
The fleet total, sourced — billions of gallons, disclosed and modeled
No operator meters every data center in the country, so the national figure is a modeled total from the federal laboratory charged with producing it. The Lawrence Berkeley National Laboratory's 2024 United States Data Center Energy Usage Report (report number LBNL-2001637, December 2024, prepared for the United States Department of Energy) puts direct, on-site water consumption by all United States data centers at 66 billion liters in 2023, up from 21.2 billion liters in 2014 — a tripling in nine years, with hyperscale and colocation facilities accounting for 84 percent of the 2023 total. Converted at 3.78541 liters per gallon, that is about 17.4 billion gallons for the year, or about 47.8 million gallons per day. The same report puts the indirect figure — water consumed at power plants to generate the electricity those data centers used — at nearly 800 billion liters in 2023, roughly twelve times the direct figure. Anyone quoting only the 17.4 billion is leaving out the larger half of the laboratory's own accounting; anyone quoting only the combined total without saying so is mixing two different places where water is consumed.
The company-level check on that scale comes from an operator's own audited disclosure. Google's Environmental Report 2025 states that in 2024 the company consumed approximately 8.1 billion gallons (31 billion liters) of water across its data centers (excluding those operated by third parties) and offices, and that overall consumption increased by 28 percent from 2023 to 2024. The report's own data table gives the inputs to check that increase: 8,135 million gallons consumed in 2024 against 6,352 million gallons in 2023 — a difference of 1,783 million gallons, which divided by 6,352 is about 28.1 percent, matching the stated increase. The same report says its stewardship projects replenished approximately 4.5 billion gallons in 2024, roughly 64 percent of its freshwater consumption, and that 72 percent of its freshwater withdrawals came from sources at low risk of depletion or scarcity. One honest boundary travels with the Google figure: it is one company's global total for data centers plus offices, not a total for artificial intelligence (AI) alone — Google does not publish an AI-only water total, and no source opened for this briefing does either. That gap is stated here rather than filled: there is no verified figure for “AI's share” of the national total, because the workloads share the same buildings and cooling systems as search, video, and cloud computing.
For the forward view, the same two primary sources carry projections, labeled as projections and not measurements: the Berkeley laboratory expects hyperscale data centers alone to consume between 60 and 124 billion liters directly in 2028, and the Riverside study (below) projects global AI demand accounting for 4.2 to 6.6 billion cubic meters of water withdrawal in 2027, of which about 0.38 to 0.60 billion cubic meters would be consumption — withdrawal being water pulled from a source, consumption being the part evaporated or otherwise not returned, a distinction that study itself defines and that this page keeps throughout.
The per-prompt figure, measured by an operator — five drops, with the method published
The smallest per-prompt figure comes with the most complete method behind it. In August 2025 Google researchers published “Measuring the environmental impact of delivering AI at Google Scale,” a production measurement — not a model built from public assumptions — of serving the Gemini Apps assistant. Its headline result: the median Gemini Apps text prompt in May 2025 consumed 0.24 watt-hours of energy and 0.26 milliliters of water, which the paper puts as the equivalent of five drops (at a standard 0.05 milliliters per drop) and less energy than watching television for nine seconds. The paper shows how the water number is built: water per prompt equals the prompt's energy minus the data center overhead portion, multiplied by the fleet's consumptive Water Usage Effectiveness (WUE — liters of water consumed per kilowatt-hour), which was 1.15 liters per kilowatt-hour for both 2023 and 2024. Redone from those stated inputs — (0.24 minus 0.02) watt-hours, converted to kilowatt-hours and multiplied by 1.15 — the arithmetic gives about 0.25 milliliters, matching the reported 0.26 within rounding. The figure is carried here as an estimate: it is the operator measuring its own fleet, it covers on-site cooling water only, and it excludes training, the network, the user's device, and the water used to generate the electricity. The paper also reports that efficiency work cut the energy per median prompt about 33 times in the year to May 2025 — which is why a 2025 production measurement and a 2023 model of an older system should never be expected to agree.
The second operator disclosure is smaller in documentation and must be labeled for exactly that. In his June 2025 essay “The Gentle Singularity,” OpenAI chief executive Sam Altman wrote that an average ChatGPT query uses about 0.34 watt-hours and “about 0.000085 gallons of water; roughly one fifteenth of a teaspoon.” Converted, 0.000085 gallons is about 0.32 milliliters (0.000085 times 3,785.41 milliliters per gallon), and one fifteenth of a teaspoon is about 0.33 milliliters — the two forms of his figure agree with each other. What the essay does not publish is any methodology, boundary, or definition of “average” — a gap Google's own paper notes explicitly when it cites the figure. It is carried here as an estimate: a company disclosure, unverifiable from outside, and not comparable to the figures below until its boundary is known.
