0.3 watt-hours. That is roughly what a simple question to a language model uses, according to current estimates. About an LED bulb running for two minutes. Even so, the electricity used by data centers is set to double by 2030. Both are true.
A hall that heats
At its core, an AI data center is a hall full of specialized chips. They compute, get hot and have to be cooled. The International Energy Agency (IEA) compares the heat of a single AI server rack to thirty gas boilers. The power density of these servers increased elevenfold between 2020 and 2025.
The single question is small
In 2025, Epoch AI estimated a typical GPT-4o query at about 0.3 watt-hours. Google puts an average text prompt to Gemini at 0.24 watt-hours. One is an outside estimate, the other a figure from the company itself. The often-quoted 3 watt-hours come from a 2023 calculation and are about ten times too high.
The task makes the difference
It only stays small for a simple text query. Generating a video, letting a reasoning model think at length or letting an agent work on its own can take a hundred to a thousand times more, according to the IEA. Very long inputs also drive consumption up: for a document of 100,000 tokens, Epoch AI estimates almost 40 watt-hours.
The sum is large
Worldwide, data centers used about 485 terawatt-hours of electricity in 2025. Consumption of AI data centers grew by half in that single year. By 2030, the IEA expects it to double to about 950 terawatt-hours, around three percent of global electricity.
Power for AI in three orders of magnitude
From the single question to the sum
- One simple text query: about 0.3 Wh, roughly an LED bulb for two minutes
- Video, reasoning model, agent: up to 100 to 1,000 times more
- All data centers worldwide: 485 TWh (2025), projection for 2030 about 950 TWh
Sources: IEA 2026, Epoch AI 2025, Google 2025
And Switzerland?
In 2024, about 3.6 percent of Swiss electricity went to data centers, around 6 percent in the canton of Zurich. According to the Swiss Federal Office of Energy, AI still plays a minor part. Large models are rarely trained in Switzerland; consumption has grown mainly because companies move into cloud and colocation data centers. By 2030, the office expects a share of 4.3 to 5.6 percent.
Two questions that matter
Because the single question is small and the sum is large, two things are worth a look: where does the power come from? And is the waste heat used? For everyday use, one simple rule is enough: a simple question does not need a reasoning model.
The single question is small. The sum is large.
Frequently asked questions
What about water?
Data centers use water for cooling. Google puts an average Gemini text prompt at 0.26 millilitres, about five drops. That is a figure from the company itself. Global figures vary widely depending on what is counted, which is why the video does not use one.
Why are the 3 watt-hours per question out of date?
The figure comes from an estimate made in 2023. In 2025, Epoch AI put a typical GPT-4o query at about 0.3 watt-hours, roughly ten times less. Google reports 33 times less energy per Gemini prompt within one year. Models and chips have become more efficient.
Sources (5)
- IEA: Key Questions on Energy and AI, Executive Summary (2026)
- IEA: Energy and AI, Executive Summary (2025)
- Elsworth et al. (Google): Measuring the environmental impact of delivering AI at Google scale, arXiv (2025)
- You (Epoch AI): How much energy does ChatGPT use? (2025)
- Swiss Federal Office of Energy (SFOE): Electricity consumption of data centres in Switzerland, summary report (2026, German)
