How much water does AI use? A single prompt uses somewhere between a few drops and half a litre, depending on whose estimate you read, and all US data centres together used about 228 billion gallons in 2023. Only about 17 billion gallons of that cooled servers. The rest was consumed by the power plants that supplied their electricity, and a third share, the water spent making the chips in the first place, is missing from most figures altogether. This post puts the three parts side by side, with every number dated and sourced.
How much water does one AI prompt use?#
A typical text prompt to a chatbot uses between about a quarter of a millilitre and 15 millilitres of water, and the larger figures that went viral are either outdated or count something different. According to Google's own figures, as reported by EcoFlow (2025), a median Gemini text prompt uses 0.26 millilitres, about five drops. OpenAI's Sam Altman has described a typical ChatGPT query as using about a fifteenth of a teaspoon. Both are company figures, and neither has been independently checked.
The best-known alarming number came from the Washington Post and UC Riverside in 2024, which estimated that a 100-word email written by GPT-4 at an average US data centre used 519 millilitres, a little more than a bottle of water. Shaolei Ren, the researcher behind that estimate, now puts a GPT-4 prompt closer to 15 millilitres, about 5 of them on site, and calls the original figure outdated, according to Forbes (2026). The same Forbes report cites a 2025 benchmark, "How Hungry is AI?", which found efficient models using under 2 millilitres a query and some reasoning models over 150.
The table below decodes the numbers that circulate most often, because each one measures something different.
| Claim you may have seen | Number | What it actually measures | Source |
|---|---|---|---|
| "Five drops of water per prompt" | 0.26 ml | Median Gemini text prompt, Google's own accounting | Google, via EcoFlow (2025) |
| "A fifteenth of a teaspoon" | about 0.33 ml | Altman's description of a typical ChatGPT query | CBS News (2026) |
| "A bottle of water per email" | 519 ml | A 100-word GPT-4 email, cooling plus electricity | Washington Post and UC Riverside (2024) |
| "Closer to 15 ml" | about 15 ml | Ren's revised GPT-4 estimate, about 5 ml of it on site | Forbes (2026) |
| Efficient against reasoning models | under 2 ml to over 150 ml | Range across models and tasks | "How Hungry is AI?" (2025), via Forbes |
Table: the viral per-prompt water numbers, decoded. Company figures are labelled as such.
The chart puts the same figures on one scale, and the spread is the point. According to CBS News (2026), the largest widely shared figure is about 2,000 times the smallest. This is because some figures count only the water evaporated at the data centre, while others add the water behind the electricity, and they cover different models doing different tasks.
The fair reading is that a text prompt on an efficient model costs a few millilitres at most, while long reasoning tasks and image or video generation cost far more. The per-prompt number is also the least useful one, for reasons the next sections explain.
How much water do data centres use in total, and where does it go?#
US data centres used about 228 billion gallons of water in 2023, and roughly nine gallons in every ten were tied to the power plants that supplied them. According to Lawrence Berkeley National Laboratory, as reported by CBS News (2026), about 17 billion gallons went to on-site cooling and about 211 billion were tied to generating the electricity the buildings used.
On-site cooling still matters, because of how the water is used. In evaporative cooling, the most common kind, water absorbs heat from the servers and evaporates into the air. According to the Environmental and Energy Study Institute (2025), about 80 per cent of the water a data centre withdraws is lost this way. That water is consumed, which is different from water withdrawn and returned, and it leaves the local watershed until it falls again as rain.
The power plants account for the larger share because generating electricity consumes water too. According to EESI (2025), citing the US Energy Information Administration, a coal plant consumes about 19,185 gallons per megawatt-hour and a gas plant about 2,800. As a result, the same AI workload carries a very different water bill on a coal-heavy grid than on one powered by wind and solar.
Most reporting stops at those two parts. There is a third, which the diagram below adds, and it is spent before a server is ever switched on.
