AI’s energy and water use, in proportion
AI runs in data centres that use electricity and water. Here’s the footprint in proportion: what one prompt costs, why images and video cost more, and habits that help.
Video transcript
Every AI answer runs on computers that need electricity, and often water. So how big is AI’s footprint really? Here it is, in proportion.
Data centres run almost everything online, and they use a small but fast-growing share of the world’s electricity. AI is a big reason for that growth.
Per prompt, it’s small. Google measured a typical text prompt to Gemini at about the energy of nine seconds of television, plus about five drops of water.
Some things cost far more. Making images and video takes much more computing than text, and training a big model is a huge one-off cost.
Location matters too. Data centres cluster in certain places and often use water for cooling, so the local strain counts most where power or water is scarce.
So use AI where it really helps, write clear prompts to avoid do-overs, pick lighter models for simple jobs, and make images and video only when you need them.
And ask AI companies to publish clear figures. Read the full guide below for the numbers and where they come from.
In 30 seconds
- Data centres use a small but fast-growing share of world electricity, and AI is a big reason why.
- A typical text prompt uses very little. Images, video and model training use far more.
- Use AI where it helps, avoid needless repeats, and ask companies to publish clear figures.
Every prompt you send is answered by computers in a Jargon busterData centre: A building full of computers that store data and run online services, including AI. They need a lot of electricity, and often water for cooling., which need electricity to run and often water to stay cool. Some headlines suggest every question costs the earth; others say AI’s footprint is nothing. Neither is right. Here’s what one company has measured, what its figures leave out, and what it all means for how you use AI.
The big picture: data centres
Data centres run almost everything online, from email to streaming, and they use a small but fast-growing share of the world’s electricity. AI is a big reason for that growth: it runs on racks of powerful chips that draw a lot of power and give off a lot of heat.
The local picture can matter more than the global one. Data centres cluster in particular places, so a single area can feel the strain on its electricity grid and its water supply, even when the worldwide share looks modest.
| Uses less | Uses more |
|---|---|
| A short text answer | Making an image, and above all a video |
| A lighter, faster model for a quick question | The biggest model, for the same quick question |
| One clear prompt | Regenerating the same request again and again |
| Using a model that’s already trained | Training a big new model |
What one prompt costs
Jargon busterWatt-hour: A small unit of energy: a thousandth of a kilowatt-hour, the unit your energy bill charges for. for a typical text prompt to Google’s Gemini app, by Google’s own 2025 measurement: about the same as watching TV for less than nine seconds.
Source: Google Cloud, How much energy does Google’s AI use? We did the math, August 2025Google put the same prompt’s water use at 0.26 millilitres, about five drops, and its emissions at 0.03 grams of Jargon busterCarbon dioxide equivalent: A way of adding up all greenhouse gases as the amount of carbon dioxide that would warm the planet by the same amount.. It also said the energy per prompt fell 33-fold in a year, as it made its systems more efficient.
That’s a useful yardstick, not the whole story. It’s one company measuring its own typical text prompt. Its water figure covers cooling in Google’s data centres, while generating the electricity can use water too. And it leaves out training the model and making images or video.
Training is a big one-off cost. Meta says pretraining two of its Llama 3 models took 7.7 million hours of computing on specialised chips, with emissions it estimated at 2,290 tonnes of carbon dioxide equivalent, which it offset. Everyday use costs little per prompt, but popular assistants answer enormous numbers of prompts, so use adds up too.
Water, images and video
Many data centres use water for cooling, often letting it evaporate, and power stations use water too. That matters most where water is scarce, so where a data centre is built, and how it’s cooled, counts as much as how many prompts it answers.
Images and video cost far more than text. A text reply is a few hundred small chunks of words. A picture is usually built up over many rounds of refinement, and a video is many pictures in a row, so each one takes far more computing than a written answer.
Habits that help, and what to ask for
- Use it when it helpsReach for AI when it saves you real time or effort, rather than out of habit for every small thing.
- Aim for one good tryA clear, detailed prompt means fewer do-overs. Plan pictures in words before you generate any.
- Go small for simple jobsMany assistants offer a lighter, faster model. It’s usually fine for quick questions and needs less computing.
- Images and video on purposeMake them when you need them, and avoid generating dozens of versions to pick from.
I want an image of [what you want, and what it’s for]. Before you generate anything, describe three different ideas in one sentence each, including the style, colours and layout. I’ll pick one, and then you can make just that image.
And ask questions, especially if you buy AI for work. Does the provider publish energy, water and emissions figures for its service, and what do they include? Where are its data centres, and how are they powered and cooled? Not every company publishes figures, and those that do may measure different things, so customers asking helps.
I’m choosing an AI tool for [my team / my organisation]. Draft six short, plain-English questions to ask suppliers about the energy, water and carbon footprint of their service, including what their published figures include and leave out, and where their data centres are.
Check yourself
3 quick questions nothing is savedTools in this guide
Sources (2)
- How much energy does Google’s AI use? We did the mathGoogle Cloud, August 2025
- Meta Llama 3 model cardMeta, April 2024
Spotted a mistake? Tell us and an editor will check it.
