How much energy and water does AI use? The cost of each ChatGPT or Gemini question, explained simply
Every time you ask a chatbot something, a data center uses electricity and water. Here are Google's and OpenAI's official figures, why to take them with a grain of salt, how much data centers use worldwide, and what you can do.

AI feels like it lives “in the cloud,” weightless. But every answer from ChatGPT, Gemini or Claude comes from a data center: huge buildings full of computers that use electricity and often water to keep cool.
How much does one question cost? Is that a lot? Here’s what the companies themselves and the International Energy Agency have published, and the limits of those numbers.
At a glance
| Figure | Number | Source |
|---|---|---|
| Energy for one Gemini text prompt (median) | 0.24 Wh | Google (May 2025) |
| Water for that prompt | 0.26 mL (about five drops) | |
| Energy for an average ChatGPT query | 0.34 Wh | Sam Altman, OpenAI CEO (June 2025) |
| Water for that query | About one fifteenth of a teaspoon | Sam Altman |
| Electricity used by all data centers in 2024 | 415 TWh, about 1.5% of the world’s total | IEA |
| Forecast for 2030 | Around 945 TWh, more than double | IEA |
One question: not much
Google says a typical Gemini text prompt uses about as much energy as watching TV for under nine seconds. Sam Altman compares a ChatGPT query to running an oven for a little over a second.
For perspective: a typical phone battery holds roughly 15 Wh. By these figures, that’s on the order of 50 text questions.
Google adds a striking detail: between May 2024 and May 2025, energy per prompt fell 33-fold thanks to more efficient models and chips.
But take it with a grain of salt
- These are the companies’ own numbers. Google says its figures haven’t been independently verified. Altman didn’t explain how he got his.
- It’s a “median” text prompt. Long answers, document analysis, “deep thinking” modes or agents working for minutes use much more.
- Images and video aren’t included. Generating an image, and especially a video, uses far more than text, but companies publish little about it.
- Training isn’t included. Building a new model takes months of computing on thousands of chips. That bill doesn’t show up in the “per question” cost.
- Anthropic (the company behind Claude) hasn’t published comparable figures.
The problem is the total
One question is small. The challenge is that there are hundreds of millions every day, and more and more tasks use AI.
According to the International Energy Agency (IEA):
- Data centers used about 415 TWh in 2024, around 1.5% of the world’s electricity. The US accounts for 45%, China 25% and Europe 15%.
- By 2030 they could reach about 945 TWh, a bit more than all of Japan uses today.
- AI is the main driver of that growth, alongside other digital services.
- One AI-focused data center already uses as much electricity as 100,000 households, and the biggest ones being built will use far more.
So the real debate isn’t “does my question pollute?” but where the electricity comes from (renewables or not), how much water is used in drought-prone areas, and whether power grids can keep up.
What about water?
Computers run hot. Many data centers cool down by evaporating water, and producing their electricity uses water too. The impact depends heavily on where the data center is: a rainy region is very different from a dry one. Google says it aims to replenish 120% of the freshwater it consumes, though that’s a company goal.
What can you do?
You don’t need to stop using AI, just use it wisely:
- Pick the right tool. For a game time or the weather, a regular search is enough.
- Ask well the first time. A clear request saves five retries. See our guide on how to write good prompts.
- Don’t generate images or videos dozens of times “just to see” if you won’t use them.
- Use lighter models for simple tasks: many apps have a fast mode that uses less.
- Keep it in perspective. Streaming a show, driving or heating your home weigh much more in your day than a few chatbot questions.
Bottom line
Each AI question uses little: a few seconds of TV and a few drops of water, according to the companies. The issue is scale: data centers could double their electricity use by 2030, with AI as the main reason. If you want to understand how it all works, start with what artificial intelligence is.


