How Much Water Does AI Use? Between Five Drops and a Tablespoon per Query, Depending on Who Is Counting

Short answer: it depends on what is counted. Google reports that a median Gemini text prompt consumes 0.26 mL of water, about five drops. That is a company figure that nobody outside has audited, and it counts on-site cooling water only. The most-cited academic estimate, from UC Riverside, is a 500 mL bottle for every 10 to 50 responses, or 10–50 mL each, because it assumes an older, hungrier model and adds the water used to generate the electricity. Either way one query is tiny beside a single 1.6-gallon toilet flush. The real issue is the total: Google alone reported consuming 10.9 billion gallons of water in 2025, and some of that is drawn in places that are short of it.

The per-query estimates side by side

The top and bottom of that table differ by a factor of about 180. Neither end is a measurement by an independent party. The company figure comes from internal telemetry that outsiders cannot inspect; the academic figures are calculated from assumed energy per request multiplied by published water-efficiency averages, because the companies had not released the underlying data when the paper was written. The paper says as much: its headline example is that training GPT-3 "can directly evaporate 700,000 liters of clean freshwater, but such information has been kept a secret."

Why does AI use water at all?

A data centre turns nearly all the electricity it draws into heat, and the heat has to leave the building. There are two places water gets used along the way.

On-site, for cooling. Many large facilities reject heat through cooling towers or evaporative systems. Water is evaporated to carry the heat off, and the evaporated portion does not come back to the local supply. Lawrence Berkeley National Laboratory's 2024 report on US data centres describes one such design as "highly energy-efficient" but adds that it "consumes substantial amounts of water, like all evaporative cooling systems," and that evaporation in cooling towers "has raised concerns regarding data center water consumption and availability at the local level." Facilities that use air cooling or closed loops use far less water on site and usually somewhat more electricity; Google's environmental report says water cooling needs less energy than air cooling, without giving a figure.

Off-site, at the power station. Thermal power plants (gas, coal, nuclear) evaporate water in their own cooling systems, and hydroelectric reservoirs lose water to evaporation. A data centre that buys that electricity is responsible for a share of that water in the same way it is responsible for a share of the emissions. This is the same accounting question that decides whether electric cars come out ahead: the answer changes depending on whether the boundary is drawn around the device or around the grid behind it.

The industry's metric for the first category is water usage effectiveness, or WUE: litres of water consumed on site per kilowatt-hour of IT electricity.

Five reasons the numbers differ by a factor of hundreds

1. On-site only, or on-site plus electricity

This is the largest single cause. In the UC Riverside paper's US-average case, on-site cooling accounts for 2.2 mL of a 16.9 mL request and electricity generation for the other 14.7 mL, or 87% of the total. Google's 0.26 mL figure is built from the company's on-site water usage effectiveness, so it covers water consumed at the data centre and not water used in generating the electricity. A reader comparing 0.26 with 16.9 is mostly comparing one scope with two.

2. How much energy a query is assumed to need

The UC Riverside estimate was built around GPT-3 and assumes about 0.004 kWh (4 Wh) per medium-length response. Google reports 0.24 Wh for a median Gemini text prompt, roughly a seventeenth of that. Google also reports that the energy per prompt fell by a factor of 33 in the twelve months to May 2025. If the company figure is accurate, an estimate built around GPT-3 overstates today's typical text query by an order of magnitude through the energy term alone. If it is not accurate, nobody outside can currently show it.

The two effects can be separated with simple arithmetic. Take the on-site parts only: 2.2 mL against 0.26 mL is a gap of about eight times, not 65. And if Google's 0.24 Wh is multiplied by the paper's US-average off-site factor of 3.14 litres per kWh, the electricity side adds about 0.75 mL, for a total near 1 mL per query. That last figure is an illustration produced for this article, not a number either source publishes, but it shows roughly where a like-for-like comparison lands for an efficient modern text model on an average US grid.

3. Withdrawal versus consumption

Water withdrawal is everything taken from a river, aquifer or mains supply. Water consumption is withdrawal minus what is returned, which in practice means what evaporates. A power plant with once-through cooling withdraws enormous volumes and returns most of it, warmer. The UC Riverside paper's per-query figures are consumption. Its headline global projection, 4.2 to 6.6 billion cubic metres in 2027, is withdrawal; the same paper puts the consumed part at 0.38 to 0.60 billion cubic metres, about a tenth as much. The larger number is the one that tends to be quoted. Both are projections, not observations.

4. Training versus inference

Training a model is a one-off cost; answering queries (inference) is continuous. The paper estimates that training GPT-3 in Microsoft's US data centres consumed about 700,000 litres on site and 5.4 million litres including electricity. Google's per-prompt figure explicitly excludes training. Spread across billions of queries, training adds little to each one, but it is real water. This article found no comparable published training figure for a current model.

