Is AI bad for the environment? Energy, water and what you can do
AI does use real electricity and water, mostly in the data centres that run it. This guide explains why published figures disagree so much, the difference between training a model and asking it a question, how a single chat compares with everyday things, and a few proportionate steps you can take.
New here? Before this guide, you might want to read How AI models are trained.
Yes, AI has a real environmental cost, but for a single typical question it is small. The electricity and water go mainly into data centres, and published estimates vary widely because they count different things. The bigger issue is scale: billions of daily questions, the one-off cost of training each new model, and how fast data centres are growing.
In this guide, you will learn where AI’s footprint comes from, why the numbers you read disagree, and what you can sensibly do about it.
Where does AI’s footprint come from?
When you type a question into ChatGPT, Claude, or Gemini, the work does not happen on your phone. It happens on powerful computers in large buildings called data centres. Those computers need electricity, and they get hot, so they need cooling, which often uses water.
There are two quite different moments when this happens:
- Training. Before a model can answer anything, it learns from a huge collection of text over weeks or months, using thousands of specialised chips. This is a one-off, very large burst of energy for each new model.
- Everyday use. Each time someone asks a question, the finished model does a smaller burst of work to produce the answer. This step is called inference. It is tiny per question but happens billions of times a day.
Why do the estimates disagree so much?
You may have seen claims that a chat “drinks a bottle of water”, and others saying it uses a few drops. Both can be honest, because they measure different things. Here are the main reasons figures vary:
| What differs | Why it changes the number |
|---|---|
| Training or use | Some figures include a share of the training cost, others count only the question itself. |
| Water on site or off site | Cooling water at the data centre is one amount. Water used by power stations to make the electricity can be far larger. |
| Where the electricity comes from | The same question has a smaller carbon footprint on a grid powered by wind and solar than on one powered by coal or gas. |
| Local climate and cooling design | A data centre in a hot, dry place may use much more water than one in a cool climate. |
| Which model and what task | A short text question, a long document, and a generated video are very different jobs. |
| When it was measured | Chips and models have become much more efficient, so older estimates are often higher. |
So when you see a single, dramatic number, ask what it includes before deciding what it means.
So how much does one question use?
Here are the most widely cited estimates, with who made them and when. Notice the range.
- Electricity per text question: Google (2025) reported about 0.24 watt-hours for a median Gemini text prompt. OpenAI’s chief executive Sam Altman (2025) gave about 0.34 watt-hours for an average ChatGPT question. The research group Epoch AI (2025) estimated around 0.3 watt-hours for a typical question, rising to between 2.5 and 40 watt-hours when you paste in very long documents.
- Older, higher estimates: the Electric Power Research Institute (2024) used about 2.9 watt-hours per ChatGPT request, which is where the popular “ten times a web search” claim comes from. Epoch AI’s later work suggests that figure is now likely an overestimate.
- Water per text question: Google (2025) reported about 0.26 millilitres, roughly five drops, counting cooling water at its data centres. Mistral AI (2025) reported about 45 millilitres for a 400-token answer from its assistant, because its study also counted water used in making the electricity and building the hardware.
- The “bottle of water” figure: researchers at the University of California, Riverside (2023) estimated that 10 to 50 medium-length answers from an older model, GPT-3, could use about 500 millilitres, depending on where and when they ran.
Most of these figures come from the companies themselves and have not all been independently checked, so treat them as useful ranges rather than final answers.
And what about training?
Training is where the big single numbers live. Mistral AI’s 2025 study found that training its Mistral Large 2 model, together with 18 months of use, accounted for about 20,400 tonnes of carbon dioxide equivalent and 281,000 cubic metres of water. It also found that training made up the large majority of those totals.
The same UC Riverside paper (2023) estimated that training GPT-3 in Microsoft’s US data centres could have consumed around 700,000 litres of water on site, and about 5.4 million litres once electricity generation was included.
How big is the bigger picture?
The International Energy Agency (2025) estimated that data centres used about 415 terawatt-hours of electricity in 2024, roughly 1.5% of the world’s total. It expects that to more than double to around 945 terawatt-hours by 2030, with AI a major driver.
That is why the honest answer has two parts. Your own chat is a small thing. The industry’s growth is a big thing, and it raises fair questions about local water supplies, electricity grids, and how quickly clean power can keep up.
Putting one chat in perspective
By the company figures above, a typical text question uses about as much electricity as a modern LED bulb running for a couple of minutes, which is the comparison OpenAI itself gave. By our rough arithmetic, boiling a full kettle uses about as much as several hundred of those questions.
That does not make it free, and heavier tasks cost much more. Generating images and video, running long “deep research” jobs, or pasting in enormous documents all take far more computing than a quick question.
What can you do?
None of this calls for guilt. A few proportionate habits make a sensible difference:
- Use AI where it earns its place. A question that saves you an hour is a good trade. Asking it to repeat something a quick search would answer is less so.
- Keep requests focused. A clear first prompt means fewer retries.
- Save the heavy tasks for when you need them. Video, large batches of images, and long research runs use the most.
- Pick the lighter option when it will do. Many tools offer a smaller, faster model that handles everyday questions well.
- Keep an eye on the bigger picture. Where companies build data centres, and how they power them, matters far more than any one person’s chats.
A clear first prompt is the easiest win. Try something like this:
I need [what you want] for [who it is for]. Please give it to me in [format],
about [length]. If anything is unclear, ask me one question before you start
rather than guessing.Next steps
If you want to understand why training takes so much computing in the first place, our guide on how AI models are trained walks through it. For the wider picture of risks and safeguards, see AI safety and ethics.
Sources
- International Energy Agency, Energy and AI (2025)
- Google, Measuring the environmental impact of AI inference and technical paper (2025)
- Sam Altman, OpenAI, The Gentle Singularity (2025)
- Epoch AI, How much energy does ChatGPT use? (2025)
- Mistral AI with Carbone 4 and ADEME, Our contribution to a global environmental standard for AI (2025)
- Electric Power Research Institute, Powering Intelligence (2024)
- Pengfei Li, Shaolei Ren and colleagues, University of California, Riverside, Making AI Less “Thirsty” (2023)
Frequently asked questions
- Does ChatGPT use water?
- Yes, indirectly. The data centres that run it use water to keep their computers cool, and power stations use water to make the electricity. Company figures for a single typical question range from a fraction of a millilitre to around 45 millilitres, depending on what is counted.
- Is one chat with AI bad for the environment?
- A single typical text question has a small footprint, roughly comparable to running a modern LED bulb for a couple of minutes, according to recent company estimates. The bigger concern is the combined effect of billions of questions a day and the fast growth of data centres.
- Which uses more energy, training AI or using it?
- Training one large model is a single, very large burst of energy and water. Everyday use is tiny per question but happens billions of times, so over a model's lifetime both matter, and studies weigh them differently.
- Should I stop using AI to help the planet?
- You do not need to give it up to make a difference. Using it for things that are worth it, keeping requests focused, and saving heavy tasks like video generation for when you need them is a proportionate approach.