How much energy does AI use?
Energy and water per task for common AI uses, from a chatbot prompt to an AI video clip, with household appliances for scale. Every figure has a low, central, and high case from published measurements or labeled estimates.
Energy and water per task
AI energy is usually measured at the data center: the servers, cooling, and power delivery behind a task. These figures leave out your phone or laptop and the network. To compare like with like, the last column uses only the data-center part of a year of everyday tech.
| Task | Energy per task, Wh | Water per task, mL | Tasks that equal your year's data-center use |
|---|
Water per task counts cooling and power-plant water. Water, data centers, and AI sets these volumes beside U.S. water use, reservoir evaporation, lawns, and homes, with what each figure measures.
Household reference points
What the figures count
- AI figures are data-center energy per task. The comparison with your year uses the same boundary: the data-center part of your apps, streaming, and cloud services.
- Devices at home, such as the router, the TV, and the smart speaker, have no data-center layer, so they're only in the full stack view.
- The AI video estimate on the results page, other people's AI video generation attributed to your viewing, is left out of both sides.
- Water is data-center cooling plus the water power plants use to make the electricity. Only one figure, Google's 0.26 mL for a median prompt, was disclosed by a company. The rest are derived.
Side by side: data center only and full stackdevices and networks added to both sides
Per-task details, facts, and research gaps are in the methodology under AI task details. Contested figures, such as the widely repeated 3 Wh per prompt, are under Contested figures.
Sources for this page
Every source on the site is on the Sources page, with its URL, date, kind, and what uses it.