How Much Electricity and Water Does ChatGPT Use? A Review of the Evidence, and How India Will Power AI
Short answer: the best published estimates put a typical ChatGPT or Gemini text prompt at about 0.24 to 0.34 watt-hours of electricity and well under 1 millilitre of water for cooling. That is far less than the 3 Wh figure that circulated in 2023, but the number rises sharply for long documents and AI agents, and broader accounting methods that include power-plant water give much higher water figures.
Individually, a prompt is tiny. Collectively, AI is reshaping electricity demand. The International Energy Agency estimates data centres used about 415 TWh in 2024 and could reach about 945 TWh by 2030. In India, data centre electricity use is projected to rise from about 10 to 15 TWh in 2025 to 40 to 45 TWh by 2030, and the government has opened nuclear power to private companies partly to meet that demand.
We reviewed 18 public sources, including company disclosures from OpenAI, Google and Mistral, independent estimates from Epoch AI and university researchers, the IEA's Energy and AI report and Indian government statements, to answer three questions: how much energy and water one AI prompt uses, what that means at India's scale, and where India's AI power will come from.
Key findings
- Per-prompt energy estimates have converged near 0.3 Wh. Sam Altman put the average ChatGPT query at about 0.34 Wh. Google measured the median Gemini text prompt at 0.24 Wh. Epoch AI estimated about 0.3 Wh for a typical GPT-4o query.
- The old 3 Wh figure was based on older assumptions. Epoch AI attributes the tenfold gap to longer assumed outputs, older A100 hardware and a larger active model in the 2023 estimate.
- Long inputs cost far more. Epoch AI estimates about 2.5 Wh for a 10,000-token input and about 40 Wh for a 100,000-token input.
- Water depends on what you count. Google reports 0.26 ml per Gemini prompt for data centre cooling. University of California researchers estimate about 17 ml per request in the US when water used by power plants is included. Mistral's full life-cycle analysis gives 45 ml for a 400-token response.
- India is betting on nuclear for AI. The SHANTI Act opened nuclear power to private companies in December 2025, and the IT Minister has called small modular reactors a practical way to power data centres.
How we reviewed the evidence
This is a review of published estimates, not a measurement. We grouped sources into company disclosures (OpenAI, Google, Mistral), independent estimates (Epoch AI, academic papers), energy-system analysis (IEA) and Indian policy sources. Company figures were not independently audited, and each source uses different boundaries, so we explain what each number includes before comparing them.
What happens to energy and water when you send a prompt
Energy is used at several stages. The GPU does the computation. The rest of the server, networking and the data centre itself add overhead. Cooling removes the heat, sometimes by evaporating water. And the power plant that generates the electricity uses water and emits carbon too. Estimates differ mainly because they draw the boundary at different points along this chain.
How much electricity does one ChatGPT prompt use?
| Source | Estimate per prompt | What it covers |
|---|---|---|
| Alex de Vries (2023), as cited by Epoch AI | About 3 Wh | Older estimate, long outputs, A100 GPUs |
| Epoch AI (February 2025) | About 0.3 Wh | Typical GPT-4o query, 500 output tokens, H100 GPUs |
| Sam Altman, OpenAI (June 2025) | About 0.34 Wh | Average ChatGPT query, no methodology published |
| Google (August 2025) | 0.24 Wh | Median Gemini text prompt, full serving overhead |
| Google, narrower method (August 2025) | 0.10 Wh | Same prompt, accelerator energy only |
Altman described 0.34 Wh as "about what an oven would use in a little over one second". Google's figure is notable for showing how much the boundary matters: the same median prompt is 0.10 Wh if only the AI chips are counted and 0.24 Wh when idle machines, cooling and other data centre overhead are included, according to CO/AI's summary of Google's report. Google also reported that the energy per median prompt fell 33 times over a 12-month period, and its carbon footprint fell 44 times.
Two caveats from the sources matter. Neither OpenAI's nor Google's figure was verified by independent third parties, and Implicator reports that outside researchers argued Google's accounting is incomplete. And Epoch AI notes that, under different assumptions, a single query could plausibly range from about 0.1 Wh to 4 Wh.
Long documents and agents change the picture
Epoch AI's estimate assumes a short question and a page-long answer. Processing long inputs costs much more:
| Input size | Estimated energy | Compared with a typical query |
|---|---|---|
| Typical short prompt | About 0.3 Wh | 1x |
| 10,000 tokens (a short paper) | About 2.5 Wh | About 8x |
| 100,000 tokens (about 200 pages) | About 40 Wh | About 130x |
Epoch AI notes that this input cost is paid once when a document is first processed, not on every follow-up message. But coding agents and research agents that repeatedly send large contexts can use many times the energy of a simple chat, which is also why they cost more in API fees.
