A single typical AI text answer uses a small amount of electricity. Google measured its own median text request to the Gemini assistant at 0.24 Wh, which it compares to watching about nine seconds of television. The large numbers come from scale: billions of requests a day, the training of new models, and all the other computing that data centres do. Data centres use about 1.5% of the world's electricity today, and that share is expected to more than double by 2030.
One request, measured
In August 2025 Google researchers published a measurement of the median Gemini Apps text prompt. Of its 0.24 Wh, they attribute the largest part to the AI chips themselves, with the rest going to the host computer, machines kept ready but idle, and the data centre's cooling and power overheads.
| Part of the system | Energy per median prompt | Share |
|---|---|---|
| AI accelerators, while working | 0.14 Wh | about 58% |
| Host processor and memory | 0.06 Wh | about 25% |
| Idle machines kept ready | 0.02 Wh | about 10% |
| Data centre overhead, such as cooling | 0.02 Wh | about 8% |
The paper notes that counting the AI chips alone would give only 0.10 Wh, which is one reason published estimates differ. It does not include training the model, networks outside the data centre or the user's own device. It also reports a 33x reduction in energy per median prompt over one year, from May 2024 to May 2025.
One country: the United States
Lawrence Berkeley National Laboratory's June 2026 report estimates that US data centres used 192 TWh in 2024, 4.7% of all US electricity consumption, a 14% increase on 2023. Its reference case for 2030 is 649 TWh, within a range of 521 to 843 TWh, and it projects that by 2030 AI servers account for 55% of total data centre energy use.
The world
The International Energy Agency puts data centres' current share of global electricity consumption at 1.5%, projected to more than double by 2030 to around 945 TWh, or approximately 3% of global electricity demand.
What is not certain
- The Google figure is one company's measurement of its own service, in a paper by its own researchers that is not peer reviewed. Other services, models and tasks such as images or video use different amounts.
- National and global figures are model estimates; projections to 2030 depend on how AI use and efficiency develop.
- Training a new model uses a large amount of electricity once; it is not included in per-request figures.
Sources and method
- Elsworth and others, Google, Measuring the environmental impact of delivering AI at Google Scale, arXiv:2508.15734, 21 August 2025, operator-authored preprint. https://arxiv.org/abs/2508.15734
- Smith and others, Lawrence Berkeley National Laboratory, United States Data Center Energy Usage Report: 2025 Update, June 2026. https://escholarship.org/uc/item/33m6w3x0 ; full text https://escholarship.org/content/qt33m6w3x0/qt33m6w3x0.pdf
- International Energy Agency with the Korea Energy Economics Institute, Energy and AI in East Asia, revised April 2026. https://iea.blob.core.windows.net/assets/431d7d22-039e-46d3-9832-090a8b4075fc/EnergyandAIinEastAsia.pdf
Method: each link was opened on 29 October 2026 and every figure was taken from the page cited beside it. No figure was estimated.
Read next
This article provides general information, not investment advice or an offer of securities.