Does AI Use Water? How Much, Why & Which Company Uses Most

Does AI Use Water?

Yes, but not in the way most people think. The AI model itself does not “drink” anything. Water is used by the buildings and power plants behind it.

When you type a question into ChatGPT, Gemini, or Claude, your request travels to a data center. Thousands of computer chips work on it. Those chips get very hot. Cooling them often uses water. The power plant that supplies the electricity may also use water to make steam or cool its equipment.

So the honest answer has two parts. Water use per prompt is small. Water use across billions of prompts, and across new data center campuses, is big. Both facts are true at once, and that is why the debate is so heated.

Why Does AI Need Water?

Because chips make heat, and water removes heat very well. AI chips, called GPUs, run much hotter than normal computer chips. If they get too hot, they slow down or break. Data centers must run cooling day and night.

Water is popular for three reasons. It carries heat better than air. It is cheap compared with land and power. And evaporating water is an efficient way to throw heat out of a building. Water is often the last thing companies think about when picking a site, because it costs less than land and electricity.

How Does AI Use Water? Direct and Indirect Use

Experts split AI water use into two buckets.

Direct (onsite) water: Water used inside the data center for cooling. The main method is evaporative cooling. Warm water goes to a cooling tower, some of it evaporates, and that carries the heat away. The evaporated water is “consumed,” because it does not go back to the local river or pipe.

Indirect (offsite) water: Water used to make the electricity. Coal, gas, nuclear, and some other plants use large amounts of water for steam and cooling. A data center that uses lots of power is therefore tied to lots of power-plant water, even if its own cooling uses little.

A third, smaller bucket is chip manufacturing, which uses ultra-pure water. It is often left out of per-prompt numbers.

This is a big deal for the debate. Many headlines only talk about cooling. But in the US, most of the water footprint comes from the power plants. It also means a clean grid can cut AI’s water use, not just better cooling.

The Main Ways Data Centers Cool Chips

The key idea is a trade-off. Saving water on site often means using more electricity. Using more electricity can mean more water at the power plant. The best choice depends on the local climate and the local grid.

How Much Water Does One AI Prompt Use?

This is the most searched question, and the answers look very different. Here is why.

Why the Numbers Disagree

• What is counted. Company numbers count only onsite cooling. Researchers often add power-plant water. That alone can multiply the figure.
• Which model and how long the answer. A short chat reply and a long report are not the same. Images and video use more.
• Where and when. A cool, northern data center at night uses less than a hot, dry one in summer.
• Old versus new data. The famous “one bottle of water per email” claim (about 519 mL for a 100-word GPT-4 email) came from 2024. Its author, Shaolei Ren of UC Riverside, has since said the GPT-4 figure is closer to 15 mL, with about 5 mL inside the data center. Models and chips have also become more efficient.

Plain summary: a normal text prompt uses somewhere between a few drops and a teaspoon or two of water. It is not a bottle. Still, the number of prompts is huge, so the total adds up.

Training vs Everyday Use

Building a model (training) uses a lot at once. A widely cited 2023 study estimated about 700,000 liters of water to train GPT-3. But after launch, billions of daily prompts (inference) keep using water every day. Over a model’s life, everyday use often becomes the larger share.

Which AI Uses How Much Water?

Be careful here. Only Google and OpenAI have given official per-prompt numbers, and they use different methods. Companies such as Anthropic, Meta, and xAI have not published a comparable per-prompt water number that I could find. So any ranking of “which chatbot uses the most water” is an estimate, not a fact.

Which AI Company Uses the Most Water?

Here are the latest company-wide numbers. Treat them as a rough guide, not a leaderboard. The companies measure different things (consumed versus withdrawn), in different years, at very different sizes.

Google reports the largest number. Its 2026 environmental report said it consumed 10.9 billion gallons in 2025, up 34% from 2024 and more than double 2021. Reporting links much of the rise to its AI data center build-out. Google also says it replenished about 78% of that (around 7.7 billion gallons) through water projects.

Amazon (AWS) disclosed for the first time in June 2026 that its data centers withdrew 2.5 billion gallons in 2025. That is lower than Google’s figure, but the number only covers direct use and measures withdrawal, not consumption.

