How Much Water Does AI Use? The Real Environmental Cost of AI

Quick Answer: How Much Water Does AI Use?

How much water does AI use? There is no single number, and anyone who quotes one without saying what it includes is oversimplifying. The honest answer has three parts.

• Per prompt, the figures are small. Google reports that a median text prompt to its Gemini app uses about 0.26 milliliters of water, roughly five drops, and OpenAI's chief executive has said an average ChatGPT query uses about 0.000085 gallons, about 0.3 milliliters. Those are company figures with limited scope, and image and video generation use far more energy.
• In total, the figures are large. US data centers directly consumed roughly 17 billion gallons of water in 2023 according to Lawrence Berkeley National Laboratory, and the water used indirectly to generate their electricity is estimated to be many times higher.
• Where the water comes from matters. A figure that is small nationally can be significant in a drought-prone area, and the data on exactly how much each facility uses is often incomplete.

The rest of this guide explains where AI's water use comes from, what the best available estimates say, why they disagree, and what you can do with the information.

Introduction

When people ask how much water AI uses, they usually have one of two questions in mind. One is personal: does a chatbot conversation have a meaningful environmental cost? The other is societal: is the growth of AI data centers straining water supplies?

Both questions deserve a straight answer, and both are harder to answer than headlines suggest, because the environmental impact of AI is measured in different ways by different groups. Water use in computing is not reported in a consistent way. Companies publish some figures, researchers estimate others, and results depend on whether you count water used inside the data center, water used by power plants that supply it, or both. This article separates what is known from what is estimated, links the sources so you can check them, and flags where the numbers are contested.

It covers how data centers use water, what researchers and companies say about training and everyday use, the national and local picture, why estimates conflict, what the industry is doing about it, and practical steps for individuals and organizations. It also lists questions you can ask any AI provider about water, and a glossary of the terms you will meet.

How AI Uses Water: The Basics

AI models run in data centers, buildings full of servers that produce a great deal of heat. That heat has to be removed, and water is one of the ways. There are two broad routes by which AI is connected to water.

Direct water use

Direct use is water consumed on site, mostly for cooling. In a common design, water absorbs heat from the equipment and then evaporates, carrying the heat away in the way sweat cools skin. Researchers at the University of Wisconsin-Milwaukee describe it this way: in evaporative systems, pumps push cold water through pipes in the data center, where it absorbs heat and vents as steam, while closed-loop systems recirculate cooled water but need more energy to run chillers (The Conversation, August 2025).

Indirect water use

Indirect use is the water consumed to produce the electricity that powers the data center. Thermal power plants use water for cooling, and hydropower reservoirs lose water to evaporation. The same article reports that Lawrence Berkeley National Laboratory estimated indirect consumption at about 12 times the direct use for cooling in 2023. Indirect figures depend heavily on which power sources are counted and how, which is one reason estimates differ.

Withdrawal versus consumption

The terms matter. Withdrawal is water taken from a source such as a river, lake or utility. Consumption is the part that is not returned to the same source, mostly because it evaporates. A facility can withdraw a lot of water and return most of it, or withdraw less and lose nearly all of it to evaporation. When you read a number, check which one it is. Most of the AI figures in this article are about consumption.

How Data Center Cooling Works

Cooling is where most direct water use comes from, so it helps to know the main approaches. The table is a general overview, not a ranking, and real facilities often combine methods.

There is a real trade-off. Using less water for cooling often means using more electricity, and electricity has its own water and carbon footprint. The right design depends on the local climate, the local water supply and the local power mix, which is why a single global number is not meaningful.

Microsoft has said it is moving to a closed-loop, chip-level liquid cooling design that does not use evaporation for cooling, with trade press reporting savings of more than 125 million liters of water per facility per year, and testing in Phoenix, Arizona and Mount Pleasant, Wisconsin (Data Centre Magazine). That is a company claim reported by trade media, so treat the figure as indicative until it is borne out in operation. The same reports say Microsoft's water usage effectiveness improved from 0.49 liters per kilowatt-hour in 2021 to 0.30 in 2024.

