How Much Water Does AI Use? The Real Cost of Each Query

You ask ChatGPT for a cake recipe and minutes later see a post claiming each query "drinks a bottle of water." The same day, another post insists it's just a drop. How much water does AI actually use?

Both sides cite real numbers. The problem is that each company measures something different, and almost no one explains what.

Here are the official numbers from Google, OpenAI and Mistral, why they range from five drops to three tablespoons, and where AI's thirst truly matters.

Quick summary: the answer in 30 seconds. A text query uses 0.26 mL to 45 mL of water, depending on what's counted. The low number covers only computer cooling; the high includes manufacturing. Most water typically goes to the power plant generating the electricity. In total, it's little; the risk is local, in an already stressed river, in summer.

Index

• How much water does AI use, anyway?
• Why does AI use water?
• How much water does ChatGPT use per query?
• How much do data centers use per year?
• Where does AI water consumption truly matter?
• What are companies doing?
• What this means for you
• Frequently asked questions

How much water does AI use, anyway?

Per query, little: from a few drops to three tablespoons, according to the companies themselves. Per company, a lot: Amazon withdrew about 9.5 billion liters in 2025. Both things are true at the same time, because billions of small queries add up to a large volume.

This water goes to data centers: warehouses full of servers, the computers that assemble the responses, running all day, like an industrial kitchen with all burners on. When you use Gemini or ChatGPT, the response is calculated there, not on your phone.

For measuring water, two words matter. Withdrawal means pulling water from a river or well. Consumption is the part that doesn't return, almost always because it evaporated. Many viral numbers get it wrong precisely by mixing the two.

Why does AI use water?

Because chips heat up and the cheapest way to cool an entire building is to evaporate water. Water that turns to vapor carries heat away, like sweat does with the body. The hotter and drier the day, the more water the system evaporates to keep servers at the right temperature.

Amazon describes the process this way: water is sprayed onto material that works like a giant sponge, and hot air passes through this sponge. As it evaporates, the water removes heat from the air, and this cooler air passes over the servers. On mild days, below about 29 °C, the company says it uses only outside air, without water.

The water-free alternative is chillers, similar to a giant air conditioner. They use 25% to 35% more electricity, Amazon says. Google reaches the same conclusion, with a smaller number: cooling with water saves about 10% of energy. It's a trade-off: less water almost always means more energy.

The water no one sees: the power plant's

Every data center needs electricity, and many power plants also drink water. Gas, coal and nuclear plants boil water to spin turbines and then need to cool that steam. In 2021, the U.S. electric sector withdrew about 180 trillion liters just for this, according to the U.S. Energy Information Administration (EIA).

On average, it was about 44 liters withdrawn per kilowatt-hour, the unit on your electricity bill; a ten-minute shower with an electric heater uses close to 1 kilowatt-hour. The type of plant changes everything. A modern gas plant withdraws about 10.6 liters per kilowatt-hour; a coal plant, 72.6. Wind and solar don't use cooling water.

Much of this water returns to the river, warmer, but it's the most forgotten part of the equation. That's why Professor Shaolei Ren of the University of California, Riverside, adds direct water (from the building) with indirect water (from the plant).

In one sentence: AI drinks water in two places, in the building that cools the chips and in the plant that generates the energy, and the second is usually larger.

How much water does ChatGPT use per query?

According to Sam Altman, president of OpenAI, an average ChatGPT query uses about 0.32 mL. Google says 0.26 mL for Gemini. Mistral, a French company, says 45 mL for its assistant, Le Chat. The difference is almost 175 times, and it comes from the method, not the technology.

Google's number is in the company's technical blog and in a study on arXiv, a site where scientists publish work before peer review. Altman's came out on his blog, in gallons: 0.000085 per query. Mistral's comes from a life cycle analysis done with consultancy Carbone 4 and the French environmental agency (ADEME).

Why the numbers are so different

Google counts only water evaporated in cooling its buildings and uses the median query, the one in the middle of the line. Power plant water is left out. An even narrower calculation, only with chips in use, would give 0.12 mL, and Google itself says this version falls below reality.

