How much energy does an AI query use?
MethodPick a published figure, set how often you use AI, and see what it adds up to, with the source and scope of every number.
Is that a lot? At 20 uses a day, a year of this adds up to 2.5 kWh: as much electricity as an average US home uses in 2 hours, or an average EU home uses in 6 hours. One person's prompts are small; the electricity adds up in the buildings that answer everyone's.
Same provider, water: 0.32 mL per query (2.3 L/year at your usage).
Scope of this figure: Not specified by OpenAI (model, boundary and query type undisclosed). Stated by CEO Sam Altman in a blog post, without methodology; compared to 'what an oven would use in a little over one second, or a high-efficiency lightbulb would use in a couple of minutes'. Source: Sam Altman (OpenAI CEO) blog
Where does this energy go?
What is a data center?
A building full of computers (servers) that store data and run online services: websites, video, payments, AI. The machines draw a lot of electricity and give it back as heat that has to be removed, which is why where one is built matters for the power grid, for water and for the neighbours. All the terms used here →
See the data centers near you →Open the world map →Where cooling them competes for water →
See all 18 published figures and the reference points
Estimates differ by an order of magnitude because they measure different things: model size, prompt length, what hardware and overheads are counted, and the year. Compare scopes, not just numbers.
| Who | What | Value | Kind | Date | Source |
|---|---|---|---|---|---|
| Gemini Apps (median text prompt) | 0.24 Wh/prompt | provider disclosure | 2025-08 | Google Cloud Blog | |
| Gemini Apps (median text prompt) | 0.03 gCO2e/prompt | provider disclosure | 2025-08 | Google Cloud Blog | |
| Gemini Apps (median text prompt) | 0.26 mL/prompt | provider disclosure | 2025-08 | Google Cloud Blog | |
| Gemini Apps (median text prompt), narrower active-accelerator-only boundary | 0.1 Wh/prompt | provider disclosure | 2025-08 | Google Cloud Blog | |
| OpenAI | ChatGPT (average query) | 0.34 Wh/query | provider disclosure | 2025-06 | Sam Altman (OpenAI CEO) blog |
| OpenAI | ChatGPT (average query) | 0.32 mL/query | provider disclosure | 2025-06 | Sam Altman (OpenAI CEO) blog |
| Mistral AI | Le Chat (Mistral Large 2), 400-token response | 1.14 gCO2e/response | provider disclosure | 2025-07 | Mistral AI |
| Mistral AI | Le Chat (Mistral Large 2), 400-token response | 45 mL/response | provider disclosure | 2025-07 | Mistral AI |
| Epoch AI | ChatGPT (GPT-4o, typical query, ~500 output tokens) | 0.3 Wh/query | independent estimate | 2025-02 | Epoch AI |
| International Energy Agency | ChatGPT (average request) | 2.9 Wh/request | agency | 2024-01 | International Energy Agency |
| Microsoft Research | Frontier-scale LLM (>200B params) on H100 nodes, median query | 0.31 Wh/query | independent estimate | 2025-09 | arXiv (Microsoft) |
| Microsoft Research | Frontier-scale LLM, long reasoning (test-time scaling, ~15x longer queries), median | 3.91 Wh/query | independent estimate | 2025-09 | arXiv (Microsoft) |
| MIT Technology Review / University of Michigan ML.Energy | Llama 3.1 8B, one text response | 0.032 Wh/response | independent estimate | 2025-05 | MIT Technology Review |
| MIT Technology Review / University of Michigan ML.Energy | Llama 3.1 405B, one text response | 1.86 Wh/response | independent estimate | 2025-05 | MIT Technology Review |
| MIT Technology Review / University of Michigan ML.Energy | Stable Diffusion 3 Medium, one standard-quality image | 0.63 Wh/image | independent estimate | 2025-05 | MIT Technology Review |
| MIT Technology Review / University of Michigan ML.Energy | CogVideoX (newer, higher-quality version), one 5-second video | 944 Wh/video | independent estimate | 2025-05 | MIT Technology Review |
| Jegham et al. | Short LLM query (benchmark example) | 0.42 Wh/query | independent estimate | 2025-11 | arXiv |
| Jegham et al. | Most energy-intensive models, long prompt | 29 Wh/prompt | independent estimate | 2025-11 | arXiv |
Reference points
- One Google web search (Google, 2009): 0.3 Wh/search: Official Google Blog
- One Google web search (Google, 2009): 0.2 gCO2/search: Official Google Blog
- Typical Google search (IEA, 2024): 0.3 Wh/search: International Energy Agency
- LED bulb (~9 W, 1,100 lumens) lit for one hour: 9 Wh/hour: U.S. Department of Energy
- Watching TV for about nine seconds (Google's comparison): 0.24 Wh: Google Cloud Blog
- Average US household electricity use, one day: 29,564 Wh/day: U.S. Energy Information Administration
- Average EU household (dwelling) electricity use, one day: 9,863 Wh/day: ODYSSEE-MURE
Questions
How much energy does one AI prompt use?
Published figures for a typical text prompt range from about 0.24 Wh (Google Gemini, median, 2025-08) and 0.34 Wh (OpenAI ChatGPT, 2025-06) to several Wh for long reasoning queries or large open models. Image and video generation use far more.
Why do estimates differ so much?
They count different things: model size, prompt and answer length, which hardware and data-center overheads are included, and when the measurement was made.