How Much Water Does AI Use?

Why Is There No Single Answer?

Ask how much water does AI use and you can find answers from a quarter of a milliliter to a full bottle. Those numbers describe different models, in different years and places, and they count different water. Three questions explain most of the gap: which water is counted, how much electricity is assumed, and where the computers are.

Direct Water vs. Indirect Water

Direct water is consumed on site, mainly to cool a data center. Indirect water is consumed at the power plants that generate its electricity.^[9]

Indirect water is usually the larger share. Researchers Shaolei Ren of UC Riverside and Amy Luers of Microsoft write that it often makes up 80 percent or more of the total.^[11] A Lawrence Berkeley National Laboratory report estimated that all U.S. data centers, not only AI, consumed 66 billion liters directly in 2023 and nearly 800 billion liters indirectly. That indirect estimate assumes each data center runs on its local grid mix, and the authors note that power purchase agreements or on-site generation could change it significantly.^[8]

In this 2023-era model, the same GPT-3 response takes 7.1 mL in Ireland and 47.5 mL in Washington state, and power generation is the larger part in every location shown.^[1]

Withdrawal vs. Consumption

These two words can change a number by more than ten times. Withdrawal is water taken from a source. Consumption is withdrawal minus what is discharged back, which is mostly water lost to evaporation.^[1] The USGS describes consumptive use as the share unavailable for immediate use.^[10]

• U.S. thermoelectric power plants withdrew 133 billion gallons a day in 2015 and consumed 4.31 billion, about 3 percent.^[10]
• At data centers, the share consumed is higher. Ren and Luers say companies have reported 45 to 60 percent,^[11] and Google reports about 80 percent for itself.^[5]
• For global AI demand in 2027, one preprint projects 4.2 to 6.6 billion cubic meters withdrawn, of which 0.38 to 0.60 billion evaporates.^[1]

The per-prompt figures from Li and colleagues and from Google are consumption. For a large total, check which of the two you are reading.

Where Did the "Bottle of Water" Claim Come From?

The bottle has two origins, and neither one describes an average chatbot prompt.

The first is a 2023 preprint by Pengfei Li, Ren and colleagues, later accepted by Communications of the ACM. It estimated that GPT-3 consumes a 500 mL bottle for roughly 10 to 50 medium-length responses, depending on when and where it runs. It assumed 4 watt-hours of server energy per request, based on an estimate in OpenAI's GPT-3 paper, and the authors said actual consumption could be higher.^[1]

The second came in 2024. Fortune, citing the Washington Post and UC Riverside, reported that GPT-4 consumes 519 mL, just over one bottle, to write a 100-word email.^[2] We could not open the Post's article, so we have not read its method.

The estimate has since been qualified. In May 2026, writer Andy Masley published an account of an email exchange with Ren. He quotes Ren saying the 2024 estimate was "time-specific, assumption-based" and "should not be used to describe general AI/ChatGPT or today's optimized systems." Masley reports Ren's current estimate for an average chatbot prompt as about 15 mL including off-site water, and about 5 mL inside the data center.^[3] That is a secondhand report, not a published paper.

Evidence and Numbers: What Has Been Published?

The table is not a ranking, because the rows are not measuring the same thing. Here is what separates them:

• Scope. A paper in the journal Patterns notes that Google's report left out indirect water, saying it "does not fully control the water consumption in electricity generation."^[9] As a rough sum of our own, 0.24 Wh^[5] times the U.S. data center average of 4.52 liters per kilowatt-hour^[8] adds about 1.1 mL, for about 1.3 mL in total. That mixes a U.S. grid average with Google's own fleet, so it only shows the scale of the omission.
• Energy. Li and colleagues assumed 4 Wh per request. Google reports 0.24 Wh for its median prompt in May 2025, after what it says was a 33-fold drop in a year.^[5]
• Measurement boundary. Google says a narrower method would give 0.10 Wh and 0.12 mL for the same prompt.^[5]
• Hosting. Jegham and colleagues model DeepSeek-R1 at 34 mL per long prompt on Azure and over 200 mL on DeepSeek's own servers.^[7]
• Task size. Google's paper does not state the length of its median prompt, so it cannot be lined up neatly against a 150 to 300 word output.^[5]

Per-prompt numbers also do not describe totals or local effects. One peer-reviewed estimate puts AI's global direct and indirect water footprint at 312.5 to 764.6 billion liters in 2025, with large uncertainty.^[9] Ren and Luers put direct U.S. data center use at about 0.3 percent of the public water supply, and say it can still strain local supplies where facilities cluster in dry areas.^[11]

Myths vs Facts

What We Still Can't Tell You

We cannot tell you how much water your own last prompt used. That depends on the model, prompt length, data center and power grid, and users cannot see those.

We ran no measurements. Every figure here comes from the sources cited, and most are modeled or self-reported. Google's and Mistral's results rest on company data that outsiders cannot fully check, and Altman's post gives no method.^[6] The Patterns paper says no company reports AI-specific metrics in its environmental disclosures, so outside estimates lean on company-wide figures.^[9]

Most per-prompt figures cover text only, so we have no comparable numbers here for image or video generation. Google's figure also leaves out training.^[5] Older estimates may not fit current models, and Ren and Luers say efficiency gains have likely reduced them but offer that as a likelihood, not a measurement.^[11]