The per-prompt figure, modeled independently — why the older study lands 65 times higher
The figure behind the famous “bottle of water” headline comes from “Making AI Less Thirsty,” by Pengfei Li, Jianyi Yang, Mohammad A. Islam, and Shaolei Ren of the University of California, Riverside and the University of Texas at Arlington (first posted 2023; revised March 26, 2025; accepted by Communications of the Association for Computing Machinery). Read in the paper itself, it does not say one prompt drinks a bottle. It says GPT-3 — the model behind the first ChatGPT, running in Microsoft's United States data centers — needs to “drink” a 500-milliliter bottle of water for roughly 10 to 50 medium-length responses, depending on when and where it runs. Its table gives the United States average in full: about 2.2 milliliters on site plus about 14.7 milliliters off site at power plants, for a total of about 16.9 milliliters per request — and 500 divided by 16.9 is about 29.6 requests per bottle, inside the paper's stated range. The same table spans about 7.1 milliliters per request in Ireland to about 47.5 milliliters in Washington state, because power-plant water intensity and cooling efficiency vary by location. Every figure from this study is an estimate: a principled model built on an assumed 0.004 kilowatt-hours of server energy per medium-length request (about 800 words in, 150 to 300 words out), using Microsoft's published efficiency figures — not a measurement of a live service. Its training figures are estimates on the same basis: training GPT-3 in Microsoft's United States data centers is modeled at about 700,000 liters of on-site water consumption, and about 5.4 million liters in total including power-plant water.
The largest per-prompt figure with a published method is also the one counting the most things. Mistral AI's lifecycle analysis (LCA — a lifecycle assessment adding up impacts from manufacturing through use), conducted with the consultancy Carbone 4 and the French ecological transition agency ADEME (the French Agency for Ecological Transition) and peer-reviewed by Resilio and Hubblo, puts the marginal impact of one 400-token Le Chat response, excluding the user's terminal, at 45 milliliters of water — alongside 1.14 grams of carbon dioxide equivalent. Mistral states the figure includes upstream impacts such as manufacturing servers, not just energy use, and calls the study a first approximation. For scale, the same analysis puts Mistral Large 2's total footprint after 18 months of use at 281,000 cubic meters of water consumed. The 45 milliliters is carried here as an estimate on a lifecycle boundary — the widest boundary of any figure on this page, which is the main reason it is the largest.
One scale, same units — the arithmetic, shown
Every comparison below is computed from the stated inputs, with the arithmetic shown, so any reader can redo it. One gallon equals 3.78541 liters; one milliliter is one thousandth of a liter.
- The national fleet, daily: 66 billion liters ÷ 3.78541 = about 17.4 billion gallons per year; ÷ 365 = about 47.8 million gallons per day consumed directly on site by all United States data centers in 2023 (computed from the Berkeley laboratory's verified total).
- Direct versus indirect: nearly 800 billion liters ÷ 66 billion liters is about 12.1 — the water consumed making the data centers' electricity was about twelve times the water consumed on site in 2023 (computed from the same report's two verified figures).
- One company's growth, checked: 8,135 million gallons (2024) minus 6,352 million gallons (2023) is 1,783 million; 1,783 ÷ 6,352 is about 0.281 — about a 28 percent increase, matching Google's stated increase (computed from the report's own data table, a second route to the narrative figure).
- One company against the national daily rate, as scale only: Google's 8,135 million gallons ÷ 365 is about 22.3 million gallons per day in 2024; 22.3 ÷ 47.8 is about 0.47. That arithmetic is shown only to give the reader a sense of scale — it is not a market share, because the two inputs cover different geographies (global versus United States), different years (2024 versus 2023), and different boundaries (data centers plus offices versus data centers only).
- The per-prompt spread: Mistral's lifecycle estimate ÷ Google's on-site measurement is 45 ÷ 0.26, about 173 times. The Riverside total ÷ Google's figure is 16.9 ÷ 0.26, about 65 times. Altman's disclosure ÷ Google's figure is 0.32 ÷ 0.26, about 1.2 times — the two 2025 operator figures for current systems land close together; the modeled and lifecycle figures for older or wider boundaries are where the spread lives (all computed from the estimates labeled above).