The totals are also growing. Lawrence Berkeley National Laboratory projects that US data-centre water use could reach 469 to 844 billion gallons by 2028. Globally, UN University researchers found that data centres consumed 4.5 trillion litres of water in 2025 and project 9.3 trillion litres by 2030, according to Reuters (2026), while AI's share of data-centre electricity rises from about a fifth to 40 per cent. For AI alone, a 2025 study in Nature Sustainability estimated that US AI servers could carry a water footprint of 731 million to 1.125 billion cubic metres a year between 2024 and 2030, most of it indirect.
Training a model has a water cost of its own. According to Li and others (2025), training GPT-3 in Microsoft's US data centres directly evaporated about 700,000 litres of fresh water, and that counts only the on-site share. Our explainer on how AI learns covers what training involves.
How much water does it take to make an AI chip?#
Making the chips inside an AI server takes a great deal of water, and by the time a single chip reaches a data centre it has already consumed thousands of gallons, according to EESI (2025). The water is used to make ultrapure water, which is purified to an extreme degree and used to clean, etch and rinse chips at microscopic scale. Our guide to what a GPU does in AI explains why AI needs so many of these chips.
Ultrapure water is expensive in water terms because purification rejects much of what goes in. According to EESI (2025), producing 1,000 gallons of ultrapure water takes about 1,500 gallons of ordinary piped water. An average chip factory, or fab, uses about 10 million gallons of ultrapure water a day, a figure both EESI and the Lincoln Institute of Land Policy (2025) take from the World Economic Forum. The Lincoln Institute compares that daily volume to the use of about 33,000 US households.
Putting those two published numbers together gives a figure neither source states. At 1.5 gallons of piped water per gallon of ultrapure water, a fab using 10 million gallons of ultrapure water a day draws roughly 15 million gallons of piped water a day. This is this post's arithmetic on averages, not a measurement of any one fab, and real plants vary with their recycling.
Recycling reduces the bill but does not remove it. According to Li and others (2025), the average recycling rate in Singapore, one of the more water-disciplined chip hubs, is 45 per cent for wafer fabrication plants and 23 per cent for semiconductor plants generally. The same paper notes that Apple reports its supply chain as 99 per cent of its total water footprint. That is Apple's own figure, and it shows how far the water used to make hardware can outweigh the water used to run it.
The chip water is also concentrated in one place, which turns it into a supply risk. According to Chatham House (2026), Taiwan produces over 90 per cent of the world's advanced semiconductors, and the island's groundwater depends on seasonal typhoons that climate change has made less predictable. A drought there would limit how many AI chips the world can make. None of this is affected by how efficiently a data centre cools its servers, because the chip arrives with its water already spent.
Why does location matter more than the per-prompt number?#
Where a data centre draws its water matters more than how much a single prompt uses, because the pressure lands on particular towns in particular weeks. According to the Guardian (2026), using Cleanview data, 517 of 809 planned US data centres are in places that spent the past year in drought. Consumer Reports (2026) found that about two-thirds of data centres built since 2022 sit in water-stressed areas, and a large facility can use up to 5 million gallons a day, which Brookings (2025) compares to the needs of a town of up to 50,000 people.
| Local pressure | Figure | Source |
|---|---|---|
| Planned US data centres in drought areas | 517 of 809 | The Guardian (2026), Cleanview data |
| US data centres in high or extreme water stress | about 40% | Florida Water and Pollution Control Operators Association |
| Phoenix-area cooling water, today against planned build-out | 385 million to 3.7 billion gallons a year | Ceres, via Consumer Reports (2026) |
| Peak-day cooling demand against the annual average | 6 to 30 times | UC Riverside and Caltech, via Forbes (2026) |
| New US water infrastructure needed by 2030 | $10 to 58 billion | UC Riverside and Caltech, via Forbes (2026) |
| Google's share of water use in The Dalles, Oregon | over 25% | Food & Water Watch (February 2026) |
Table: the local figures that per-prompt estimates hide.
Phoenix shows how fast this can change. According to Consumer Reports (2026), citing Ceres, data centres around Phoenix use about 385 million gallons of water a year for cooling, and planned facilities could raise that to 3.7 billion, an 870 per cent increase. The chart sets that local jump beside the national projection.