How Much Water Does Training an AI Model Use?

Training a large model takes weeks of heavy computing, and the cooling water used during that time has been studied.

In a widely cited paper, "Making AI Less 'Thirsty'", Pengfei Li, Jianyi Yang, Mohammad Islam and Shaolei Ren estimated that training the GPT-3 language model in Microsoft's US data centers could directly evaporate 700,000 liters of clean freshwater (arXiv). The paper was later accepted by Communications of the ACM and has been revised several times, with the latest version dated March 2025.

A few cautions about that number:

• It is an estimate for one model, trained in specific facilities, under assumptions about cooling and location.
• It describes direct on-site evaporation, not the water behind the electricity.
• Newer models differ in size, hardware and efficiency, so it cannot simply be scaled up or down.
• Training happens once, while serving a model to millions of people happens continuously. For a popular model, everyday use may add up to more than training over time, although there is no reliable public figure for any specific system.

How Much Water Does a Single AI Prompt Use?

This is the figure most people want, and it is where the best-documented company numbers exist.

Google's estimate for Gemini

Google researchers published a paper measuring the energy, emissions and water used to serve its Gemini app in production. They report that the median Gemini Apps text prompt consumes 0.24 watt-hours of energy and the equivalent of about five drops of water, 0.26 milliliters, and that over one year the energy per median prompt fell 33 times and the carbon footprint fell 44 times (Elsworth et al., arXiv, August 2025). This is a measurement from within a company, covering a defined set of activities, so it is useful but not an independent audit. Check what the figure includes before comparing it with others: it describes a median text prompt in Google's own serving environment, within the scope the paper defines, and it should not be read as the full footprint of every kind of AI use.

OpenAI's statement for ChatGPT

OpenAI's chief executive, Sam Altman, wrote in a blog post that the average query uses about 0.34 watt-hours of electricity and about 0.000085 gallons of water, which he described as roughly one fifteenth of a teaspoon. Converted, 0.000085 gallons is about 0.32 milliliters. The post gives no methodology or caveats for these numbers, so they are best treated as a claim to be examined, not an audited measurement.

What those numbers add up to

The arithmetic below uses the company figures, not independent measurements. It shows scale, not a verified footprint.

On these figures, one person's everyday chatbot use accounts for a very small amount of water. That is true, and it is also incomplete. The total depends on how many people use these systems and how often, and the figures cover text prompts, not the heaviest uses.

Images, video and long tasks

Not all AI tasks are equal. MIT Technology Review's 2025 analysis, written by James O'Donnell and Casey Crownhart, estimated the energy for different outputs using open models. It put a standard-quality 1024 by 1024 image from Stable Diffusion 3 Medium at about 2,282 joules and a higher-quality image at about 4,402 joules, and a five-second video from CogVideoX at about 3.4 million joules, which the article describes as more than 700 times the energy required to generate a high-quality image (MIT Technology Review). The authors stress that these are open models, measured as GPU energy with totals estimated by doubling to account for cooling and other equipment, and that the figures cannot serve as a proxy for closed systems such as ChatGPT. Water use tracks energy use in a rough way, because more computing means more heat to remove and more electricity to generate, so heavier tasks generally carry a larger footprint. Long conversations, large documents and tools that run many steps in the background also use more than a single short prompt.

If you are choosing between image generators or want to understand how they compare in other ways, see our guide to the best AI image generators.

The Bigger Picture: AI Water Consumption and Data Centers in Total

Per-prompt numbers describe an individual interaction. The total matters for planning and policy.

United States

According to Lawrence Berkeley National Laboratory's 2024 report on US data center energy use, US data centers directly consumed about 17 billion gallons of water in 2023, roughly 64 billion liters, a figure some summaries round to 66 billion. The same report projects that direct consumption could double or even quadruple by 2028, and the researchers who summarized it say the indirect consumption from electricity was about 12 times greater than the direct amount in 2023 (The Conversation). Not all of this is AI, since data centers also run streaming, storage, search and ordinary cloud services, and there is no clean public split.

