July 10, 2026

Representative data center image. Photo by panumas nikhomkhai on Pexels.

The unsettling part of AI’s water footprint is not that one chatbot answer drains a reservoir. It does not. The unsettling part is that a tiny, hidden cost becomes less tiny when it is repeated across millions or billions of requests, then added to the water needed to build and power the data centers behind them.

A 2024 Washington Post analysis, produced with researchers at the University of California, Riverside, estimated the water and electricity needed for ChatGPT using GPT-4 to write an average 100-word email at an average American data center. The Post framed the result as roughly a bottle of water per email.

That number is useful, but only if it is read carefully. It is not a permanent meter attached to every ChatGPT answer. Water use changes with the model, the data center, the local weather, the cooling system, the electricity supply and the accounting method. The same prompt can have a different footprint depending on where and when it is processed.

The broader scientific work behind many of these 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 model such as GPT-3 could consume about 500 milliliters of water for roughly 10 to 50 medium-length responses, depending on where and when it was deployed.

That range matters. It is the difference between treating water use as a single universal fact and treating it as an infrastructure problem with geography inside it.

What counts as AI water use?

Data centers use water in two main ways. The first is direct water use, often for cooling. Servers generate heat. In many facilities, evaporative cooling systems use water to carry that heat away. Some of that water is consumed because it evaporates into the air.

The second is indirect water use. Electricity generation can require water, especially in thermal power plants that use water for steam cycles or cooling. A data center that looks water-efficient on site may still be tied to water use elsewhere through the grid that powers it.

This is why the language in these studies is careful. Water withdrawal means water taken from a source, such as a river, aquifer or municipal system. Water consumption usually means water removed from immediate reuse, often through evaporation. Both matter, but they are not the same number.

Li and colleagues projected that global AI demand could account for 4.2 to 6.6 billion cubic meters of water withdrawal in 2027 under the scenarios they examined. In the abstract, they compared that range with the total annual water withdrawal of several Denmarks or about half of the United Kingdom.

That is a model-based projection, not a measurement of all AI systems today. It should not be read as destiny. But it gives scale to a problem that is otherwise easy to hide inside the smooth surface 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 text prompt used 0.24 watt-hours of energy and 0.26 milliliters of water under Google’s accounting framework. The authors said that was the equivalent of about five drops of water.

That figure is far smaller than the bottle-scale estimates often cited for ChatGPT. The gap does not necessarily mean one number is simply true and the other false. The studies are not measuring the same thing. They involve different systems, time periods, assumptions, workloads and boundaries.

Google’s paper looked at serving Gemini text prompts inside Google’s production infrastructure. The Li and Ren work tried to estimate a broader AI water footprint, including direct and off-site water use. The Washington Post calculation focused on GPT-4 producing a 100-word email at an average U.S. data center.

The lesson is not that AI uses either five drops or one bottle. The lesson is that without transparent, comparable reporting, the public is left comparing unlike numbers.

The local problem

Water is not like carbon dioxide. A ton of carbon dioxide has a global effect no matter where it is released. A liter of water taken in a wet region is not equivalent to a liter taken from an aquifer under stress.

That is why the location of AI infrastructure matters. A 2026 Guardian analysis reported that 517 of 809 planned U.S. data centers were in locations that had been in drought conditions during the previous year. The article also noted that an earlier version had incorrectly paraphrased the 500-milliliter estimate as applying to each 100-word prompt, then corrected the text to match the underlying study’s range of roughly 10 to 50 medium-length GPT-3 responses.

That correction is worth mentioning because it shows how easily these numbers can slide. The local risk remains real even when the per-prompt shorthand is too neat. If a large data center is built in a dry region and uses water-based cooling, the question is not only how much water an individual user consumed. It is whether the facility is competing with households, farms, rivers or groundwater systems in a place already under pressure.

Associated Press reporting in 2023 found that Microsoft-backed OpenAI infrastructure in West Des Moines, Iowa, had drawn attention because Microsoft facilities used about 6 percent of the local water district’s supply in July 2022, a month before OpenAI finished training GPT-4. Microsoft told the AP it was working to reduce resource intensity.

Not just a chatbot habit

It is tempting to turn the issue into a simple personal rule: write fewer AI emails, save water. There is some truth in that. Shorter prompts, shorter answers and smaller models can reduce resource use. But individual restraint cannot substitute for infrastructure disclosure.

Most people cannot choose which data center handles a query. They cannot tell whether a response came from a water-cooled facility in a dry region or a more efficient system in a water-secure one. They cannot see whether the electricity behind the request carried its own water footprint.

A 2026 arXiv paper by Yuelin Han, Pengfei Li, Adam Wierman and Shaolei Ren, Small Bottle, Big Pipe, argued that U.S. data centers could require hundreds of millions of gallons per day of new water capacity through 2030 if 2024 water-use intensity persists. The authors framed the issue as a public water-system constraint, not just a private efficiency problem.

That framing is important. A data center does not only consume resources after it is built. It asks a community to reserve water capacity for peak demand, often during the hottest days of the year, when cooling demand is high and public systems may already be strained.

The real question

The point is not that every use of AI is reckless. The point is that the industry has grown faster than public understanding of its physical demands.

AI feels weightless because its interface is text on glass. The machinery is not weightless. It is chips, servers, cooling towers, substations, power plants, supply chains and local water systems. A 100-word email is small. The infrastructure built to answer millions of such requests is not.

The bottle-of-water comparison works as a warning, but it should not be treated as a universal conversion table. The better question is more concrete: where is the computation happening, what model is being used, what water is being counted, and who else depends on the same supply?

Until those answers are reported consistently, the public will keep seeing AI as a clean digital service while communities near the pipes, pumps and cooling systems deal with the physical cost.

Produced with AI assistance. Reviewed by the ScienceBlog.com editorial team before publication.