Agents, long documents and video can use 10× to 1,000× more energy than a short AI prompt
A short prompt uses about 0.3 Wh, a very long document about 40 Wh, an agent task about 50 Wh. New data centres are built for heavy uses.
Energy: AI requests vs everyday things
Watt-hours. Log scale: each gridline is 10× the previous. Blue = AI.
What this measures: How much electricity different AI requests use, next to everyday activities.
AI figures are company disclosures (Google, OpenAI) or published models (Microsoft-led Joule study, Epoch AI, IEA); none is independently audited. Household figures assume typical UK appliances (Hannah Ritchie). The web-search figure is Google's from 2009.
Show data table and sources
| Item | Value | Note |
|---|---|---|
| One AI text prompt (Gemini median) | 0.24 Wh | Google, all overheads |
| One Google web search (2009 figure) | 0.3 Wh | old figure |
| Average ChatGPT query | 0.34 Wh | OpenAI, method unstated |
| Long reasoning-style AI query | 3.91 Wh | modelled median |
| LED bulb for 1 hour | 10 Wh | |
| Microwave for 1 minute | 16.7 Wh | |
| Charging a smartphone | 19.4 Wh | |
| Laptop for 1 hour | 20 Wh | |
| AI query over a 100,000-token document | 40 Wh | estimate, paid once per upload |
| Agentic AI task with reasoning | 50 Wh | IEA, GPU-only |
| TV for 1 hour | 60 Wh | |
| Streaming video for 1 hour | 77 Wh | 2019, incl. device |
| Boiling a full kettle | 100 Wh |
- How much energy does Google's AI use? We did the math (Google Cloud)
- The Gentle Singularity (Sam Altman)
- Energy Use of AI Inference, Efficiency Pathways, and Test-Time Scaling (Microsoft et al. (Joule))
- How much energy does ChatGPT use? (Epoch AI)
- Key Questions on Energy and AI (IEA)
- Energy use comparisons (Hannah Ritchie)
- Greenhouse gas equivalencies calculator: calculations and references (US EPA)
- Factcheck: What is the carbon footprint of streaming video on Netflix? (Carbon Brief)
- Green-minded Google gets red-faced over search energy consumption claims (quotes Google's 2009 figure) (eWeek)
The point that a prompt uses almost nothing holds for short text questions. It holds less well for how AI is increasingly used.
Long documents. Epoch AI estimates about 2.5 Wh for a query with 10,000 input tokens (word pieces) and almost 40 Wh with 100,000. That cost is paid once per uploaded document, not for each follow-up question. The Microsoft-led Joule study models queries about 15× longer than typical at a median of 3.91 Wh.
Reasoning and agents. Reasoning models work through a problem step by step before answering. Agents carry out multi-step tasks on their own. The IEA’s rough figures, counting only the AI chips (GPUs), are 0.05 Wh for a medium model, 0.31 Wh for a large model, 7.6 Wh for a reasoning model, 1.1 Wh for an agent task and about 50 Wh for an agent task with reasoning. Video generation can take hundreds to thousands of times more energy.
For scale. The IEA calculates that 10 billion queries a day at a generous 1 Wh each would be about 3.6 TWh a year, under 1% of today’s data-centre use. It concludes that most planned data-centre capacity is for training, video and agents rather than chat.
What it means
Whether AI’s energy use is small depends on what it is doing. A quick chat question uses very little; long documents, reasoning, agents and video use tens to thousands of times more, and that is where data-centre building is aimed. When someone quotes an energy figure for “an AI query”, check which kind they mean.
Caveats
- The IEA figures count AI chips only, from standard tests of open-weight models rather than optimised production systems.
- “Agentic” covers very different designs; agents can run for long periods without a person waiting.
- Energy per task has been falling by at least an order of magnitude a year, per the IEA.