The indirect figure is contested. Analyst Brian Potter, writing in Construction Physics, argued that estimating the water behind data center electricity is complicated, because hydropower reservoirs lose far more water to evaporation per kilowatt-hour than thermal plants consume, and that the answer depends on how those sources are allocated (Construction Physics). He offered a lower alternative estimate of roughly 200 to 275 million gallons per day under different assumptions. The disagreement shows why the indirect number should be treated as a range.

Electricity demand is the driver behind indirect use

The International Energy Agency estimated in its 2025 report on energy and AI that global data centers used about 415 terawatt-hours of electricity in 2024, around 1.5% of global electricity use, and that the figure could more than double to about 945 terawatt-hours by 2030, with AI the most important driver of that growth (IEA, summarized by Data Center Dynamics). More electricity generally means more indirect water use, unless the power comes from sources that use little water, such as wind and solar.

What companies report

Large technology companies publish environmental reports, and the numbers are growing.

• Google reports that it consumed about 8.1 billion gallons of water across its data centers and offices in 2024 and replenished about 4.5 billion gallons, which it describes as 64% of its freshwater consumption, against a goal of replenishing 120% of the freshwater it consumes by 2030 (summary of Google's 2025 Environmental Report). Company totals can still rise even as efficiency per prompt improves, so it is worth looking at both numbers.
• Individual facilities can be large. The Conversation article cites Google's Council Bluffs, Iowa data center as consuming about 1 billion gallons a year, equivalent to the residential water use of Iowa for about five days.

These are company disclosures. Reporting is voluntary, and the researchers behind The Conversation article note that companies report different statistics in ways that make them hard to combine or compare.

Why Location Matters More Than the Average

A national total can hide local stress. Water is a local resource, and the same consumption means very different things in a wet region and a dry one.

A Bloomberg News analysis found that about two-thirds of new data centers built or in development since 2022 are in places already under high water stress, with Arizona among the states seeing heavy construction (Bloomberg, summarized by Sherwood News). Reports based on that work say mid-sized facilities can use up to about 300,000 gallons a day and large ones as much as 5 million gallons daily, which is comparable to what a small town uses. Treat those as upper-end examples, not a typical figure.

Local effects depend on several things:

• The water source. Using treated municipal water, groundwater, surface water or reclaimed wastewater has different consequences.
• The local supply situation. A drought, an overdrawn aquifer or competing demand from farms and households makes new consumption harder to absorb.
• Timing. Cooling demand peaks in hot weather, which is also when water is scarcest.
• Agreements and permits. Some facilities sign agreements with utilities, and some communities have pushed back or asked for disclosure.

This is why researchers and community groups keep asking for facility-level reporting, not just company totals.

Why the Numbers Disagree

If you have seen very different figures for AI's water use, the cause is usually one of these differences in method.

A good habit is to ask four questions of any claim: Is it direct, indirect or both? Is it consumption or withdrawal? What activity does it cover? Who measured it and how?

Is AI's Water Use a Big Problem?

The evidence supports two statements that can both be true.

It is small for an individual. On the company figures above, ordinary text use by one person amounts to a few liters a year. Lecturing people about each chatbot question is unlikely to move the numbers.

It can be significant in aggregate and in specific places. Total data center water use is large, it is projected to grow, and a large share of new capacity is going into regions with water stress. The indirect water behind electricity is substantial under most estimates, even if the exact figure is disputed.

Comparisons with other sectors can be useful but are easy to misuse. Brian Potter's essay makes the point directly: water use figures are easy to take out of context or to compare misleadingly. A comparison with agriculture, golf courses or thermal power plants depends on exactly what is counted, so check the underlying data before you repeat one. A fair summary is that AI's water footprint is real, measurable and worth managing, even though it is one of many pressures on water supplies. The question that matters is whether growth is occurring in places and in ways that the local water supply can bear.