- Where the Riverside spread comes from: of its 16.9-milliliter United States total, 14.704 milliliters is off-site power-plant water — 14.704 ÷ 16.904 is about 87 percent. Google's 0.26 milliliters counts none of that category. Most of the apparent contradiction between the two studies is that one line item (computed from the Riverside paper's own table).
- The bottle, corrected: a 500-milliliter bottle ÷ 0.26 milliliters is about 1,923 median Gemini prompts on Google's on-site accounting; ÷ 16.9 milliliters is about 30 medium-length GPT-3 responses on the Riverside total accounting. “One prompt drinks a bottle of water” is not supported by any source opened for this briefing — the study the claim traces to gives a bottle per 10 to 50 responses, for an older model, on the widest of the two accountings in that study (computed).
The pattern the arithmetic leaves is the one the queue for this briefing predicted: the fleet total and the per-prompt figure sound like they contradict each other because one is billions of prompts added up and the other is one prompt divided out — and because “water per prompt” has no agreed boundary, so each published figure is silently answering a slightly different question.
Why the numbers do not contradict each other — the honest caveats
Four differences keep these figures from being interchangeable, and each cuts against a different side's headline.
What gets counted decides the answer. On-site cooling only (Google's 0.26 milliliters) is the narrowest boundary. Add the water consumed generating the electricity and the same class of request rises to the Riverside total (16.9 milliliters, 87 percent of it off site). Add server manufacturing and the full life cycle and it rises again (Mistral's 45 milliliters). A figure quoted without its boundary is not a smaller or larger answer to the same question — it is an answer to a different question wearing the same words.
Which model, and which year, matters as much. The Riverside model describes GPT-3-era serving at an assumed energy per request roughly seventeen times Google's measured 0.24 watt-hours for a median 2025 Gemini prompt (0.004 kilowatt-hours is 4 watt-hours; 4 ÷ 0.24 is about 16.7). Efficiency gains of that size in two years are exactly what Google's paper documents on its own fleet. Comparing the two per-prompt figures as if they described the same system in the same year is the error; reporting both, dated and bounded, is the fix this page applies.
A median is not a medium-length request, and neither is an image. Google's figure is a median text prompt — half of prompts use less, and long reasoning, image, and video work use more, by amounts no source opened for this briefing quantifies per prompt. The Riverside figure is a defined medium-length conversation request. Altman's “average” is undefined. None of the three can stand in for a heavy workload, and this page claims none for one.
Small per prompt does not mean small in total — and large in total does not mean large per prompt. The disconfirmation runs both ways. Against the reassuring reading: billions of prompts at a fraction of a milliliter each are precisely how a fleet reaches 17.4 billion gallons a year on site, and the Berkeley laboratory's tripling since 2014 happened while per-unit efficiency was improving — volume outran efficiency, which is the fleet story in one line. Against the alarming reading: the per-prompt bottle claim survives nowhere in its own source paper, and the two current operator figures agree with each other within about 20 percent. The local question — what one more large withdrawal does to one specific aquifer — is decided by siting and permits, not by either national figure, which is the same conclusion this stream reached for data centers generally: How much water does a data center use? and yesterday's data center versus golf course comparison.
What happens next — and what would change this briefing
The next verifiable events are publications, not predictions: (1) a per-prompt disclosure from OpenAI with a published methodology and boundary, which would make the ChatGPT figure comparable to Google's for the first time; (2) the Berkeley laboratory's next national update carrying water figures for a year after 2023 — its June 2026 electricity update, as noted in this stream's earlier research, carried no new water figures; (3) a second operator publishing a production measurement on Google's comprehensive boundary, which would test whether 0.26 milliliters is typical of current fleets or specific to one company's. What would change this briefing's conclusions: a metered — rather than modeled — national total materially different from 66 billion liters for 2023, or a peer-reviewed production measurement of a current model landing near the lifecycle figures, either of which would move the per-prompt range this page reports. Until then, the honest answer to the search is the one in the keyline: billions of gallons a year as fleets, drops to tablespoons per prompt depending on what is counted — both sourced, neither canceling the other. The permit side of the same story is in Florida's data center water law, explained, and your own city's supply picture starts at Your city's water quality report, with local treatment help in the license-verified provider directory.
Sources
- National fleet totals: Shehabi and others, 2024 United States Data Center Energy Usage Report, Lawrence Berkeley National Laboratory (LBNL-2001637), December 2024 — opened directly for this briefing. Verified from the report text: direct on-site water consumption of 66 billion liters in 2023 (21.2 billion liters in 2014; hyperscale and colocation at 84 percent of the 2023 total); indirect water consumption through electricity use of nearly 800 billion liters in 2023, at a national average of 4.52 liters per kilowatt-hour; hyperscale direct consumption projected at 60 to 124 billion liters in 2028. All gallon and daily figures on this page are computed from those liter figures at 3.78541 liters per gallon, with the arithmetic shown.
