted as consumed, because it does not go back to where it came from. Other designs recirculate the water in a closed loop and use far less of it, but they need more electricity to run chillers. So cutting one often raises the other.

The energy: tiny per question, large in total

A watt-hour is a small unit of energy: a 10-watt LED bulb burning for six minutes uses one. A TWh (terawatt-hour) is a billion kilowatt-hours, the unit used for whole countries. The International Energy Agency (IEA) estimates that data centres worldwide used about 415 TWh in 2024, around 1.5% of the world’s electricity. In its April 2025 Energy and AI report it projects roughly 945 TWh by 2030, slightly more than all of Japan uses today, with AI as the main driver. It also says the picture is uncertain, as it depends on how fast AI is adopted and how fast it becomes more efficient.

Seen from one user, it is much smaller. In August 2025 Google published its own estimate for a median text question to its Gemini app: 0.24 watt-hours, which Google compares to about nine seconds of TV, 0.03 grams of CO2 equivalent and 0.26 millilitres of water, which it compares to about five drops. Google also says these figures fell 33 times for energy and 44 times for carbon in the 12 months to May 2025.

The figures disagree because they count different things. Counting only the active chips, Google’s own comparison gives 0.10 watt-hours. Counting the whole system, including idle machines kept ready and the building’s overhead, it gives 0.24. An outside estimate by the research institute Epoch AI, from February 2025, puts a typical ChatGPT question at about 0.3 watt-hours, ten times less than the roughly 3 watt-hours quoted since 2023. Epoch says its figure is approximate, with about a factor of two of uncertainty, because OpenAI has not published the details.

The water: two kinds, and the second is bigger

Researchers separate direct water, used inside the data centre to cool it, from indirect water, used by the power stations that make its electricity. A 2024 report by the Lawrence Berkeley National Laboratory in the United States estimated that US data centres consumed about 64 billion litres (17 billion gallons) directly in 2023. It put the indirect water at about 800 billion litres (211 billion gallons), around 12 times more. It projects that the direct figure could double or even quadruple by 2028.

Those are American totals. The point to keep is that the water behind AI is not only the water you can see at the data centre: a large part is used where the electricity is made.

Before and after: what the big companies report

The clearest view comes from the companies that publish their numbers. Google reports that its data centres consumed about 4.3 billion gallons of water in 2021. For 2023 it reported about 6.4 billion gallons across all its operations, about 95% of it in data centres. Coverage of its latest report says consumption grew 34% to about 10.9 billion gallons in 2025. Microsoft reports a similar jump: about 4.8 million cubic metres of water in 2021 and about 6.4 million in 2022, up 34%, which the press links in part to generative AI.

In plain terms, Google’s reported water use went from the equivalent of about 6,500 Olympic swimming pools a year in 2021 to about 16,500 in 2025, which is two and a half times as much. The years, the scope and the way the companies count are not identical, so read it as a direction, not as a precise comparison.

Why location matters more than the total

Water is a local resource. The same amount of water means little in a wet region and a lot in a dry one. The researchers who summarised this field give two examples from Google in 2024. One data centre in Iowa consumed about 1 billion gallons, which they say is enough to supply all of Iowa’s homes for about five days. One in Texas used about 10,000 gallons, about what one Texas home uses in two months. They add that in some places data centres use more than 25% of the local community’s water supply. For a supplier choosing a cloud region, where the servers sit is part of the answer.

When one question gets bigger

• Long documents. Epoch estimates about 2.5 watt-hours for an input of 10,000 tokens, roughly a long report, and almost 40 watt-hours for 100,000 tokens, roughly a short book. A token is a small piece of a word.
• Reasoning models. They write much more before they answer, and likely use more energy per question.
• Images and video. The figures above are for text. Epoch’s estimate leaves out image generation.
• Everything around the question. Training the models and making the chips are not in these per-question figures.

What it means for a supplier: an illustration

Take a simple case with stated assumptions: a team sends 1,000 text questions a day over 250 working days. At Google’s figures, that is about 60 kilowatt-hours of electricity a year, about 7.5 kilograms of CO2 equivalent and about 65 litres of water, roughly six and a half ten-litre buckets. These are our own calculations from Google’s numbers, for Google’s systems. Compare 60 kilowatt-hours with your electricity meter. A plant that makes steel, aluminium, cement or fertiliser is almost certainly in a different league.

The conclusion for an exporter is a matter of proportion. For goods covered by CBAM, the emissions that count are those of production: the furnaces and the electricity supply, not the chatbot. Heavy use of AI on whole documents, images or video is different, and worth knowing about.

What to say when a customer asks

1. Know what you use: which tools, who uses them, for what, and whether it involves long documents, images or video.
2. Ask your AI or cloud provider which energy and water data it publishes, and in which region your data is processed. Many publish little, and you can say so.
3. If you calculate your emissions, a bought AI or cloud service is a purchased service, and the provider’s data is the place to start. Do not invent a factor.
4. Keep the claim proportionate. Do not say that your AI is green, carbon neutral or water neutral unless you can show how.
5. Write a short statement and keep it with your other ESG documents.
6. Re-check it regularly. These figures are moving fast, in both directions.

If you export goods covered by CBAM, the free CBAM Compass Report shows where you stand on the emissions that actually set your customers’ bill, in about three minutes.