In short
A typical short text prompt to a chatbot such as ChatGPT or Gemini uses roughly 0.2 to 0.4 watt-hours of electricity, by the providers' own figures and independent estimates: about as much as a TV uses in a few seconds. AI energy use rises steeply for long inputs, reasoning models, images, video and coding agents, which can use tens to thousands of times more per task. In total, data centres used about 485 TWh of electricity in 2025, according to the IEA, and AI is the main reason that figure is growing.
Key points
- Google reports 0.24 Wh for a median Gemini text prompt in May 2025; OpenAI's chief executive gave about 0.34 Wh for an average ChatGPT query. Both are self-reported.
- Epoch AI estimates a typical GPT-4o query at about 0.3 Wh, rising to about 2.5 Wh with a 10,000-token input and almost 40 Wh with a 100,000-token input.
- Hugging Face's AI Energy Score tests found reasoning models used on average 30 times more energy than standard models.
- In tests of open models, generating an image used about 2.9 Wh on average, and one short AI video about 90 Wh.
- One analysis of heavy use of an AI coding agent estimated about 150 Wh per typed prompt, roughly 600 times a median chat prompt.
- The IEA estimates data centres used about 485 TWh in 2025, up 17% in a year, and expects around 950 TWh by 2030.
How much energy does AI use? For one short question to a chatbot, very little: a fraction of a watt-hour. For long documents, reasoning modes, images, video and coding agents, far more. And across billions of daily requests, AI energy consumption is now large enough to shape electricity planning. This page brings together the best public figures, says where each comes from, and explains why they differ.
How much energy does a ChatGPT or Gemini prompt use?
A short text prompt to a large chatbot uses roughly 0.2 to 0.4 watt-hours (Wh), on the best current figures. A watt-hour is the energy a 1-watt device uses in an hour.
The most quoted numbers come from the providers themselves:
- Google reported in 2025 that a median Gemini Apps text prompt used 0.24 Wh. Its paper compares that with watching television for less than nine seconds.
- OpenAI's chief executive wrote in 2025 that an average ChatGPT query uses about 0.34 Wh. No method was published.
- Epoch AI, an independent research group, estimated about 0.3 Wh for a typical GPT-4o query, working from public information about hardware and model size.
These are close, but they are not like-for-like: a median is not an average, and each draws the boundary of what is counted in a different place. They also cover short text prompts only.
Energy per task: the figures compared
Energy per AI task spans more than a thousandfold, from under a watt-hour for a short prompt to hundreds of watt-hours for an agent session. The table gives the main published figures. Self-reported figures come from the company running the service; estimates and measurements come from independent researchers.
| Task | Energy | Source and year | Type |
|---|---|---|---|
| Median Gemini text prompt | 0.24 Wh | Google, 2025 | Self-reported, published method |
| Average ChatGPT query | about 0.34 Wh | OpenAI chief executive, 2025 | Self-reported, no method |
| Typical GPT-4o query | about 0.3 Wh | Epoch AI, 2025 | Independent estimate |
| GPT-4o, short prompt (100 tokens in, 300 out) | 0.42 Wh | Jegham et al., 2025 | Independent estimate |
| GPT-4o-class query with a 10,000-token input | about 2.5 Wh | Epoch AI, 2025 | Independent estimate |
| GPT-4o-class query with a 100,000-token input | almost 40 Wh | Epoch AI, 2025 | Independent estimate |
| Long prompt (10,000 in, 1,500 out): Llama 3.1 8B vs DeepSeek-R1 | about 0.44 Wh vs about 29 Wh | Jegham et al., 2025 | Independent estimate |
| Reasoning vs standard models | 30 times more on average | Hugging Face AI Energy Score, 2025 | Measured, open models |
| One generated image | about 2.9 Wh (2.9 kWh per 1,000) | Luccioni, Jernite and Strubell, 2024 | Measured, open models |
| One short generated video | about 90 Wh for one model; 0.14 Wh to over 415 Wh across models | Delavande, Pierrard and Luccioni, 2025 | Measured, open models |
| Median coding-agent session | about 41 Wh | Couch, 2026 | Independent estimate |
| Coding agent, per typed prompt (heavy use) | about 150 Wh (range 60–290 Wh) | Hausfather, 2026 | Independent estimate |
A token is a chunk of text, roughly three-quarters of an English word. The figures for images and video come from open models run on research hardware; commercial services may be more or less efficient, and most do not publish figures.
Why do reasoning models, images and agents use more?
Because they do much more computation per request. Energy rises with the number of tokens a model reads and writes, and with the size of the model.
- Reasoning models write long hidden chains of working before they answer. Hugging Face's AI Energy Score team found that, run with reasoning switched on, the same model used between 150 and 700 times more energy than with it off, and wrote 300 to 800 times more tokens.
- Long inputs must be processed in full. Epoch AI's estimates rise from about 0.3 Wh to almost 40 Wh as the input grows to 100,000 tokens, the size of a very long pasted document.
- Images and video are generated in many steps over many pixels. Delavande, Pierrard and Luccioni found video energy grew quadratically with resolution and length, and put one short clip at roughly 2,000 times a text answer.
