How much water does AI use? A single text prompt uses anywhere from about 0.26 mL, Google’s official figure for a median Gemini prompt, to a few milliliters for a typical chatbot reply. The scary viral ~500 mL “per email” number comes from also counting the electricity supply chain, which is why estimates span more than 1,000x. At scale it adds up, with US data centers consuming about 17 billion gallons of water directly in 2023 and that figure projected to double or even quadruple by 2028.
This guide gives you verified numbers for every version of the question, so per prompt, per question, per day, per year, and per data center. Every figure is dated and attributed to its source. You will see why credible estimates differ so wildly, which viral claims were debunked, and how AI water use compares to a burger or a cotton t-shirt. We lead with data, not panic, and we tell you what the data does not yet show.
The Key Takeaways
- Per prompt: ~0.26 mL to a few mL. Google’s official 2025 figure for a median Gemini text prompt is 0.26 mL (about five drops); independent estimates for heavier models land in the low tens of milliliters.
- The “bottle of water per email” figure is ~500 mL, from a 2023 UC Riverside study, but it bundles in the water used to generate the electricity, which is why it is more than 1,000x higher than cooling-only numbers.
- US data centers used about 17 billion gallons directly in 2023, plus an estimated 211 billion gallons indirectly through power generation; direct use is projected to double or quadruple by 2028.
- xAI’s Grok supercomputer in Memphis can draw around 1 million gallons of drinking water a day, with projected demand near 5 million, comparable to a town of up to 50,000 people.
- The single biggest lever you control is model choice. A lean model on a simple task can use a fraction of the water of a frontier model answering the same prompt.
How Much Water Does AI Use Per Prompt?
Estimates range from about 0.26 mL per prompt (Google’s official 2025 figure for a median Gemini text prompt, roughly five drops) to around 500 mL for a session of 10 to 50 questions (a 2023 UC Riverside study). The gap exists because the low number counts only data-center cooling, while the high number adds the water used to generate the electricity that powers the chips. A realistic figure for a single, ordinary chatbot prompt sits in the low single-digit milliliters, with heavier reasoning models pushing into the low tens.
| Model / estimate | Energy per prompt | Water per prompt | What’s counted | Source and year |
|---|---|---|---|---|
| Google Gemini (median text prompt) | 0.24 Wh | ~0.26 mL (≈5 drops) | Cooling + full data-center overhead | Google, Aug 2025 |
| Mistral (Le Chat, ~400-token reply) | n/a | ~45 mL | First-party LCA (training + inference + hardware) | Mistral / Carbone 4, Jul 2025 |
| ChatGPT, GPT-4o (medium reply) | ~0.3 Wh | ~1.5 to 4 mL | On-site data-center cooling | Epoch AI / A. Masley, 2025 |
| ChatGPT, GPT-5 (medium reply) | ~18 Wh (up to 40) | ~a few to tens of mL | On-site cooling estimate | Univ. of Rhode Island, 2025 |
| ChatGPT (Altman’s stated average) | 0.34 Wh | ~0.32 mL (1/15 tsp) | Unspecified “average” | S. Altman, 2025 |
| GPT-3 era (“bottle” figure) | n/a | ~500 mL per 10 to 50 replies | Cooling + electricity (lifecycle) | UC Riverside, 2023 to 2025 |
| Claude (Anthropic, medium reply) | ~0.8 to 5.5 Wh | No first-party data | Third-party energy estimate; no water disclosed | Jegham et al., 2025 |
| DeepSeek-R1 (reasoning reply) | ~24 to 34 Wh | No first-party data | Third-party energy estimate; among the heaviest | Jegham et al., 2025 |
| Meta Llama 3.x (medium reply) | ~0.25 to 1.6 Wh | No first-party data | Third-party energy estimate; no water disclosed | Jegham et al., 2025 |
Google Gemini: about 0.26 mL per prompt
In August 2025 Google published the most detailed first-party measurement so far, based on production data from May 2025. It found that a median Google Gemini text prompt uses 0.24 watt-hours of energy, emits 0.03 grams of CO2 equivalent, and consumes 0.26 milliliters of water, which Google describes as about five drops. Crucially, that figure is not a cherry-picked best case; Google’s method includes cooling, idle reserve capacity, CPU and RAM, and data-center overhead, which makes it one of the more honest single numbers available. You can read the methodology in Google’s environmental impact report.
ChatGPT: a few milliliters to a few dozen
OpenAI has not published a Google-style breakdown, so ChatGPT figures are independent estimates. A medium GPT-4o reply of 150 to 200 words works out to about 0.3 watt-hours and roughly 1.5 to 4 mL of cooling water. A comparable GPT-5 reply is far heavier, at about 18 watt-hours (up to 40) and an estimated few to tens of mL.
OpenAI CEO Sam Altman has claimed the average query uses “roughly one fifteenth of a teaspoon,” about 0.32 mL. He did not define “average” or say whether training was included, which is why independent researchers treat that figure cautiously. We will cover ChatGPT in depth in a dedicated companion piece; for the broad picture, newer reasoning models are thirstier per answer, not leaner.