Peaks matter as much as totals. According to Forbes (2026), researchers at UC Riverside and Caltech found that peak-day cooling demand can reach 6 to 30 times the annual average, which is why US water systems may need $10 to 58 billion in new infrastructure by 2030. Annual averages hide the hot afternoon when a utility actually runs short. Our post on why AI data centres are being built where they are covers how sites are chosen, and what an AI data centre is explains the cooling systems inside.
Is AI's water use large compared with other things?#
Nationally, AI's water use is small next to farming and lawns. Locally, it can still be a serious problem, and the two facts sit together without contradiction. According to CBS News (2026), US residential lawns used about 2,900 billion gallons of irrigation water and golf courses 531 billion, against 228 billion for every US data centre combined. Fodder crops grown for animal feed take about 5.5 trillion gallons of irrigation water a year. CBS also checked the viral claim that a single hamburger uses more water than thousands of AI prompts, and found it true however it is counted.
Those comparisons can make the problem look smaller than it is. According to the International Energy Agency, as cited by CBS News (2026), AI is about 15 to 20 per cent of data-centre electricity today, and the UN University projects that share rising to 40 per cent by 2030. Lawns and golf courses are not doubling in four years, and they are not being built in clusters in the driest parts of Arizona. The useful comparison is with a region's own water budget, not with the national total.
Reuse helps, though it is not yet common. According to Brookings (2025), closed-loop cooling can cut freshwater use by up to 70 per cent. Amazon Web Services said in 2024 that it was expanding recycled water to 100 data centres, Google says it uses non-potable water at about a quarter of its sites, and Microsoft is piloting designs that evaporate no water for cooling. These are the companies' own statements. Closed-loop systems often need more electricity, which can raise the indirect water bill on a fossil-fuelled grid, so reuse moves part of the footprint rather than erasing it.
How should you read the next AI water headline?#
Ask what is being counted before you react to the number. A figure for cooling alone, a figure that adds electricity, and a figure that includes chip manufacturing can differ by orders of magnitude while all being accurate. Then ask where and when it was measured, since a desert data centre on a hot afternoon is a different machine from one in a cool climate, and finally how old the estimate is, since the 519 millilitre email is already outdated by its own author's account.
Your own choices matter less than the overall pattern, though they are not nothing. According to UN University (2026), the water tied to the electricity for a single AI image is about two tablespoons (29 millilitres), and a complex AI video can reach 4.1 litres. Plain text requests on a default model are the cheapest thing you can ask for. Running a model on your own computer moves the footprint to your local grid rather than removing it, as our post on the electricity one AI prompt uses explains.
The larger fix sits with operators and regulators. Mandatory reporting of all three parts of the footprint, siting decisions that weigh local water stress, and chip supply chains that publish their water use would answer most of the questions above. Until that reporting exists, the honest answer to how much water AI uses is a range with its assumptions written next to it, and any single precise number deserves a second look.
Sources#
- Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models (Li et al., arXiv 2304.03271, 2025)
- 2024 United States Data Center Energy Usage Report (Lawrence Berkeley National Laboratory)
- How much water does AI really use? (CBS News, 2026)
- Data Centers and Water Consumption (EESI, 24 June 2025)
- Environmental impact and net-zero pathways for sustainable artificial intelligence servers in the USA (Nature Sustainability, 2025)
- The carbon and water footprints of data centers (Patterns, 2026)
- AI to double data centre power and water consumption by 2030, UN researchers say (Reuters, 3 June 2026)
- Environmental Cost of AI's Energy Use, Carbon, Water and Land Footprints (UNU-INWEH, 3 June 2026)
- AI water usage requires governments to rethink their approach to water (Chatham House, June 2026)
- Data Drain, the land and water impacts of the AI boom (Lincoln Institute of Land Policy, 17 October 2025)
Last reviewed: 18 October 2026. Merged with the former chip manufacturing water post. Per-prompt estimates and the 2028 and 2030 projections are the figures most likely to move, and the post is re-checked quarterly.