How much water does one AI prompt use?
| Source | Water per prompt | What it counts |
|---|---|---|
| Google (August 2025) | 0.26 ml, about five drops | Data centre cooling for a median Gemini prompt |
| Sam Altman (June 2025) | About 0.32 ml (0.000085 gallons) | Average ChatGPT query, no methodology published |
| Li, Yang, Islam and Ren (UC Riverside) | About 17 ml (US average) | Cooling plus water used to generate the electricity |
| Mistral (July 2025) | 45 ml per 400-token response | Full life cycle, including hardware and training share |
The New Stack describes Google's 0.26 ml as about five drops of water. The gap between 0.26 ml and 45 ml is mostly about scope, not disagreement. The UC Riverside paper "Making AI Less Thirsty" finds that in the US, about 2.2 ml per request is evaporated in data centre cooling, while about 14.7 ml is consumed by the power plants supplying the electricity. The same paper says about 500 ml of water is consumed for roughly 10 to 50 medium-length GPT-3 responses, depending on where and when the model runs, and projects global AI water withdrawal of 4.2 to 6.6 billion cubic metres in 2027.
Mistral's July 2025 life-cycle assessment, summarised by The Decoder, adds 1.14 g of CO2 equivalent per 400-token response and attributes 85.5% of its model's lifetime emissions and 91% of its water use to training rather than everyday use.
What this means at India's scale
India is now one of the largest AI markets. Sam Altman said in February 2026 that India had about 100 million weekly ChatGPT users. To show how per-prompt figures add up, here is an illustrative calculation. It is our arithmetic on the published figures, not a measurement, and it assumes all prompts are served at the per-prompt rates above, wherever the data centre is.
| Assumption or result | Value |
|---|---|
| Prompts per day (100 million users, 10 prompts each) | 1 billion |
| Electricity at 0.34 Wh per prompt | About 340 MWh a day, or about 0.12 TWh a year |
| Share of India's estimated 2025 data centre use (10 to 15 TWh) | Roughly 1% |
| Cooling water at about 0.3 ml per prompt | About 300,000 litres a day |
| Water at 17 ml per prompt, including power plants | About 17 million litres a day |
Two conclusions follow. First, everyday chatbot use is a small slice of total data centre demand. Second, the method matters enormously for water: counting power-plant water raises the estimate about fiftyfold. Training new models, long-context work and AI agents add further demand that this simple calculation does not capture.
Water is a local issue as well as a global one. Down To Earth reported in November 2025 that a data centre can use about 26 million litres of water per megawatt of capacity a year, and that many Indian data centres are being built in water-stressed regions.
How fast is AI electricity demand growing?
The IEA's Energy and AI report, as summarised by Down To Earth in April 2025, found that:
- Data centres used about 415 TWh in 2024, about 1.5% of global electricity, with demand growing about 12% a year over five years.
- Use could reach about 945 TWh by 2030, slightly more than Japan's total electricity use today.
- A single AI-focused data centre can exceed 100 MW, using as much electricity as about 100,000 households.
- Data centres currently draw about 30% of their power from coal, 27% from renewables, 26% from natural gas and 15% from nuclear.
In India, PTI reported in 2026, citing government goals, that data centre electricity use could rise from about 10 to 15 TWh in 2025 to 40 to 45 TWh by 2030, as capacity grows towards 10 GW. We covered the capacity build-out in India's AI Data Centres in 2026.
Where will India get the power for AI? The nuclear bet
India's policy response has focused on clean, round-the-clock power, and nuclear has become central to it:
- February 2025: Nuclear Energy Mission. World Nuclear News reports the Union Budget set up a Nuclear Energy Mission with ₹20,000 crore for small modular reactors (SMRs), with at least five indigenously developed SMRs to be operational by 2033, and set a goal of at least 100 GW of nuclear capacity by 2047.
- December 2025: the SHANTI Act. The Sustainable Harnessing and Advancement of Nuclear Energy for Transforming India Bill was introduced on 15 December 2025, according to Business Standard. It allows private companies to apply for licences to build and operate nuclear plants, while enrichment, spent fuel and heavy water stay with the government. It replaces the 2010 nuclear liability law. News On AIR reports it received presidential assent after Parliament's Winter Session.
- January 2026: SMRs for data centres. IT Minister Ashwini Vaishnaw said small modular reactors are "feasible and practical solutions" for the energy needs of data centres and that new designs allow a nuclear plant on 14 acres of land, according to News On AIR.