5. Model, modality, place and season

"A median text prompt" hides a wide spread. Long outputs, reasoning modes, image generation and video take more computation than a short text answer, and none of the per-query figures above cover them. Location matters as much: in the paper's table, the same request costs 0.56 mL of on-site water in Virginia and 6.5 mL in Arizona, and the off-site part ranges from 6.6 mL in Texas to 43.7 mL in hydro-heavy Washington. On-site use also rises in hot weather, when evaporative systems work hardest; the paper notes that water efficiency varies with time as well as place. The phrase "10 to 50 responses per bottle" is that location-and-timing spread.

How much water does ChatGPT use per query?

There is no documented company figure for ChatGPT that this article could cite. Per-query numbers attributed to OpenAI circulate, but none was found with a published method, scope or model mix in a source this site accepts, so none is repeated here.

What does exist is the UC Riverside range of 10 to 50 mL per medium-length response. It was calculated for GPT-3, with an assumed 4 Wh per response, and includes electricity-generation water. It was a reasonable estimate given what was public at the time. If Google's measured energy figure is any guide to current text models, it is probably too high for a typical short text answer today, and it may not be too high for a long reasoning-mode answer or an image.

A fair reading, and it is this article's reading, not a published figure: a fraction of a millilitre of on-site water for an ordinary text query on an efficient current model, around a millilitre once power-plant water is added on an average US grid, with the top of the old range still possible for heavy requests in unfavourable locations. Anyone who states a single number with confidence is choosing a scope and not saying so.

AI data centre water usage in total

Per-query figures make the subject look trivial, and totals make it look alarming. Both views are needed.

The Berkeley Lab figures are for all data centres, not AI alone, and the report is candid about its limits: "the lack of direct energy data available in a sector with rapidly evolving technologies limits the analysis." The report does show the direction: electricity use roughly tripled between 2014 and 2023, and it attributes the more-than-doubling of demand between 2017 and 2023 to the rapid growth of accelerated servers, the kind used for AI, with further substantial increases possible by the end of the decade.

Google's 10.9 billion gallons is the company's own total for its data centres and offices, so it covers everything Google runs, including search, video and cloud customers, not AI alone. Divided by 365 it works out to about 30 million gallons a day (this article's arithmetic), or the water of roughly 19 million toilet flushes at the federal standard. Google's report offers its own comparison: about what it takes to irrigate 73 golf courses a year in the southwestern United States. That makes it a large industrial water user, and still a small one next to irrigated agriculture.

"Replenished" is not the same as "did not use." Replenishment projects restore or save water somewhere in a watershed; they do not necessarily return it to the aquifer or utility that supplied the data centre. The volumes are calculated by the companies and their partners.

One query against a flush, a dishwasher cycle and a T-shirt

The table below gives the number of AI text queries that equal one everyday use, at both ends of the published range: Google's 0.26 mL and the UC Riverside US average of 16.9 mL.

The query counts are this article's division of each volume by 0.26 mL and by 16.9 mL.

Sources for the right-hand side of each comparison: the toilet figure is the EPA WaterSense statement of the federal standard, and the dishwasher figure is the current ENERGY STAR ceiling for standard-size machines; real cycles vary, as the piece on how much water a dishwasher uses sets out. For showers, see how much water a shower uses. The T-shirt figure is from Chapagain and colleagues' 2006 analysis: 1,230 litres of irrigation and process water, 1,110 of rainfall and 380 of dilution water for pollution.

The beef rows need a caveat, and it is the same caveat that applies to AI. Mekonnen and Hoekstra's global average for beef is 15,400 litres per kilogram, of which about 94% is "green" water, rain that fell on pasture and feed crops and would have fallen anyway. Only about 4% (550 litres per kilogram) is "blue" water taken from rivers and aquifers. Cooling water drawn from a mains supply or an aquifer is blue water. So the fair comparison for a burger is the 62-litre row, not the 1,750-litre one. Neither row is a published per-burger figure: both are this article's multiplication of the global per-kilogram averages by a 113 g (quarter-pound) patty. The environmental cost of fast fashion runs into the same problem with cotton.

With that said, the per-person arithmetic is clear. Someone who sends 50 text queries every day for a year sends 18,250. At Google's figure that is under 5 litres a year. At the UC Riverside US average it is about 310 litres, roughly 80 gallons, or about 50 toilet flushes. On no published estimate is an individual's chatbot habit a meaningful part of their water footprint.

Does AI use a lot of water?

Per query, no. In aggregate, yes, in the sense that any industry consuming billions of gallons a year uses a lot of water. Whether that total is a problem depends almost entirely on where it is drawn.

A data centre cooled with evaporative towers in a wet, cool region with a healthy river is a modest industrial customer. The same facility on a stressed aquifer in the desert Southwest competes directly with households and farms, and it draws most in the hottest weeks, when supply is tightest. The UC Riverside table puts Arizona's on-site water per request at nearly twelve times Virginia's. National and global totals average this away, which is why they are less informative than they look, and why the useful disclosure is per site and per watershed. Few operators publish that. Google says it applies a water risk framework to its data centres to evaluate local watershed health and guide technology choices; that is a company process, described by the company.

