The unsettling a part of AI’s water footprint will not be that one chatbot reply drains a reservoir. It doesn’t. The unsettling half is {that a} tiny, hidden price turns into much less tiny when it’s repeated throughout hundreds of thousands or billions of requests, then added to the water wanted to construct and energy the info facilities behind them.
A 2024 Washington Post analysis, produced with researchers on the College of California, Riverside, estimated the water and electrical energy wanted for ChatGPT utilizing GPT-4 to put in writing a median 100-word e-mail at a median American knowledge middle. The Put up framed the end result as roughly a bottle of water per e-mail.
That quantity is helpful, however solely whether it is learn rigorously. It isn’t a everlasting meter connected to each ChatGPT reply. Water use adjustments with the mannequin, the info middle, the native climate, the cooling system, the electrical energy provide and the accounting methodology. The identical immediate can have a special footprint relying on the place and when it’s processed.
The broader scientific work behind many of those estimates comes from Pengfei Li, Jianyi Yang, Mohammad A. Islam and Shaolei Ren. Their paper, Making AI Less “Thirsty”, was first posted to arXiv in 2023 and later accepted by Communications of the ACM. It estimated {that a} mannequin comparable to GPT-3 may devour about 500 milliliters of water for roughly 10 to 50 medium-length responses, relying on the place and when it was deployed.
That vary issues. It’s the distinction between treating water use as a single common reality and treating it as an infrastructure drawback with geography inside it.
What counts as AI water use?
Knowledge facilities use water in two fundamental methods. The primary is direct water use, usually for cooling. Servers generate warmth. In lots of services, evaporative cooling methods use water to hold that warmth away. A few of that water is consumed as a result of it evaporates into the air.
The second is oblique water use. Electrical energy technology can require water, particularly in thermal energy crops that use water for steam cycles or cooling. An information middle that appears water-efficient on web site should still be tied to water use elsewhere by the grid that powers it.
This is the reason the language in these research is cautious. Water withdrawal means water taken from a supply, comparable to a river, aquifer or municipal system. Water consumption often means water faraway from quick reuse, usually by evaporation. Each matter, however they aren’t the identical quantity.
Li and colleagues projected that world AI demand may account for 4.2 to six.6 billion cubic meters of water withdrawal in 2027 below the eventualities they examined. Within the summary, they in contrast that vary with the full annual water withdrawal of a number of Denmarks or about half of the UK.
That may be a model-based projection, not a measurement of all AI methods right this moment. It shouldn’t be learn as future. But it surely provides scale to an issue that’s in any other case straightforward to cover inside the graceful floor of a chat window.
Why estimates disagree
Public estimates of AI water use differ sharply. In 2025, a Google-authored arXiv paper, Measuring the environmental impact of delivering AI at Google Scale, reported that the median Gemini Apps textual content immediate used 0.24 watt-hours of vitality and 0.26 milliliters of water below Google’s accounting framework. The authors mentioned that was the equal of about 5 drops of water.
That determine is way smaller than the bottle-scale estimates usually cited for ChatGPT. The hole doesn’t essentially imply one quantity is solely true and the opposite false. The research aren’t measuring the identical factor. They contain completely different methods, time intervals, assumptions, workloads and bounds.
Google’s paper checked out serving Gemini textual content prompts inside Google’s manufacturing infrastructure. The Li and Ren work tried to estimate a broader AI water footprint, together with direct and off-site water use. The Washington Put up calculation targeted on GPT-4 producing a 100-word e-mail at a median U.S. knowledge middle.
The lesson will not be that AI makes use of both 5 drops or one bottle. The lesson is that with out clear, comparable reporting, the general public is left evaluating not like numbers.
The native drawback
Water will not be like carbon dioxide. A ton of carbon dioxide has a worldwide impact regardless of the place it’s launched. A liter of water taken in a moist area will not be equal to a liter taken from an aquifer below stress.
That’s the reason the situation of AI infrastructure issues. A 2026 Guardian analysis reported that 517 of 809 deliberate U.S. knowledge facilities had been in places that had been in drought circumstances in the course of the earlier 12 months. The article additionally famous that an earlier model had incorrectly paraphrased the 500-milliliter estimate as making use of to every 100-word immediate, then corrected the textual content to match the underlying research’s vary of roughly 10 to 50 medium-length GPT-3 responses.
That correction is price mentioning as a result of it reveals how simply these numbers can slide. The native danger stays actual even when the per-prompt shorthand is just too neat. If a big knowledge middle is in-built a dry area and makes use of water-based cooling, the query will not be solely how a lot water a person consumer consumed. It’s whether or not the power is competing with households, farms, rivers or groundwater methods in a spot already below stress.
Related Press reporting in 2023 discovered that Microsoft-backed OpenAI infrastructure in West Des Moines, Iowa, had drawn consideration as a result of Microsoft services used about 6 p.c of the native water district’s provide in July 2022, a month earlier than OpenAI completed coaching GPT-4. Microsoft advised the AP it was working to cut back useful resource depth.
Not only a chatbot behavior
It’s tempting to show the difficulty right into a easy private rule: write fewer AI emails, save water. There’s some fact in that. Shorter prompts, shorter solutions and smaller fashions can scale back useful resource use. However particular person restraint can not substitute for infrastructure disclosure.
Most individuals can not select which knowledge middle handles a question. They can not inform whether or not a response got here from a water-cooled facility in a dry area or a extra environment friendly system in a water-secure one. They can not see whether or not the electrical energy behind the request carried its personal water footprint.
A 2026 arXiv paper by Yuelin Han, Pengfei Li, Adam Wierman and Shaolei Ren, Small Bottle, Big Pipe, argued that U.S. knowledge facilities may require a whole bunch of hundreds of thousands of gallons per day of recent water capability by 2030 if 2024 water-use depth persists. The authors framed the difficulty as a public water-system constraint, not only a personal effectivity drawback.
That framing is vital. An information middle doesn’t solely devour assets after it’s constructed. It asks a group to order water capability for peak demand, usually in the course of the hottest days of the 12 months, when cooling demand is excessive and public methods could already be strained.
The actual query
The purpose will not be that each use of AI is reckless. The purpose is that the business has grown quicker than public understanding of its bodily calls for.
AI feels weightless as a result of its interface is textual content on glass. The equipment will not be weightless. It’s chips, servers, cooling towers, substations, energy crops, provide chains and native water methods. A 100-word e-mail is small. The infrastructure constructed to reply hundreds of thousands of such requests will not be.
The bottle-of-water comparability works as a warning, but it surely shouldn’t be handled as a common conversion desk. The higher query is extra concrete: the place is the computation occurring, what mannequin is getting used, what water is being counted, and who else will depend on the identical provide?
Till these solutions are reported persistently, the general public will preserve seeing AI as a clear digital service whereas communities close to the pipes, pumps and cooling methods cope with the bodily price.