Angel One and the defence think tank MP-IDSA both describe the SHANTI Act as a way to supply data centres with round-the-clock clean power, which solar and wind cannot provide on their own. The scale of the task is clear from Business Standard's figures: India has about 8.8 GW of nuclear capacity from 23 reactors, with targets of 22 GW by 2032 and 100 GW by 2047. Nuclear plants take years to build, so renewables and the existing grid will carry most new AI demand this decade. The IEA similarly expects renewables to meet nearly half of new data centre demand globally, with nuclear and SMRs playing a larger role towards the end of the decade.
How to use AI more efficiently
- Keep context short. Long inputs multiply energy use, according to Epoch AI. Start a new chat instead of keeping very long threads.
- Use smaller models for simple tasks. Smaller models need less computation per token. Our guide on when thinking mode is worth it shows how to avoid unnecessary reasoning tokens.
- Cache repeated context. Prompt caching avoids reprocessing the same instructions and documents on every call.
- For builders, measure tokens. Tokens are the best available proxy for energy and cost in your own app.
Frequently asked questions
How much electricity does one ChatGPT query use?
About 0.34 watt-hours on average, according to OpenAI CEO Sam Altman in June 2025. Epoch AI independently estimated about 0.3 Wh for a typical GPT-4o query, and Google reported 0.24 Wh for a median Gemini text prompt.
How much water does ChatGPT use per prompt?
Sam Altman said about 0.000085 gallons, roughly 0.32 ml or one fifteenth of a teaspoon. Estimates that include water used by power plants are higher: UC Riverside researchers estimate about 17 ml per request in the US.
Is the "ChatGPT uses a bottle of water" claim true?
Not per prompt. The UC Riverside paper estimates about 500 ml for 10 to 50 medium-length GPT-3 responses, including power-plant water, which is the likely source of the bottle comparison. Per-prompt cooling water reported by Google and OpenAI is under 1 ml.
How much energy does Gemini use per prompt?
Google reported that a median Gemini text prompt uses 0.24 Wh of energy, emits 0.03 g of CO2 equivalent and consumes 0.26 ml of water, in August 2025.
Why did older estimates say 3 Wh per ChatGPT query?
The 3 Wh figure from 2023 assumed long outputs, older A100 GPUs and a larger active model. Epoch AI's 2025 analysis says newer hardware, shorter typical answers and more efficient models bring the figure to about 0.3 Wh.
How much electricity will data centres use in India?
Government goals cited by PTI project India's data centre electricity use rising from about 10 to 15 TWh in 2025 to 40 to 45 TWh by 2030.
What is the SHANTI Act and how does it relate to AI?
The SHANTI Act opens India's nuclear power sector to private companies. The government has linked it to providing clean, round-the-clock power for AI data centres, and the IT Minister has called small modular reactors a practical solution for data centre energy needs.
Limits of this review
- No measurement of our own. All figures come from the sources below.
- Company figures are unaudited. OpenAI published no methodology for its figure, and Google's report was not independently verified.
- Boundaries differ. Some figures count only chips, some the whole data centre, some power plants and some the full life cycle.
- The India calculation is illustrative. It uses assumed usage and published per-prompt rates, not data on where or how Indian prompts are served.
References
- The Gentle Singularity, Sam Altman, June 2025
- How much energy does ChatGPT use?, Epoch AI, 7 February 2025
- Google reveals Gemini AI uses 0.24 watt-hours per query, CO/AI, 21 August 2025
- Google says the median Gemini prompt uses 5 drops of water, The New Stack, August 2025
- Google says a prompt uses 0.24 Wh, researchers say the math is incomplete, Implicator
- Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models, Li, Yang, Islam and Ren, arXiv, 2025
- Mistral AI publishes the first comprehensive life cycle assessment of a large language model, The Decoder, July 2025
- Our contribution to a global environmental standard for AI, Mistral AI, July 2025
- Energy and AI: energy demand from AI, International Energy Agency, April 2025
- Energy demand for data centres to double by 2030, AI biggest driver: IEA, Down To Earth, 10 April 2025
- India's data centre boom and grid planning, Outlook Business, April 2026
- India has 100M weekly active ChatGPT users, Sam Altman says, TechCrunch, 15 February 2026
- Indian budget launches Nuclear Energy Mission, World Nuclear News, 3 February 2025
- SHANTI Bill 2025: private nuclear sector entry explained, Business Standard, 15 December 2025
- Vaishnaw says small modular nuclear reactors offer feasible solutions for data centres, News On AIR, 2 January 2026
- SHANTI Bill to support data centres and 24x7 clean power in India, Angel One
- Atoms for AI: the promise of the SHANTI Act, MP-IDSA
- India's digital thirst: data centres are rising in water-scarce regions, Down To Earth, November 2025