How much energy does one AI prompt use? The only measured figure comes from Google, and it is 0.24 watt-hours

In August 2025 Google published a technical paper measuring its own production systems and reported that the **median Gemini text prompt consumes 0.24 watt-hours of energy** — along with 0.26 millilitres of water.

That is the figure to start from, with two conditions attached immediately.

It is a **median**, not an average and not a maximum. Prompt cost varies enormously with the length of the output, whether the model reasons before answering, and whether images or video are involved. A median text prompt is the middle of a distribution with a long tail to the right.

And it is **Google measuring Google**. It covers Gemini on Google's own hardware in Google's own datacentres. OpenAI has published no comparable figure for ChatGPT, which means anyone quoting a per-prompt energy cost for ChatGPT specifically is extrapolating from someone else's infrastructure.

To make 0.24 watt-hours concrete: it is roughly what a 60-watt bulb draws in about fifteen seconds. Running a thousand such prompts consumes about 0.24 kilowatt-hours — less than a typical clothes dryer cycle.

Which is why the per-prompt number, though it is the one everyone asks for, is not where the significance sits.

The aggregate is a different object. The International Energy Agency puts total data centre electricity demand at roughly **460 to 490 terawatt-hours in 2025**, growing about 17 per cent that year, with consumption at AI-focused facilities specifically rising around 50 per cent. Data centres overall account for just over **1 per cent of global electricity** and about **0.5 per cent of global CO2 emissions**.

In the United States the share is higher: around **4 per cent of national electricity in 2023**, with Lawrence Berkeley National Laboratory projecting **7 to 12 per cent by 2028**.

So both of these are true at once. A single prompt is trivial. The aggregate is a measurable and fast-growing share of national electricity. There is no contradiction — it is what happens when a very small unit cost is multiplied by an extremely large and rising number of units, and concentrated into a small number of physical locations.

That concentration is the part per-prompt arithmetic hides completely. A national percentage says nothing about which grid absorbs the load. The IEA estimates Irish data centres could reach around a third of that country's electricity, which is a different problem entirely from 1 per cent globally.

Two things remain genuinely unresolved. Efficiency is improving quickly and demand is growing quickly, and which wins over a decade is not established by any of the figures above. And no independent party can verify Google's measurement — it is a company disclosure about its own systems, which does not make it wrong but does make it something other than an independent finding.