Mistral, Claude, DeepSeek and Llama: what each lab actually discloses
Only two labs publish a first-party water number. Google gives 0.26 mL for a median Gemini prompt, and Mistral published a full lifecycle analysis in July 2025, audited with Carbone 4 and France’s ADEME, putting one ~400-token Le Chat reply at about 45 mL of water and 1.14 g of CO2e. For Claude, DeepSeek and Meta Llama there is no first-party water figure at all; the only numbers available are third-party energy estimates that researchers then back-convert to water using regional cooling factors.
That gap matters for one popular myth. DeepSeek is often called “efficient,” but that reputation is about its training cost; an independent 2025 benchmark found DeepSeek-R1 among the thirstiest models at inference, at roughly 24 to 34 watt-hours per reasoning answer, while a lightweight non-reasoning model can sit under 1 watt-hour. “Lean” describes how a model works on a given task, not a brand name, which is exactly why your choice of model and mode moves the water number more than anything else you do.
The “bottle of water per email” claim: about 500 mL
The number that went viral comes from “Making AI Less Thirsty,” a 2023 study by Pengfei Li and colleagues at UC Riverside, later popularized by the Washington Post as roughly 519 mL to draft a 100-word email with GPT-4. The Lincoln Institute restated it in October 2025 as “a chat session of about 20 queries uses up to a bottle of freshwater.” It is not wrong, but it measures a much wider system than cooling alone, which is exactly why it dwarfs Google’s 0.26 mL.
The same study estimated that training GPT-3 in Microsoft’s US data centers directly evaporated about 700,000 liters of clean freshwater. You can check the figures in the original UC Riverside paper.
Why Do AI Water Estimates Vary So Wildly?
If you have seen “five drops” and “a whole bottle” for the same prompt and assumed someone is lying, neither side is. The numbers differ by more than 1,000x because they measure different things, and once you know the boundary, the figures stop fighting each other.
On-site water is what a data center evaporates to cool its chips. Indirect water is what power plants consume to make the electricity those chips run on, which is often larger than the cooling itself. The low estimates count only the first; the viral high estimates count both, plus chip manufacturing in some cases. Jonathan Koomey of Lawrence Berkeley National Laboratory argues the indirect electricity water should not sit in a data center’s footprint at all, since no other industry is measured that way.
Methodology matters enough to flip headlines. In December 2025, investigative reporting reviewed a widely shared claim that a Google data center in Chile would use more than a thousand times a city’s water. It found the figure was off by a factor of 1,000 due to a unit error, though the facility still requested more than a local population’s residential use. The lesson is not that AI water use is fake; it is that you should always ask what is being counted, where, and when before repeating a number.
How Much Water Do AI Data Centers Use Per Day?
Per prompt the numbers feel trivial. But data centers run billions of prompts, so the daily totals are where the impact becomes concrete. A typical data center uses around 300,000 gallons of water per day, about the same as 1,000 households. A large facility can use up to 5 million gallons per day, comparable to a town of up to 50,000 people.
Specific sites bring it home. Google’s thirstiest single site, in Iowa, used roughly 2.7 million gallons per day in 2024. A Meta facility in Newton County, Georgia, draws about 500,000 gallons per day, around 10% of that county’s water.
Aggregated, the scale is national. US data centers directly consumed about 17 billion gallons in 2023 and an estimated 211 billion gallons indirectly through electricity generation, a figure the LBNL report projects could double or quadruple by 2028. Google alone reported using more than 5 billion gallons across its data centers in 2023, with 31% drawn from water-stressed watersheds. Microsoft’s reported water use climbed from 6.4 million cubic meters in 2022 to 7.8 million in 2023, about a 22% jump in a single year as it expanded AI infrastructure.
How Much Water Does AI Use Per Day and Per Year?
There is no official global annual total, and anyone quoting one precisely is guessing. Credible 2025 estimates put AI-related data-center water consumption in the hundreds of billions of liters range and rising fast. The UC Riverside team projects global AI water withdrawal of 4.2 to 6.6 billion cubic meters by 2027, which is more than the total annual water withdrawal of several small countries. The trajectory, not any single year’s figure, is the real story.
| Scope | 2023 | 2025 | 2028 (projected) | Source |
|---|---|---|---|---|
| US data centers, direct (cooling) | ~17B gal | rising | 2x to 4x (≈34 to 68B gal) | LBNL 2024 report |
| US data centers, indirect (electricity) | ~211B gal | rising | higher | LBNL 2024 report |
| Texas data centers (state only) | n/a | ~49B gal | ~399B gal by 2030 | HARC / Univ. Houston |
| Google (all data centers, direct) | >5B gal | n/a | n/a | Google 2023 report |
| Global AI water withdrawal | n/a | hundreds of billions of L | 4.2 to 6.6B m³ by 2027 | UC Riverside |
Why Does AI Need Water at All?
