How much water does ChatGPT use? OpenAI’s CEO says an average ChatGPT query uses about 0.000085 gallons of water (0.32 mL, roughly one fifteenth of a teaspoon) and 0.34 watt-hours of electricity. Researchers who also count the water power plants use to make that electricity get much bigger numbers: about 7 to 48 mL per response for GPT-3, depending on where it runs.
This free AI water calculator puts the published figures side by side: enter how you use AI and see a low, middle and high estimate of water and electricity per prompt, per day and per year, with everyday comparisons and a source for every number.
Free AI water calculator
How much water and electricity does your AI use?
Tell us how you use AI on a typical day. You get a low, middle and high estimate of the water and electricity it takes, per prompt, per day and per year, built only from published figures, with the source behind every number.
Which assistant do you use most?
Your estimate in three scenarios
Where your water comes from (middle estimate)
Per prompt, image or clip
Your year in everyday terms
Unlock the source-by-source table
See your year of everyday prompts under each of 13 published estimates side by side: Google, OpenAI, Microsoft, Epoch AI, Mistral, and six data center locations from the UC Riverside study.
Unlocked: the source-by-source table is just below. You are also on the weekly HOW TO // AI newsletter.
Free. You also get the HOW TO // AI newsletter once a week. Unsubscribe any time.
How the calculator works
The formula: for each kind of use, items a day times the published figure per item, added up. A year is that daily total times your days a week times 52.14 (365 divided by 7).
- Low: for everyday prompts, the company's own figure (OpenAI: 0.34 Wh and 0.32 mL; Google: 0.24 Wh and 0.26 mL), which counts only the water that cools the data center. For thinking prompts, images and video, the lowest research figure we use, with cooling water at Google's 2024 fleet average of 1.15 mL per Wh.
- Middle: the same company energy figure for everyday prompts and typical research figures for the rest, counting all the water: cooling plus the water power plants evaporate to make the electricity. We use the US average from Li et al., about 3.6 mL per Wh of data center electricity.
- High: the highest research figures, all water counted: Li et al.'s GPT-3 estimate for everyday prompts (4.7 Wh, 16.9 mL), DeepSeek-R1 on its own servers for thinking prompts (29 Wh), the least efficient image model tested (13.4 Wh) and the most power-hungry of 7 open video models (128 Wh per clip).
- Thinking prompts use Jegham et al.'s estimates for a long prompt (about 10,000 tokens in and 1,500 out): o3-mini for low, o3 for middle. Images come from 2023 open models (Luccioni et al.) and video from 7 open models making clips of about 5 seconds (Hugging Face). Images and video were measured at the server, so middle and high add 17% for the data center building (Li et al.'s US average PUE of 1.17).
- The scenarios cover running the model. They do not include training it or your own phone or computer.
Why the range is so wide
- Which water counts. Google and OpenAI count the water that cools the data center. Li et al. add the water power plants evaporate to make the electricity, which is most of their total (14.7 of 16.9 mL at the US average). Mistral's 45 mL per response is a lifecycle figure that also counts making the servers.
- Where it runs. In Li et al., the same GPT-3 request ranged from 7.1 mL in Ireland to 47.5 mL in Washington state.
- Which model and mode. Hugging Face found reasoning models used about 30 times more energy on average than models without reasoning.
- How new it is. Google says its energy per Gemini text prompt fell 33 times from May 2024 to May 2025, and Microsoft's June 2026 study puts a typical query to large models at 0.16 to 0.60 Wh with 0 to 0.067 mL of cooling water. Li et al. studied an older model, GPT-3.
- Who checked it. Company figures are self-reported. Google says its figures were not verified by an independent third party, and OpenAI has not published its method.
These are published estimates for learning about AI's footprint, not a measurement of your own account.
Sources: Google (Gemini, Aug 2025), OpenAI CEO Sam Altman (June 2025), Li et al., Making AI Less Thirsty (rev. Mar 2025), Microsoft (June 2026), Mistral (July 2025), Jegham et al. (Nov 2025), Hugging Face AI Energy Score (Dec 2025), Luccioni et al., Power Hungry Processing (2024), Hugging Face text-to-video tests (July 2025). Comparisons: EPA WaterSense (2.5 gallons a minute per standard showerhead, 1.6 gallons per flush, 82 gallons a day per person at home), ENERGY STAR (14 gallons per laundry load), EPA (0.019 kWh per phone charge), EIA (10,791 kWh a year per US home, 2022) and Google (0.3 Wh per search, 2009). Checked October 18, 2026.
How to use the AI water calculator
- Pick your assistant. Only OpenAI and Google have published both an energy and a water figure per prompt for their own assistant that we found, so other assistants use their range.
- Enter a typical day: everyday prompts, long thinking prompts, images and short video clips, plus how many days a week you use AI.
- Read your low, middle and high totals, then copy your summary to share it.
Why the estimates differ so much
The published numbers mostly measure different things:
- Which water counts. Google (0.26 mL per Gemini text prompt) and OpenAI (0.32 mL per ChatGPT query) count the water that cools the data center. Researchers at UC Riverside and UT Arlington add the water evaporated at power plants, which is most of their total: 14.7 of 16.9 mL per GPT-3 request at the US average.