What the Industry Is Doing

Several approaches are in use, and they do not all point the same way.

• Better efficiency per task. Google's paper reports a 33 times reduction in energy per median prompt over a year, through software improvements and cleaner energy procurement. Efficiency gains reduce both direct and indirect water, although they can be offset by growth in total usage.
• Closed-loop and chip-level cooling. Microsoft's reported design uses recirculating liquid cooling that does not evaporate water for cooling. These designs can lower water use but may raise electricity use.
• Reusing and treating water. Some facilities use reclaimed or recycled water, which reduces pressure on drinking water supplies.
• Replenishment projects. Google and others fund projects that return water to watersheds. Critics point out that replenishment in one place does not remove local impacts in another, and that the accounting is voluntary.
• Better siting. Choosing cooler climates, areas with abundant water or sites near clean power can reduce the footprint, though many new facilities are going where land, power and connectivity are available.
• Disclosure. Some companies publish water usage effectiveness and total consumption. Researchers want more consistent, facility-level data.

Water use effectiveness, or WUE, is the metric most often used. It is data center water consumption in liters divided by IT equipment energy in kilowatt-hours, so a lower value means less water per unit of computing. Reports put typical enterprise values anywhere from about 0.2 to 1.8 liters per kilowatt-hour depending on climate, design and water reuse, and Microsoft's reported figure of 0.30 in 2024 sits near the low end.

What You Can Do

As an individual

• Don't worry about each prompt. The per-prompt figures are small. Using AI for things that are useful to you is not where the main impact lies.
• Be deliberate with heavy uses. Large image batches, long videos and repeated regenerations use more resources than short text tasks.
• Ask for transparency. Support companies that publish water and energy data, and ask the ones that don't.
• Look at the bigger levers. Home water use, energy use and what you buy are likely to have far larger effects than a chatbot habit.

As a business or team

• Ask your AI vendors about water and energy. Use the questions in the next section.
• Avoid wasteful automation. Workflows that call a model thousands of times for little value cost money and resources. Our guide to building an AI workflow for your business explains how to design and test workflows so they only do work that is needed.
• Match the model to the task. A smaller or cheaper model can handle many routine jobs.
• Cache and reuse. If you generate the same output repeatedly, store it.
• Include sustainability in procurement and reporting. If you publish an environmental report, decide how you will account for the footprint of the cloud and AI services you buy.
• Plan governance. Environmental impact can be part of your AI policy. See our guide to what AI governance is.

As a developer or builder

• Measure your own usage. Track calls, tokens and the cost of workflows.
• Choose efficient architectures. Use smaller models where they are good enough, batch work and avoid unnecessary regeneration.
• Pick regions and providers with care. If your provider publishes data by region, use it.
• Be honest in claims. If you say your product is sustainable, state what the claim covers.

Questions to Ask Any AI Provider About Water

These questions are a practical way to separate marketing from measurement.

1. Do you publish water consumption for your data centers, and is it direct, indirect or both?
2. Is the figure consumption or withdrawal?
3. What is your water usage effectiveness, and how is it measured?
4. How much of your water comes from drinking water supplies versus reclaimed or non-potable sources?
5. In which regions do you run the services I use, and are any of them in water-stressed areas?
6. Do you disclose per-prompt or per-task water and energy figures, and what do they include?
7. What is your plan to reduce water use as usage grows, and how will you report progress?
8. Can you share facility-level data, not only company totals?

A provider that answers these clearly is signaling it takes the question seriously. One that cannot answer is telling you something too.

Glossary

• Water consumption: water withdrawn that is not returned to the same source, usually because it evaporates.
• Water withdrawal: water taken from a source, part of which may be returned.
• Direct water use: water used on site, mostly for cooling.
• Indirect water use: water used to generate the electricity that powers a facility.
• Evaporative cooling: cooling that works by evaporating water.
• Closed-loop cooling: cooling in which liquid circulates in a sealed system.
• WUE (water usage effectiveness): liters of water consumed per kilowatt-hour of IT equipment energy.
• Training: the process of building a model, which takes large amounts of computing once.
• Inference: using a trained model to respond to requests, which happens continuously.
• Water stress: a measure of how much demand for water competes with available supply in a region.