Microsoft withdrew about 2.7 billion gallons in 2024. It says new AI data centers use designs that need no water for cooling, and it reports reaching “water positive” in fiscal 2025.

Meta used about 813 million gallons in 2023, with about 95% going to data centers. Newer figures may exist in later reports.

Which Company Is the Most Water-Efficient?

A better yardstick is WUE, or water use effectiveness, which is liters of water per kilowatt-hour of electricity. Lower is better.

On this measure, AWS and Microsoft look strong. But these are company-reported numbers, and critics note the comparisons are not always apples to apples. Google does not publish a WUE number, which makes direct comparison hard. Note too that WUE counts only onsite water, not the power plant’s.

Fair conclusion: Google reports the biggest total, partly because it is large and AI-heavy. AWS and Microsoft report efficient onsite use. Nobody discloses the full picture across onsite, power, and chips in the same way. The 2026 push for disclosure is aimed at fixing that.

How Much Water Do AI Data Centers Use in Total?

• One facility: A medium data center uses about 110 million gallons a year (roughly 300,000 gallons a day). Hyperscale sites can use up to 5 million gallons a day, similar to a town of 10,000 to 50,000 people, according to Brookings as relayed by secondary reports.
• US direct use: About 17.4 billion gallons in 2023. Projections for 2028 range from 38 to 73 billion gallons, according to EPA figures cited by Fortune.
• US indirect use: About 211 billion gallons via electricity in 2023.
• Global: The IEA is cited as putting data center water use at about 560 billion liters in 2023, though sources differ in what they include.

Can AI Cause a Water Crisis?

This is the key question. The honest answer depends on the scale you look at.

The Case That the Fear Is Overblown

Agriculture uses most water. In Arizona, farming is estimated at around 86% of use, while industry as a whole is around 8%. US golf courses use roughly 2 billion gallons a day. One advocate’s analysis argues all data centers use about 0.05% of US freshwater and that in Maricopa County, Arizona, they use about 0.12% of the county’s water versus 3.8% for golf. That author is a pro-industry voice, so treat the numbers as one side of the argument. Per-prompt use is also falling as hardware improves.

The Case That the Concern Is Real

Averages hide local pain. The problem is where data centers are built, and when.

• Site choice: A Guardian analysis found 517 of 809 planned US data centers are in areas that had been in drought over the past year. Bloomberg found about two-thirds of new facilities since 2022 sit in high water stress zones.
• Newton County, Georgia: A Meta data center was reported to disrupt nearby private wells, leaving families hauling water. Another report says a data center there uses about 500,000 gallons a day, around 10% of the county’s use, per the Lincoln Institute. A critic of the story notes the wells issue was tied to construction sediment before the site was running, so causes are debated.
• May 2026 cases: Two developers, one in Georgia and one in Arizona, were reported to have taken public water they were not allowed to take. In Tucson, it was about 650,000 gallons for dust control.
• The US West: Western Resource Advocates estimates data centers across five western states could use 7 billion gallons a year by 2035, enough for up to 194,000 people. Lake Mead has a 2026 shortage declaration that forces Arizona to cut about 18% of its apportionment.
• Texas: Researchers calculated data centers could reach 9% of Texas’s water use by 2040.
• Business impact: Ceres says water and grid opposition disrupted about $130 billion of data center projects in the first quarter of 2026.

My Verdict

AI is not draining the planet. But it is adding a new, fast-growing demand in places that are already short of water, and it often arrives with little warning and little public data. That is why water is an issue for AI, and not simply a myth.

The risk is highest where three things meet: a dry climate, a fast-growing cluster of data centers, and a power grid that uses lots of water. The risk is lowest where data centers use closed-loop cooling, reclaimed wastewater, or cool climates.

What Are Companies and Governments Doing?

One caution: “water positive” means a company funds projects that return water elsewhere. It does not mean the local river is unaffected. Replenishment far from the data center does not help the town next to it.

What Can You Do as a User?