- Where it runs. In the same study, one request ranged from 7.1 mL in Ireland to 47.5 mL in Washington state.
- What else is counted. Mistral’s 45 mL per Le Chat response is a lifecycle figure that includes making the servers.
- Which model and mode. Hugging Face found reasoning models used about 30 times more energy on average than models without it.
- How new it is. Google says its energy per prompt fell 33 times from May 2024 to May 2025, and a June 2026 Microsoft study says widely cited estimates are 4 to 20 times too high. The highest research figures come from older models.
Company figures are self-reported. Google says its numbers were not verified by an independent third party, and OpenAI has not published how it got its figure.
Published water and energy figures per text prompt
| Source and date | Electricity | Water | What it counts |
|---|---|---|---|
| Google, median Gemini Apps text prompt (Aug 2025) | 0.24 Wh | 0.26 mL | Cooling water only, with idle capacity and data center overhead |
| OpenAI CEO Sam Altman, average ChatGPT query (June 2025) | 0.34 Wh | 0.32 mL | Company statement, method not published |
| Microsoft, typical query to large models (June 2026) | 0.16 to 0.60 Wh (median 0.31) | 0 to 0.067 mL | Company study, cooling water only, data center overhead included |
| Epoch AI, typical GPT-4o query (Feb 2025) | About 0.3 Wh | Not given | Independent estimate |
| Mistral, Le Chat 400-token response (July 2025) | Not given | 45 mL | Lifecycle, including making the hardware |
| Li et al., GPT-3 medium-length request (revised Mar 2025) | 4 Wh at the server | 7.1 to 47.5 mL (16.9 mL US average) | Cooling plus power plant water, by location |
Other uses cost more: about 3.5 to 29 Wh for a long prompt to a reasoning model (Jegham et al.; Microsoft’s median for long reasoning queries is 3.91 Wh), 1.4 to 11.5 Wh per generated image (Luccioni et al., 2023 open models) and 4 to 109 Wh per 5-second video clip (Hugging Face).
How much water and electricity does AI use overall?
There is no official worldwide total for AI. Li et al. projected that global AI demand could account for 4.2 to 6.6 billion cubic meters of water withdrawal in 2027 (withdrawal includes water that is later returned). For electricity, the IEA says data centers used about 485 TWh in 2025 and expects roughly 950 TWh by 2030, around 3% of the world’s electricity.
How to cut your AI water and energy use
- Turn off thinking modes for simple questions. Microsoft estimates that switching reasoning off can cut the energy of a query by 5 times or more.
- Use a smaller, faster model when it is good enough. Mistral reports that a model 10 times bigger has impacts about 10 times larger.
- Get it right in fewer tries. Our free AI prompt improver turns a rough prompt into a clear one.
- Paste only what is needed. Epoch AI estimates a query with a 10,000-token input uses about 2.5 Wh, around 8 times a typical query.
- Make images and video on purpose. Hugging Face measured one clip from the largest open video model it tested at about 380 times the energy of a short chatbot reply.
Related on HOW TO // AI
- ChatGPT vs Claude vs Gemini: which assistant to use.
- AI chatbot pricing compared: what each plan costs.
- Free AI tools: more calculators and checkers like this one.
Want practical AI tips like these every week? Join the free HOW TO // AI newsletter.
Checked against Google, OpenAI, Microsoft, Mistral, Epoch AI, Li et al., Jegham et al., Hugging Face, Luccioni et al., the IEA, EPA WaterSense, ENERGY STAR, the EPA equivalencies calculator and the US EIA on October 18, 2026.
Frequently asked questions
How much water does one ChatGPT prompt use?
OpenAI's CEO said in June 2025 that an average ChatGPT query uses about 0.000085 gallons, or about 0.32 mL, which counts the water used to cool the data center. Researchers who also count power plant water estimated about 7 to 48 mL per response for GPT-3, depending on the data center's location.
Does ChatGPT really use a bottle of water?
That idea comes from a study (first published in 2023, revised in 2025) estimating that GPT-3 used a 500 mL bottle of water for roughly 10 to 50 medium-length responses, counting power plant water. So it is one bottle for 10 to 50 responses, not one per prompt.
Why does Google say five drops while others say much more?
Google's 0.26 mL per Gemini text prompt counts only the water that cools its data centers, at its 2024 fleet average. Higher estimates add the water evaporated at power plants, older and larger models, less efficient locations, or the lifecycle of the hardware.
How much water does AI use in total?
There is no official worldwide figure. Researchers at UC Riverside and UT Arlington projected that global AI demand could account for 4.2 to 6.6 billion cubic meters of water withdrawal in 2027. For electricity, the IEA says data centers used about 485 TWh in 2025 and could reach about 950 TWh by 2030.
Do AI images and videos use more water than text?
Yes, in the research we cite. A generated image used about 1.4 to 11.5 Wh in a study of 2023 open models, and a video clip of about 5 seconds used about 4 to 109 Wh in Hugging Face's tests, compared with 0.24 to 0.34 Wh for a company-reported text prompt. Water rises with the electricity used.