AI Impact

How much electricity does one AI prompt use?

Between 0.16 and 0.70 watt-hours for an ordinary question, with a median near 0.31. A reasoning model working on a hard problem can use a hundred times that. The gap is the answer, not a problem with the answer.

Kind of messageLow, WhMedian, WhHigh, Wh
Short question0.160.310.70
Ordinary text answer0.501.002.00
Reasoning or thinking mode2.004.007.00
Long or agentic task10.0020.0040.00

“The new Joule paper estimates 0.31 Wh median for standard frontier-model queries, with an interquartile range of 0.16-0.60 Wh.”

What 0.31 Wh actually means

A watt-hour is a small unit, so the honest comparison is a small one. At the US residential average of 18.17 cents per kWh, one thousand short prompts cost about $0.06 in electricity. That is the number that makes the debate strange: a query is almost free, and a hundred million queries a day is a power station.

The interesting quantity is never the single prompt. It is prompts multiplied by users multiplied by the models those users prefer.

Why reasoning models break the comparison

A thinking model may generate thousands of internal tokens before it answers. Benchmarks put the most intensive models above 33 Wh for a long prompt, more than seventy times a small model doing the same short task. If you are comparing your own usage with a friend's, the model choice matters more than the number of messages.

“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.”

The 2.9 Wh figure, and why it will not die

It came from 2024 estimates on older infrastructure, and it is repeated in thousands of articles. The 2026 production measurements are roughly ten times lower for ordinary text. Both numbers get published side by side, usually without the note that one measures a system that no longer exists.

Put your own usage into the calculator or see the water side.