There is also a trade-off that has no free option. Air cooling saves on-site water and costs electricity, and on many grids more electricity means more water evaporated at a power station somewhere else, plus more carbon. The paper found that carbon efficiency and water efficiency are only weakly related, with a correlation coefficient of 0.06 in Virginia. A facility can be scheduled to look good on one metric and poor on the other.

What is not worth doing

Rationing chatbot questions to save water. At a fraction of a millilitre to a few millilitres each, skipping a hundred queries saves somewhere between a mouthful and a glass. One avoided toilet flush, at about six litres, saves more than that many times over.

Quoting "a bottle of water per conversation" as settled fact. The figure is 10 to 50 responses per 500 mL, for GPT-3, including power-plant water, from a modelled estimate. It was careful work on the data available. Used without those qualifiers for every AI product in 2026, it is no longer what the authors wrote.

Quoting "five drops" as settled fact either. That is one company's median text prompt, on-site water only, excluding training, measured by the company. It says nothing about image or video generation, other providers, or power-plant water.

Comparing AI with beef or cotton using total water footprints. Most of those footprints is rainfall. Compare blue water with blue water or leave the comparison out.

What to actually do

Keep the proportions straight. For a household, the water that matters is in the shower, the toilet, the washing machine and above all the garden. Fixing a running toilet does more than any change in software habits.

If the concern is AI's footprint, look at the heavy uses. Text queries are the cheap end. Bulk image and video generation, and automated agents that run thousands of calls unattended, are where an individual or a business can move their own number, on energy as well as water.

For anyone living near a proposed data centre, ask the local questions. What is the cooling design: evaporative, air or closed loop? What is the source: potable mains, groundwater or reclaimed wastewater? What is the peak daily draw in summer, not the annual average? Is the utility's contract public? These are answerable and they matter more than any national statistic.

Read company figures as company figures. They are the best data available and they are unaudited. A number becomes more credible when it comes with a stated boundary and method, as Google's per-prompt figure does, and less so when it arrives as a single line with neither. The same habit applies to other claims AI companies make about their products, including what happens to the conversations themselves.

Expect the numbers to move. Energy per query has fallen fast, by Google's account 33-fold in a year, while total data centre electricity is projected to roughly double or triple by 2028. Per-query efficiency and total consumption are heading in opposite directions, and both statements will be true at once for some time.

Questions people ask

How much water does AI use? For one text query, published figures run from about 0.26 mL (Google's own on-site figure) to 10–50 mL (a UC Riverside estimate for GPT-3 that includes electricity generation). In total, Google reported consuming 10.9 billion gallons of water in 2025 across its data centres and offices.

How much water does ChatGPT use? There is no per-query figure for ChatGPT published with a method that this article could verify. The best-known estimate, a 500 mL bottle per 10 to 50 responses, is an academic calculation for GPT-3 that counts power-plant water as well as on-site cooling, and it is probably high for a short text answer from a current model.

How much water does one ChatGPT query use? Probably somewhere between a fraction of a millilitre and a few millilitres for an ordinary text answer, depending on whether electricity-generation water is counted; that is a reading of the published estimates, not a measurement. Long reasoning answers and image generation use more. No independently audited per-query figure exists for any major chatbot.

Why does AI use water? Data centres turn electricity into heat, and many reject that heat by evaporating water in cooling towers. Additional water is evaporated at the power stations that generate their electricity. In the UC Riverside US-average estimate, the power-station share was about 87% of the total.

Does AI use a lot of water? Per query, very little: between a few hundred and about 23,000 text queries equal one toilet flush, depending on the estimate. In aggregate, large operators consume billions of gallons a year, which is significant where facilities draw on stressed local supplies.

How much water does AI use per day? There is no audited global daily figure for AI alone. Google's company-reported total works out to about 30 million gallons of water a day in 2025 across all its services, not only AI. An academic projection put global AI water consumption at 0.38–0.60 billion cubic metres for 2027.

How much water do AI data centres use? It varies enormously with cooling design and climate, and few operators publish per-site figures. As a guide to scale, Google reported 10.9 billion gallons for its data centres and offices in 2025, and Lawrence Berkeley National Laboratory estimates that all US data centres used 176 TWh of electricity in 2023, 4.4% of the national total.

Should I use AI less to save water? Not for water's sake. Fifty text queries a day for a year comes to between about 5 and 310 litres on the published estimates, at most roughly 50 toilet flushes. Household fixtures, leaks and outdoor watering are far larger levers.

This article summarises published estimates of water use by AI systems and data centres for general information. Most per-query and company-level figures are self-reported by the companies concerned and have not been independently audited; modelled academic estimates rest on stated assumptions. Figures change quickly as models, hardware and cooling designs change.