Servers running AI models generate intense heat, and many large data centers cool them by evaporating water, the same physics that makes sweating cool your skin. That evaporated water leaves the local system, which is why researchers track it as consumption, not just withdrawal. A lot of it is treated, drinking-quality water, although operators are increasingly shifting to reclaimed or non-potable sources.
There is also a hidden upstream cost. Manufacturing the chips themselves is extraordinarily water-intensive, with a single advanced semiconductor fab using on the order of 10 million gallons of ultrapure water per day. Training a model is a one-time spike, but inference, the everyday business of answering your prompts, is now the dominant and fastest-growing share of AI’s water demand. Understanding how AI systems actually work under the hood makes the cooling bill a lot less mysterious.
Grok and the Memphis Water Fight
The clearest example of AI water becoming a local problem is xAI’s Colossus supercomputer in Memphis, which trains and runs Grok. Memphis Light, Gas and Water has confirmed the site can draw on the order of 1 million gallons of drinking water per day from the Memphis Sand aquifer to cool its chips, with projected demand climbing toward 5 million gallons a day as the campus expands. That aquifer is also the city’s primary source of drinking water, which is why local utilities, the NAACP and the Southern Environmental Law Center have all pushed back.
xAI broke ground in October 2025 on an $80 million water recycling plant designed to supply up to 13 million gallons a day of treated wastewater instead, which MLGW estimates would spare roughly 5 billion gallons of aquifer water a year. As of April 2026, Elon Musk has put that plant on hold, saying xAI needs to finish its second site, Colossus 2, first, while Memphis officials are publicly pressing for a restart. It is the single best illustration of this article’s core point, that the water question is rarely about your prompt and almost always about where the data center sits.
Is AI’s Water Use Actually a Big Deal?
Context matters in both directions. Per prompt, AI is a rounding error against everyday consumption. A single beef burger takes more than 600 gallons of water to produce, a cotton t-shirt more than 700 gallons, and a single US golf course can use 100,000 to 2 million gallons a day. By that yardstick, asking a chatbot a question is negligible.
The legitimate concern is not the per-prompt cost; it is concentration and timing. Data centers cluster in a handful of regions, often water-stressed ones, and draw heavily at specific times. A facility pulling millions of gallons a day in Arizona or Texas is a local problem even if it is a global rounding error. Both things are true at once, and a good answer to “is this a big deal” is “not for your usage, yes for where the buildout is happening.”
What You Can Actually Do About It
The single most effective lever an individual controls is model choice, because the spread between models is enormous. Sending a simple question to a heavyweight reasoning model can use roughly ten times the energy and water of a lean model giving an equally good answer. Matching the model to the task does more than any amount of prompt-trimming. Comparisons like DeepSeek vs ChatGPT and our guide to the most capable AI models show how widely those profiles vary, and why a model’s “efficient” reputation usually describes its training cost, not the energy it burns answering you.
Practically, that means defaulting to a lighter model for everyday questions and reserving frontier reasoning models for problems that actually need them. Batch related questions into one session instead of many cold starts, and do not regenerate answers reflexively. Running several models from one place makes “pick the right tool” realistic instead of theoretical, which is the whole point of a multi-model setup like the one in our getting started guide. None of this requires guilt; it just requires using the heavy option deliberately rather than by default.
Conclusion
So how much water does AI use? Per prompt, realistically a few milliliters; Google’s official Gemini figure is 0.26 mL, and the viral 500 mL claim is a much broader lifecycle measure, not a contradiction. At scale the numbers are large and growing fast, with US direct data-center use at about 17 billion gallons in 2023 and projected to double or quadruple by 2028, concentrated where water is scarcest. The most useful next step is not to stop using AI; it is to be deliberate about which model you reach for, starting with a clear-eyed look at how the leading models actually differ.
FAQ
Does writing an email with AI really use a whole bottle of water?
That figure, roughly 519 mL, comes from a 2023 UC Riverside study and counts cooling plus the water used to generate the electricity. Google’s official cooling-inclusive figure for a median prompt is about 0.26 mL, so the honest answer is “it depends entirely on what you count,” with a typical prompt closer to a few milliliters.
Does AI actually destroy the water, or is it returned?
Most data-center cooling water is evaporated, so it leaves the local watershed rather than being returned downstream, which is why researchers measure it as consumption. Some of it is treated drinking water, though operators are shifting toward reclaimed and non-potable sources.
Does training an AI model use more water than chatting with it?
Training is a large one-time spike, for example roughly 700,000 liters to train GPT-3 in US data centers, but the everyday cost is inference. As of 2026, answering user prompts is the dominant and fastest-growing share of AI water demand.
Which company uses the most water for AI?
By disclosed totals, Google reported more than 5 billion gallons across its data centers in 2023 and Microsoft about 7.8 million cubic meters (roughly 2 billion gallons) the same year, but direct comparison is unfair because companies count and disclose differently. Single facilities matter more locally than corporate totals do.
What is the simplest way to reduce my own AI water footprint?
Use a lighter model for routine questions and save frontier reasoning models for tasks that need them, since the per-prompt difference between models is far larger than anything else you control. Batching questions into one session also helps more than shortening prompts.
