Skip to content
Prompt Corpus
Model behaviour19 August 20262 min read

The 3 watt-hour number was wrong by about ten times

The most-repeated figure about AI energy use turned out to be a large over-estimate. What the measurements actually say, and why the correction matters more than the number.

For a couple of years, the standard figure for the energy cost of one AI query was 3 watt-hours. It appeared in serious publications. It got repeated until it felt established.

Then people with access to the actual hardware measured it.

Google published a detailed measurement in 2025: the median text prompt to Gemini used 0.24 watt-hours. Independent estimates for ChatGPT land around 0.3 Wh, and a 2025 benchmarking study across models reported a median of 0.31 Wh.

Different labs, different methods, converging around a tenth of the number everyone was quoting.

What 0.24 Wh actually is

Numbers mean nothing without a comparison, so:

  • One prompt: 0.24 Wh
  • A microwave running for one second: about 0.3 Wh
  • Boiling a kettle: about 100 Wh, or 400 prompts
  • Charging a phone: about 15 Wh, or 60 prompts
  • One hot shower: about 2,000 Wh, or 8,000 prompts

If you sent 400 prompts today, you used roughly the electricity of making one cup of tea.

Where it stops being small

This is the part the reassuring version leaves out, and it is where the actionable information is.

That figure is for a typical text query. The spread is enormous:

  • Simple text prompt: ~0.2–0.6 Wh
  • Long context, whole documents in the prompt: several times more
  • Reasoning model on a hard problem: substantially more, because you pay for all the internal thinking
  • Agentic task looping with tools: several to tens of watt-hours

That is one to two orders of magnitude. So the honest statement is not "AI queries are cheap." It is a simple question is cheap and the expensive things are expensive — and the single most wasteful common habit is sending a trivial task to a reasoning model because that is what the interface had open.

Why the correction matters more than the number

The 3 Wh figure was a reasonable guess made without access to production hardware. It was repeated for years because it was the only number available, and repetition made it feel verified.

That is exactly the failure mode this whole subject warns about. Fluency and repetition are not evidence — and that applies to humans writing about AI at least as much as to AI writing about anything.

So the transferable habit is not "remember 0.24." It is: check the date, check whether it was measured or estimated, check which model and which task, and check what it is being compared to. A number without those is not information.

Including the numbers in this post. They will move.

The full picture, including the aggregate — which genuinely is large and growing — is in what one prompt actually costs and where the energy actually goes.

All notes