- Agents call the model again and again, re-reading their growing context each time. Anthropic reported that its agents typically used about four times the tokens of a chat, and multi-agent systems about fifteen times.
How much energy does an AI coding agent use?
Estimates run from tens of watt-hours for a typical session to over a kilowatt-hour for a heavy day. No AI company publishes energy figures for its coding agents, so the best available numbers are careful independent estimates from token counts.
Simon Couch estimated a median Claude Code session, with 24 model calls and about 590,000 tokens, at about 41 Wh. Climate scientist Zeke Hausfather logged eight weeks of his own heavy use: 1,138 typed prompts set off more than 14,000 model calls and 3.2 billion tokens, 96% of them re-reads of cached context. He estimated about 170 kWh in total (range 70–330 kWh), about 150 Wh per prompt, and a median session of about 0.6 kWh, which he compares to charging a phone about 50 times. See agentic AI and token waste.
Is training or inference the bigger cost?
For a widely used model, running it (inference) usually outweighs training it. Training is a large one-off cost; inference is small per request but repeated billions of times.
Published training figures show the scale:
- GPT-3 (2020): Patterson and colleagues at Google estimated training used 1,287 MWh of electricity.
- Llama 3.1 (2024): Meta's model card reports 39.3 million GPU-hours on H100 chips for its three model sizes, of which 30.84 million were for the largest, 405-billion-parameter model.
On the other side, Patterson and colleagues cite NVIDIA estimates that 80–90% of machine learning workload is inference, and Amazon Web Services' claim that 90% of machine learning demand in its cloud is inference. That balance is why falling energy per prompt matters so much, and also why it is not enough on its own: use is growing faster. See efficiency is not sufficiency.
How much electricity do data centres and AI use in total?
The International Energy Agency (IEA) estimates that data centres used about 485 TWh in 2025, and expects around 950 TWh by 2030, about 3% of global electricity. A terawatt-hour (TWh) is a billion kilowatt-hours.
- The IEA put data-centre use at about 415 TWh in 2024, around 1.5% of the world's electricity. Its April 2026 update reported 485 TWh in 2025, up 17% in a year, with electricity used by AI-focused data centres up 50%.
- It expects the United States and China to account for nearly 80% of growth to 2030, and accelerated servers, mainly for AI, to drive almost half of the increase.
- In the United States, Lawrence Berkeley National Laboratory estimated data centres used about 178 TWh in 2023, 4.4% of national electricity, and could reach 6.7–12% by 2028.
There is no reliable global figure for AI alone, because data centres do not separately meter AI work. Company reports show the trend: Microsoft's electricity use rose from about 10.8 TWh in its 2020 financial year to about 37.0 TWh in 2025, and Google reported a 37% rise in electricity demand in 2025, the largest in its history. See why green AI matters now.
Why do AI energy figures vary so much?
Because studies count different things, on different hardware, for different models and tasks. Four factors explain most of the spread.
- System boundary. Google's own paper shows that counting only the active AI chips gives 0.10 Wh per median prompt; adding host processors and memory, idle capacity kept ready for demand, and data-centre cooling gives 0.24 Wh, more than twice as much. See measurement is the bottleneck.
- Hardware and utilisation. Newer chips do more work per watt, and a busy server shares its overheads across more requests than a quiet one.
- Model size. In Jegham and colleagues' estimates, a long prompt cost about 0.44 Wh on the small Llama 3.1 8B model and about 29 Wh on DeepSeek-R1, a large reasoning model.
- Tokens. Longer inputs, longer answers, hidden reasoning and repeated agent steps all add tokens, and energy rises with them.
Two further cautions. Figures date quickly: Google reported a 33-fold fall in energy per median prompt over one year. And self-reported figures cannot be checked independently. When you see a number, ask what was measured, on what boundary, and when. See how to check a claim about AI energy use.
How does AI energy compare with everyday things?
A short text prompt is small next to most household uses of electricity; heavy agent or video use is not. These comparisons are the sources' own:
- Google compares its 0.24 Wh median Gemini prompt to watching television for less than nine seconds.
- Hausfather compares a median coding-agent session in his heavy use (about 0.6 kWh) to charging a phone about 50 times, and an average day of that use (about 3 kWh) to running more than two refrigerators.
- Couch compares a median day of his coding-agent use (about 1,300 Wh) to running a dishwasher an extra time each day.
- Google said in 2009 that an average web search used about 0.3 Wh, a similar order to today's short text prompts. The figures are far apart in time and method.
The point is not that any one request is large. It is that habits matter: choosing a reasoning model, pasting long documents, generating batches of images or leaving an agent running can each multiply the energy of a task many times over. See put my AI energy use in perspective.
What can you do to cut AI energy use?
Use the smallest tool that does the job, and ask less of it. The largest savings come from avoiding needless work, not from switching providers.
- Decide whether you need AI for the task: see decide if I need AI for a task.
- Pick a smaller or faster model for everyday questions: see choose a smaller AI model and choosing the right model.
- Switch reasoning off when you don't need it: see use reasoning modes only when needed.