Frequently Asked Questions

How much water does AI use? There is no single figure. Per prompt, Google reports about 0.26 milliliters of water for a median Gemini text prompt and OpenAI's chief executive has said about 0.3 milliliters for an average ChatGPT query. In total, US data centers directly consumed about 17 billion gallons in 2023, with much more used indirectly through electricity. Training a large model such as GPT-3 was estimated to evaporate about 700,000 liters.

How much water does ChatGPT use per question? OpenAI's chief executive has said an average query uses about 0.000085 gallons, about one-fifteenth of a teaspoon, or roughly 0.3 milliliters. The post that gave the figure includes no method, and the number covers an average text query, not images or long tasks.

How much water does one AI prompt use? On company figures, roughly 0.26 to 0.32 milliliters for a typical text prompt. A thousand such prompts would be about a cup of water. Heavier tasks such as image and video generation use much more energy and likely more water.

Does AI use more water than other industries? It depends on what is counted and where. Data centers are one of many large water users, and any comparison depends on what each side includes. Data center use can be significant locally, and the total is growing. Be cautious with comparisons that don't state what they include.

Is AI's water use getting better or worse? Both. Efficiency per task has improved quickly, as Google's reported 33 times reduction in energy per median prompt over a year shows. At the same time, total data center electricity and water use are projected to rise because of growth in usage.

Why does AI need water at all? Servers produce heat, and many data centers use evaporating water to remove it. Water is also used to generate the electricity that powers them.

What is the difference between direct and indirect water use? Direct use is water consumed at the data center, mainly for cooling. Indirect use is water consumed to generate the electricity that powers it. Indirect use is often estimated to be larger than direct use, but the estimates vary widely.

Are data centers being built in places with water shortages? A Bloomberg News analysis found that about two-thirds of new data centers built or in development since 2022 are in places already under high water stress. Local conditions vary, and some facilities use reclaimed water or closed-loop cooling.

Can data centers use no water for cooling? Some designs use closed-loop liquid cooling that does not evaporate water for cooling. Microsoft has said it is adopting such designs. These can trade lower water use for higher electricity use, and they still have an indirect water footprint through the power supply.

What is water usage effectiveness? WUE is liters of water consumed divided by IT equipment energy in kilowatt-hours. A lower number means less water per unit of computing. Reports put typical values between about 0.2 and 1.8 liters per kilowatt-hour.

Should I stop using AI because of its water use? For an individual, text-based use accounts for a very small amount of water on current company figures. A more useful response is to be deliberate with heavy uses, support transparency and pay attention to where large-scale growth is happening.

Why are estimates of AI's water use so different? They differ in what they count, such as direct or indirect water, consumption or withdrawal, training or inference, and in the location, date and method used. Always check the scope before comparing numbers.

Conclusion

How much water does AI use? On the best available figures, a single text prompt uses a fraction of a milliliter, training a large model can evaporate hundreds of thousands of liters, and the data centers behind it consume billions of gallons a year in the United States alone, with a larger indirect footprint through electricity. All of those statements can be true at once, because they measure different things at different scales.

The useful way to think about AI's water footprint is to look past the headline number. Ask whether it is direct or indirect, consumption or withdrawal, and where it occurs. The risk is not that each chatbot question is wasteful. It is that fast growth in energy-hungry computing is concentrated in places and systems where water is already tight, and that the data to judge it is incomplete. Better disclosure, efficient design, smart siting and informed choices by people who buy and build AI services are the practical levers.

Sources

• Li, Yang, Islam and Ren, Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models (arXiv, 2023)
• Sam Altman, The Gentle Singularity (June 2025; no methodology given for the per-query figures)
• MIT Technology Review, We did the math on AI's energy footprint (May 2025)
• The Conversation, Data centers consume massive amounts of water - companies rarely tell the public exactly how much