How much water does AI use? Real numbers per prompt

One AI text prompt uses somewhere between a few drops and a few tablespoons of water. Google puts a median Gemini prompt at 0.26 ml. Older academic estimates for GPT-3 land closer to 17 ml. Across the US data centers used about 17 billion gallons of water for cooling in 2023, before counting the power plants that feed them.

The gap between those numbers isn't a mistake. How much water AI uses depends on what each estimate counts. Below you'll see what each one measures, how AI uses water in the first place and when it becomes a real local problem. Updated November 2026.

How does AI use water?

AI doesn't drink water. The computers that run it get hot. Water is the cheapest way to carry that heat away. AI uses water in three places.

Cooling the data center

Most large data centers use evaporative cooling. Warm water from the server halls runs through cooling towers, part of it evaporates and that evaporation pulls heat out. The evaporated water is gone from the local supply. Newer designs cool the chips directly with liquid in a closed loop and use much less.

Making the electricity

Thermoelectric power plants (coal, gas and nuclear) also cool themselves with water. Every kilowatt-hour a data center buys carries a water cost at the plant. Researchers call this off-site or indirect water.

Making the chips

Chip factories use large amounts of ultrapure water. That share is the hardest to measure and most per-prompt estimates leave it out.

Two terms matter when you read any number. Water withdrawal is all the water taken from rivers or groundwater, even if most of it goes back. Water consumption is the part that doesn't come back, mostly because it evaporated. The UC Riverside study on AI's water footprint uses these definitions. In its 2027 projection AI's global water withdrawal is about 11 times its consumption. Check which one a headline quotes.

How much water does one AI prompt use?

It depends on who's counting. This table puts the best-known estimates side by side. The last column turns each one into something easier to picture, using a standard US showerhead that runs up to 2.5 gallons (about 9.5 liters) a minute, the federal limit noted by the EPA.

So why is the highest estimate about 170 times the lowest? Four reasons:

• What's counted. Google's figure covers water used for cooling in its own data centers. The UC Riverside estimate adds the water used to generate the electricity, which is usually the bigger share. Mistral counts the whole life cycle, including making the servers.
• Which model. The UC Riverside numbers are for GPT-3, a model from 2020. Google says energy per Gemini prompt fell 33 times between May 2024 and May 2025, as models and hardware got more efficient.
• Where and when. A data center in a hot, dry place evaporates more water than one in a cool climate. In the UC Riverside paper a 500 ml bottle covered about 10 GPT-3 responses in Washington state and about 70 in Ireland.
• What a prompt is. A short question and a 2,000-word report aren't the same job. Image and video generation take far more computing than text.

The honest answer for a typical text prompt today is a fraction of a milliliter to a few tens of milliliters. Even at the high end that's a few sips.

How much water does ChatGPT use?

OpenAI hasn't published a method. The only official figure comes from a June 2025 blog post by CEO Sam Altman, "The Gentle Singularity." He wrote that the average ChatGPT query uses "about 0.000085 gallons of water." That's roughly one fifteenth of a teaspoon, or about 0.32 ml.

That's close to Google's Gemini figure. But without a method nobody can check whether it includes power plant water or only cooling. The often-quoted "a bottle of water per conversation" comes from the 2023 UC Riverside study and refers to GPT-3, not to the models ChatGPT runs today.

If you want to understand the privacy side of ChatGPT too, here's whether ChatGPT is safe to use.

How much water do AI data centers use?

Per prompt the numbers are tiny. In total they aren't. A 2024 Lawrence Berkeley National Laboratory report estimated that US data centers used:

• 17 billion gallons of water for cooling in 2023, about 46 million gallons a day
• about 211 billion gallons more through the electricity they used, roughly 12 times the cooling figure

Those numbers cover all data centers, not just AI. The report expects direct water use could double or quadruple by 2028 as AI grows. Looking further out, the UC Riverside team projected that global AI could withdraw 4.2 to 6.6 billion cubic meters of water a year by 2027.

Single companies give a sense of scale. Google reported more than 6 billion gallons of cooling water across its data centers in 2023. One Google data center in Iowa used about 1 billion gallons in 2024, roughly five days of the whole state's residential water use.

Can data centers reuse the water? Is it drinkable afterward?

Mostly not. In an evaporative system the cooling works because water turns into vapor. That vapor leaves the site and comes back as rain somewhere else, not in the same town.

The 20% to 30% that doesn't evaporate picks up minerals as it circulates. Over time it can scale and clog the equipment. So it's flushed out as "blowdown" and sent to a wastewater plant, according to Data Center Knowledge. You wouldn't drink it. Recycling it on site costs more than most operators are willing to pay today.

Closed-loop systems are the exception. The water stays sealed in pipes and goes round and round. Microsoft says all its new data center designs since August 2024 cool chips this way and avoid more than 125 million liters of evaporated water per data center each year, per its own announcement. The trade-off is that these systems use more electricity, which has its own water cost at the power plant.

Is AI's water use a real problem?

For you as one user it isn't. Producing 1 kg of beef takes about 15,400 liters of water by the Water Footprint Network's estimate, cited by the UN Convention to Combat Desertification. That equals about 900,000 AI prompts even at the 17 ml estimate.

For some towns it is. Water is local. A data center that takes over 25% of a community's supply, which has happened, competes with homes and farms during a drought no matter how efficient each prompt is. That's why where new AI data centers are built matters more than the size of your prompts.

The useful questions to ask aren't "should I stop using AI" but "where is it running, how is it cooled and what powers it."

How to cut the water footprint of your AI use

None of these will change much on their own. But they're easy:

• Ask for what you need. A short answer takes less computing than a 2,000-word one.
• Skip the regenerate button when the first answer is fine.
• Use a smaller model for simple tasks. Many chat apps let you pick a lighter one.
• Save image and video generation for when you need it. They use far more computing than text.
• Run a local model on your own computer for everyday prompts. Nothing goes to a data center, which means no data center cooling. Your computer still uses electricity, which has its own small footprint.

New to local AI? Here's what local LLMs are and how they work and how two popular tools compare in LM Studio vs Ollama.

Run AI on your own computer with Sigma

Sigma is a free browser with Eclipse built in, a local AI model that runs on your device. It works offline and your prompts stay on your computer instead of going to a cloud server. That's a privacy win first. It also means everyday questions don't touch a data center at all.

Eclipse is free and available on Windows and macOS. Its speed depends on your hardware and a local model is usually less capable than the largest cloud models. It fits quick everyday tasks best.