- Keep conversations short and start a new chat for a new topic: see keep long AI chats efficient.
- Generate fewer images and videos: see cut the impact of AI image and video generation.
- Set limits on agents: see stop AI agents wasting tokens and use an AI coding assistant with a smaller footprint.
For organisations, the starting point is measurement: see measuring AI energy, estimate the energy of AI API calls and using AI efficiently.
What to do next
- Read what is AI's carbon footprint? to turn energy into emissions.
- Read how much water does AI use? for the water side.
- Try reduce the energy of my ChatGPT use for everyday habits.
- Ask your AI provider for per-task energy figures: see ask an AI company for energy figures.
- Compare open models on measured energy with the AI Energy Score.
Frequently asked questions
How much energy does a ChatGPT prompt use?
OpenAI's chief executive said in 2025 that an average ChatGPT query uses about 0.34 watt-hours, and Epoch AI independently estimated about 0.3 Wh for a typical GPT-4o query. Long inputs, reasoning modes and image generation use much more.
How much energy does AI use per prompt?
For a short text prompt, published figures cluster around 0.2 to 0.4 Wh: Google reports 0.24 Wh for a median Gemini prompt. Independent estimates for large models with long prompts reach about 29 Wh, and agentic coding sessions can use hundreds of watt-hours.
Does AI use more energy than Google search?
Not for a short text prompt, on the published figures. Google said in 2009 that an average search used about 0.3 Wh; it now reports 0.24 Wh for a median Gemini text prompt. The figures are 16 years apart and measured differently, so treat this as a rough guide. Long, reasoning, image and agent tasks use far more.
How much electricity does AI use in total?
There is no reliable figure for AI alone. The International Energy Agency estimates all data centres used about 485 TWh in 2025, up 17% in a year, that AI-focused data centres' use grew 50%, and that the total will reach around 950 TWh, about 3% of global electricity, by 2030.
Is training or using AI more energy-intensive?
Training one large model is a big one-off cost, such as an estimated 1,287 MWh for GPT-3. But a popular model is used billions of times, and industry estimates cited by Google researchers put inference at 80–90% of machine learning workloads.
Why do AI energy estimates vary so much?
Because they count different things. Including data-centre cooling, idle servers and host computers more than doubled Google's own figure, from 0.10 Wh to 0.24 Wh. Model size, hardware, prompt length and the number of tokens generated also change the answer.
How much energy does AI image generation use?
In tests of open models by Luccioni, Jernite and Strubell, generating images used about 2.9 kWh per 1,000 images on average, or about 2.9 Wh each, more than a thousand times as much as classifying a piece of text.
How can I reduce the energy my AI use takes?
Use a smaller or faster model for simple tasks, keep prompts and conversations short, switch reasoning modes off when you don't need them, and generate fewer images and videos. Asking whether you need AI for the task at all saves the most.
Sources
- Measuring the environmental impact of delivering AI at Google scale, Elsworth et al., Google (2025)
- The gentle singularity, Sam Altman (OpenAI) (2025)
- How much energy does ChatGPT use?, Epoch AI (2025)
- How hungry is AI? Benchmarking energy, water and carbon footprint of LLM inference, Jegham et al. (arXiv 2505.09598, v6) (2025)
- AI Energy Score v2: refreshed leaderboard, now with reasoning, Luccioni and Gamazaychikov, Hugging Face (2025)
- Power hungry processing: watts driving the cost of AI deployment?, Luccioni, Jernite and Strubell, ACM FAccT 2024 (2024)
- Video killed the energy budget: characterizing the latency and power regimes of open text-to-video models, Delavande, Pierrard and Luccioni (arXiv 2509.19222) (2025)
- The real energy use of agentic AI, Zeke Hausfather, The Climate Brink (2026)
- Electricity use of AI coding agents, Simon P. Couch (2026)
- How we built our multi-agent research system, Anthropic (2025)
- Carbon emissions and large neural network training, Patterson et al. (arXiv 2104.10350) (2021)
- Powering a Google search (Italian edition: Dare energia alle ricerche su Google), Google (2009)
- Llama 3.1 8B model card, Meta (2024)
- Energy and AI, International Energy Agency (2025)
- Key questions on energy and AI, International Energy Agency (2026)
- 2024 United States Data Center Energy Usage Report, Lawrence Berkeley National Laboratory (2024)
- 2026 environmental data fact sheet, Microsoft (2026)
- Read our 11th annual Environmental Report, Google (2026)
Related
What is AI's carbon footprint?
AI carbon footprint explained: emissions per prompt, training emissions from model cards, embodied carbon, grid effects and big tech emissions trends.
How much water does AI use?
How much water does AI use? AI water use per prompt and per model, data centre water, WUE, company figures, local water stress and what is still unknown.
Measurement is the bottleneck for green AI
Why AI energy figures are hard to trust or compare: self-reported numbers, estimates not meters, different boundaries, and what good looks like.
How to measure the energy and carbon of AI
A practical guide to metering or estimating the energy and carbon of AI work: tools, system boundaries, per-task metrics and honest uncertainty.