How much energy and water does your AI use?
Pick your habits and see the yearly figures. Every number is a published range with a source, not a guess made up for a headline.
Why the same question has answers a thousand times apart
Ask how much water one AI prompt uses and you will find 0.26 millilitres and 500 millilitres. Neither is fabricated. The small figures count only the water evaporated on site to cool the servers that answered you. The large ones also count the water used to generate the electricity, and sometimes the water used to make the chips.
That is the whole argument, and almost every article skips it. Switching the third dropdown above changes the boundary, not the accuracy. When a source says 0.32 mL it is answering a narrower question than a source that says 10 mL.
“OpenAI ChatGPT, disclosed - direct cooling only (0.32 mL/query).”
“About 500 mL for 20 to 50 exchanges, roughly one bottle of water, once power-plant water use is included.”
What one message costs in energy
| Kind of message | Low, Wh | Median, Wh | High, Wh |
|---|---|---|---|
| Short question | 0.16 | 0.31 | 0.70 |
| Ordinary text answer | 0.50 | 1.00 | 2.00 |
| Reasoning or thinking mode | 2.00 | 4.00 | 7.00 |
| Long or agentic task | 10.00 | 20.00 | 40.00 |
Median is the number used in the calculator. The range is not noise; it is the spread between models and between short and long answers of the same kind.
“The new Joule paper estimates 0.31 Wh median for standard frontier-model queries, with an interquartile range of 0.16-0.60 Wh.”
The number everyone still quotes is out of date
The figure that made AI's energy use famous was 2.9 Wh per query, from 2024 analysis of GPT-3-era hardware. A 2026 study of production traffic puts the median near 0.31 Wh, roughly a tenth of it. If you have read that a prompt uses as much power as a light bulb left on for a few seconds, you are reading the old estimate.
“A single short GPT-4o query consumes 0.43 Wh ... o3 and DeepSeek-R1 emerge as the most energy-intensive models, consuming over 33 Wh per long prompt, more than 70 times the consumption of GPT-4.1 nano.”
Where the rest of the numbers come from
Carbon uses the US average of 0.37 kg CO2 per kWh, which is what you get from dividing published national emissions by national generation. Money uses the US residential average price of 18.17 cents per kWh. Both are in the method page with the arithmetic shown.