What one AI prompt really costs: electricity, water and the number to show your class

Numbers of every size are doing the rounds online: every ChatGPT question drinks a glass of water, AI uses more power than whole countries, or it is all a fuss about nothing. None of them is any use for a presentation if you cannot say where it comes from. Below are only the figures the companies themselves have published or that come from the International Energy Agency, with links, so you can check them before they go on a slide.

What a prompt costs, according to the companies

On 21 August 2025 Google published its measurement for the Gemini apps: the median text prompt uses 0.24 watt-hours of electricity, 0.26 millilitres of water (about five drops, Google says) and emits 0.03 grams of CO2. That is less than nine seconds of watching TV, in Google's words. The count includes chips sitting idle in reserve, the machine's CPU and memory, and the data centre's cooling, and Google claims that over twelve months the energy per prompt fell 33 times and the carbon footprint 44 times. Sam Altman, the head of OpenAI, wrote on his blog on 10 June 2025 that the average ChatGPT query uses about 0.34 watt-hours, what an oven uses in a little over a second or an efficient light bulb in a couple of minutes, and 0.000085 gallons of water, roughly one fifteenth of a teaspoon, which is about 0.32 millilitres. No word on how it was measured.

• An 800 watt microwave: 0.24 watt-hours is just over one second of running. With Altman's figure, a second and a half.
• A 10 watt LED bulb: almost a minute and a half of light per prompt. Two minutes with the ChatGPT figure.
• A phone battery (around 15 watt-hours): about sixty prompts equal one full charge, not counting charging losses.
• Water: 0.26 millilitres is five drops. A 250 millilitre glass covers almost a thousand prompts.
• CO2: a thousand prompts emit 30 grams.
• Your whole class: 25 students, 20 prompts a day, 180 school days, that is 90,000 prompts a year, around 21.6 kilowatt-hours. One LED bulb left on for ninety days straight.

Why the small number still gets big

The International Energy Agency works out in its 2025 report Energy and AI that data centres used around 415 terawatt-hours in 2024, about 1.5 percent of the world's electricity, and that by 2030 it will be around 945 terawatt-hours, slightly more than all of Japan uses today. That figure holds everything that lives in a data centre: streaming, online games, your school's cloud, and AI too. The same report applies the brakes: data centres make up only a tenth of the growth in global electricity demand to 2030, less than industrial motors, air conditioning or electric cars. Both things are true at once: per prompt it is almost nothing, added up it is a country.

Images are a different league. A study by Alexandra Sasha Luccioni, Yacine Jernite and Emma Strubell, presented at the ACM FAccT conference in 2024, measured open models and found that generating images uses on average over 60 times the energy of generating text: 2.9 kilowatt-hours per thousand images against 0.047 per thousand texts. That is about 2.9 watt-hours per image, a dozen text prompts at Google's figure, and the least efficient model in the study used half a phone charge per image. For video no company has published any figure, so all that can be said is the obvious: a video is many images per second.

The slide for your presentation

• One text prompt: between 0.24 watt-hours (Google, Gemini, median, 2025) and 0.34 (Altman, ChatGPT, average, 2025).
• In other words: a good second of microwave or a minute and a half of LED light.
• Water: five drops per prompt. A glass covers almost a thousand.
• CO2: 0.03 grams per prompt, 30 grams per thousand.
• One image: on average over 60 times the energy of a text (Luccioni et al., 2024, open models).
• All the world's data centres: 415 terawatt-hours in 2024, 1.5 percent of global electricity; around 945 in 2030, more than Japan today (IEA).
• But: only a tenth of the growth in consumption to 2030, less than motors, air conditioning or electric cars (IEA).
• The small numbers come from the companies. The big one comes from the IEA. Saying so is what separates a presentation from a rumour.

Three rules for using less

• Model: for a word or a formula, the fast, small model your app offers. The reasoning mode, which computes far longer for each answer, only when the task genuinely needs it.
• Length: one well-built question with all the context instead of five vague tries and three regenerations. And whatever a glance at the book settles does not get asked.
• Images: only the ones you will use. Three variants, not thirty, and video only for a real project. This is where one decision of yours outweighs a hundred text prompts.
• And if you want your own figure for the presentation: count your prompts for a week and multiply by 0.24 and by 0.34. Give both, the range is the honest answer.