Researchers criticized the scope. "They're hiding the essential information," Shaolei Ren told The Verge. Alex de Vries-Gao of Vrije Universiteit Amsterdam compared the number to seeing only the tip of the iceberg. At the time, Google's study had not yet undergone peer review.

Mistral went to the other extreme: the 45 mL includes server manufacturing and energy for everything running in the building. Altman didn't explain how he arrived at his number.

A simple calculation shows the weight of the invisible part. A Gemini query uses 0.24 watt-hours, the same as watching TV for less than nine seconds, Google says. If that energy came from the average U.S. grid, the plant would withdraw about 10 mL, almost 40 times the building's 0.26 mL. This is our estimate, of withdrawn water, and part of it returns to the river.

In one sentence: no one is lying in the per-query numbers, but each counts a different piece of the same equation.

How much do data centers use per year?

Billions of liters per company, but little compared to other uses. Amazon withdrew about 9.5 billion liters in 2025. U.S. lawns and gardens alone take about 12.5 trillion liters per year, according to the U.S. Environmental Protection Agency (EPA): more than 1,300 times Amazon's volume. Google, citing a U.S. government lab, says all U.S. data centers use less than 1% of the water from those lawns.

Before responding, each model (the program behind the chatbot) spends months being trained on thousands of chips, and that also drinks. In 18 months of training and use, Mistral's Large 2 consumed 281 thousand cubic meters, about 112 Olympic pools. Ren's study calculates that training GPT-3 evaporated 700 thousand liters in Microsoft's U.S. data centers, more than a quarter of an Olympic pool.

Where does AI water consumption truly matter?

In place and time. A data center doesn't pull water from the planet, it pulls from a river, a well or a city's network. If that river is already stressed, a large building competes with homes, farms and industries. And peak use coincides with heat, when everyone also needs water.

Amazon, for example, only uses water for cooling when outside air exceeds about 29 °C. In other words, water goes out precisely in the hottest hours of summer. Ren's study shows the same effect: the same query uses more water on a hot day, in a dry region.

That's why Ars Technica calls the total a "drop in the bucket" and, in the same article, warns of the local effect. Residents don't drink from the national account, they drink from the city's reservoir.

In practice, Google acknowledges this. The company says it evaluates the watershed of each location (the rivers and reservoirs that supply the region) before building. If the source is at risk, it chooses air cooling or reclaimed water, treated wastewater that isn't drinkable.

What are companies doing?

Three fronts: cooling without evaporation, using water no one would drink, and funding projects that "return" water to watersheds. All help at the building. None alone solves power plant water, which depends on replacing gas and coal with wind and solar.

Closed loop. Nvidia launched servers cooled by liquid at up to 45 °C, which circulates in a closed loop, filled once. Warm like this, it cools with outside air, without evaporating, like a car radiator. In favorable climates, the company says it cuts cooling water from about 9.8 million liters per megawatt per year, almost four Olympic pools, to nearly zero. TechCrunch noted the limit: the savings stop at the building wall.

Hotter building. Amazon started letting servers work at higher temperatures, which reduces evaporation hours. It says it uses 0.12 liters per kilowatt-hour, versus 0.84 for the industry average. In 26 buildings, it only uses reclaimed water.

Returning water. Google and Amazon promise to replenish more water than they consume by 2030, with projects like restoring wetlands and fixing leaks. Amazon says it returned 3 gallons for every 4 used in 2025. Replenishing water elsewhere, however, doesn't restore the neighbor's river.

What this means for you

Day to day, little: stopping chatbot use won't save a river. It matters to know how to read the next viral number and what to demand from those building data centers near you.

If the energy topic interests you, the article about AI data center energy bills shows the other side of the same coin. For more explanations like this, see Kalkazu.org's artificial intelligence section.

Frequently asked questions

Does water evaporated by the data center disappear?

It doesn't disappear from the planet: it becomes vapor and returns as rain, but in another place and at another time. For the city that supplied the water, it left the account. That's why evaporated water is called consumed.

Do data centers use drinking water?

Often, yes, because treated water has less dirt and clogs equipment less. But reclaimed water use is growing. Amazon says it has 26 buildings using only it and another 130 with contracts to start using it.

The numbers will keep fighting in feeds until companies measure the same way; until then, the right question isn't "how much," but "what was counted."