- Company fleet total: Google, Environmental Report 2025 (covering 2024) — opened directly for this briefing. Verified from the report text and its data tables: total water consumption of approximately 8.1 billion gallons (31 billion liters) in 2024 across data centers (excluding third-party-operated) and offices; an increase of 28 percent from 2023 to 2024; table values of 8,135 million gallons (2024) and 6,352 million gallons (2023), from which the 28 percent increase is recomputed on this page; replenishment of approximately 4.5 billion gallons in 2024, roughly 64 percent of freshwater consumption; 72 percent of freshwater withdrawals from low-risk sources.
- Per-prompt, operator measurement: Elsworth and others (Google), “Measuring the environmental impact of delivering AI at Google Scale,” August 2025 — opened directly for this briefing. Verified from the paper: median Gemini Apps text prompt in May 2025 at 0.24 watt-hours and 0.26 milliliters of water (comprehensive boundary; 0.12 milliliters on the narrower existing approach as tabulated, described as 0.15 milliliters in the paper's text); consumptive Water Usage Effectiveness of 1.15 liters per kilowatt-hour for 2023 and 2024; about 80 percent of withdrawn water consumed on average; a 33-times reduction in energy per median prompt over 12 months. Carried on this page as an estimate — the operator's own production measurement, on-site cooling water only, excluding training.
- Per-prompt, operator disclosure: Sam Altman, “The Gentle Singularity,” June 2025 — opened directly for this briefing. Verified as stated there: an average ChatGPT query at about 0.34 watt-hours and about 0.000085 gallons of water, described as roughly one fifteenth of a teaspoon. No methodology or boundary is published with the figure; the milliliter conversion (about 0.32 milliliters) is computed on this page. Carried as an estimate.
- Per-prompt, independent model: Li, Yang, Islam, and Ren, “Making AI Less Thirsty: Uncovering and Addressing the Secret Water Footprint of AI Models,” arXiv:2304.03271 (first posted 2023; version 5 revised March 26, 2025; accepted by Communications of the Association for Computing Machinery) — opened directly for this briefing. Verified from the paper: a 500-milliliter bottle per roughly 10 to 50 medium-length GPT-3 responses; the United States average table entry of 2.200 milliliters on site plus 14.704 milliliters off site, totaling 16.904 milliliters per request, at an assumed 0.004 kilowatt-hours of server energy per request; location entries from 7.107 milliliters (Ireland) to 47.506 milliliters (Washington); GPT-3 training modeled at 700,000 liters on site and 5.4 million liters in total in Microsoft's United States data centers; global AI water withdrawal projected at 4.2 to 6.6 billion cubic meters in 2027, with 0.38 to 0.60 billion cubic meters as consumption. All figures from this study are modeled estimates, labeled as such on this page.
- Per-prompt, lifecycle analysis: Mistral AI, “Our contribution to a global environmental standard for AI,” July 22, 2025, reporting its lifecycle analysis with Carbone 4 and ADEME (the French Agency for Ecological Transition), peer-reviewed by Resilio and Hubblo — opened directly for this briefing. Verified from Mistral's publication: a 400-token Le Chat response, excluding users' terminals, at 45 milliliters of water and 1.14 grams of carbon dioxide equivalent, including upstream impacts such as server manufacturing; Mistral Large 2's total at 281,000 cubic meters of water consumed as of January 2025 after 18 months of usage. Carried on this page as an estimate on a full lifecycle boundary.
- Search record: the target phrase for this briefing — “how much water does ai use” — was returned by Google autocomplete on November 18, 2026, with the variants “per day,” “per year,” “per prompt,” and “per question,” and from the ChatGPT seed as “how much water does chatgpt use per question / per prompt / to generate an image,” as recorded in this site's topic queue. No search volume is stated for it; none has been verified. No per-image or per-video water figure is stated on this page; no primary source quantifying one was found in this run, and the absence is reported rather than filled.
Keep going
- How much water does a data center use? — the full data center guide
- Data center vs. golf course: the comparison everyone is actually searching — yesterday's briefing
- Florida's new data center water law (SB 484), in plain words
- Is Florida running out of water? — the full scarcity guide
- Who owns the water? — the full ownership guide
- Your city's water quality report
- Find a license-verified water treatment provider
- More Latest briefings