• Don’t panic about each prompt. One chat reply is tiny in water terms.
• Be mindful with heavy tools. Images and video use much more than text.
• Ask for disclosure. Prefer companies that publish water numbers and per-site data.
• Support local oversight. Data center permits and water rules are decided locally.
• Keep perspective. Food, clothing, and energy use far more water than chatbots do.

Common Myths About AI and Water

• Myth: Every prompt uses a bottle of water. Fact: Modern estimates are from a fraction of a mL to a few mL for typical text prompts.
• Myth: AI water use is made up. Fact: Company reports show rising totals, and local conflicts are documented.
• Myth: Only cooling matters. Fact: In the US, the electricity behind data centers carries most of the water footprint.
• Myth: “Water positive” means no local impact. Fact: It is an accounting goal, not a guarantee for nearby communities.

Frequently Asked Questions

Does AI use water?

Yes. Data centers use water to cool AI chips, and power plants use water to make the electricity that runs them.

How much water does ChatGPT use per question?

OpenAI’s CEO said about 0.32 mL (one-fifteenth of a teaspoon) per average query, counting onsite cooling. Independent estimates that include power-plant water are higher, roughly 1 to 15 mL.

How much water does Google Gemini use per prompt?

Google reports a median of about 0.26 mL per text prompt.

Does an AI image use more water than text?

Yes. One mid-2026 assessment put a standard image at about 28.6 mL, over 100 times a short text reply. Estimates vary by tool.

Which AI company uses the most water?

Among companies that publish figures, Google reports the largest total (10.9 billion gallons in 2025). But companies measure differently, so it is not a perfect comparison.

Which AI company is most water-efficient?

AWS reports 0.12 liters per kWh and Microsoft 0.27, both well below the cited industry average of 0.84. Google does not publish a WUE number.

Will AI cause a global water shortage?

Probably not on its own. Farming and other sectors use far more. But AI can worsen local shortages in dry regions.

Why do data centers use so much water?

AI chips produce lots of heat. Evaporative cooling is cheap and efficient, so many sites use it, despite the water cost.

Is AI water use increasing?

Yes in total. Google’s rose 34% in 2025. Per-prompt use may fall as efficiency improves, but more prompts and bigger models push totals up.

Can AI data centers use no water at all?

Some new designs use closed-loop or dry cooling with almost no onsite water. They may use more electricity, so the power-plant water still matters.

Final Takeaway

AI does use water, and the numbers have two sides. A single prompt uses a tiny amount, often less than a teaspoon. But with billions of prompts and fast-growing data centers, the total is large and rising, with Google up 34% in one year.

The biggest risk is local, not global. The worry is dry places where new data centers compete with farms and towns, often without clear public data. The most useful answers are more disclosure, smarter siting, better cooling, and cleaner, less thirsty power.

So the fair view is neither “AI is draining the planet” nor “it’s all hype.” AI water use is a real but manageable problem, and it depends on where and how we build.

Related Reading

• AI industry liability exposure: who pays when AI infrastructure and products cause harm
• The AI kill switch debate: another side of AI risk and regulation
• Bessent rejects an AI liability shield: the policy fight over who is responsible for AI

Sources

1. AI Water Usage and Environmental Impact Stats, Exploding Topics
2. AI Data Center Water Usage Statistics 2026, Axis Intelligence (cites Google’s 2026 Environmental Report)
3. Behind Amazon’s Industry-Leading Water Efficiency Score, Trellis
4. How Much Water Do Data Centers Use? GPUSmith (LBNL, IEA, Meta figures)
5. Does AI Use a Lot of Water? TechJournal (Ren’s revised estimate)
6. How Much Water Does AI Use? Analytics Insight
7. How Much Water Does AI Use? MillionMiner
8. The Data Center Water Footprint, AKCP
9. America’s Data Centers Are Thirsty, Fortune
10. Most New US Data Centers Are Slated for Drought-Plagued Areas, Mother Jones (Guardian analysis)
11. AI Data Center Water Use Collides With Drought in the West, Quartz
12. Many Ways Data Centers Affect US Communities, World Resources Institute
13. AI Data Centers Tied to Water Use Across 7 States (Ceres report coverage)
14. The Myth of Water Usage in Datacenters, Julian Estevez (opinion, used for the skeptic view)