Every time you ask ChatGPT a question, a data center somewhere burns through roughly half a liter of freshwater to help make it happen. That’s not a metaphor. Researchers at UC Riverside measured it. And as more Americans reach for AI tools every day to write emails, answer medical questions, prep for job interviews, or understand a lease — the water tab for all that convenience is adding up faster than most people realize.

The connection between your phone screen and a town’s water supply might sound abstract. For families living near data centers in Arizona, Nevada, or Texas, it’s already hitting home.

So Where Does the Water Actually Go?

AI doesn’t drink water the way a person does. But the massive buildings that run AI called data centers do, and the reason is straightforward: computers generate heat, and that heat has to go somewhere.

When you type a question into ChatGPT, that question gets processed by thousands of powerful chips firing simultaneously. Those chips produce enormous amounts of heat. Much of that heat is removed by cooling systems that pull in water, run it through the hot equipment, and release that heat by letting the water evaporate like sweating, but on an industrial scale.

The Two Water Drains You Don’t See

There are actually two ways AI consumes water, not one. The first is the cooling systems inside data centers. Water gets pulled directly from municipal supplies or groundwater, cycled through the hot equipment, and released as steam. About 80% of it evaporates and never comes back.

The second drain is indirect the power plants that generate electricity to run those data centers. Power plants burn fuel to make steam, which turns turbines, which generates electricity. Water is part of that process at every step. By the time a data center gets the power it needs and uses it, water has been consumed twice before you see a single answer on your screen.

How Much Water Does AI Use? The Real Numbers

The numbers that researchers have actually measured are striking.

What the UC Riverside Study Found

A team from UC Riverside’s Bourns College of Engineering led by associate professor Shaolei Ren  published one of the most-cited studies on this topic. Their finding: running 20 to 50 questions through ChatGPT, a typical back-and-forth conversation, consumes roughly 500 milliliters of freshwater. That’s the size of a standard water bottle.

For a single prompt of about 100 words, the estimate lands around 519 milliliters. More than a 16-ounce bottle, for one question. Those numbers cover both the cooling towers at Microsoft’s US data centers and the indirect water use from electricity generation.

The Scale Across the Whole Industry

Zoom out from a single conversation and the scale changes completely.

Sources: Yale E360, UC Riverside, UK Government Digital Sustainability Alliance

For comparison: that large data center number 5 million gallons per day — is roughly what a city of 50,000 people uses.

Why Are Data Centers Built in Deserts?

This is where things stop making sense on the surface. You’d expect water-hungry facilities to be placed near rivers or in rainfall-heavy regions. Instead, roughly two-thirds of data centers built in the US since 2022 have been placed in water-stressed regions, according to a Bloomberg analysis. That includes Arizona, Nevada, and parts of Texas states already living with drought and shrinking groundwater supplies.

The reasons aren’t mysterious. Land is cheaper in these areas. Power infrastructure already exists. Local governments often offer tax breaks to attract the construction investment and jobs. But the water math doesn’t add up for communities that have been managing scarce supplies for decades.

In Arizona, data centers now account for 7.4% of the state’s total power consumption. Texas data centers are projected to use 49 billion gallons of water in 2025 alone — and up to 399 billion gallons by 2030, according to research from the University of Houston.

What This Means for the Town Near You

If you live within range of one of these facilities, the impact isn’t hypothetical. Several communities across the Mountain West have raised concerns about data centers competing with agriculture, residents, and wildlife for the same water supply. A 2026 study published in April found that data center water use has become increasingly hard to track, because many facilities don’t disclose how much they pull from local sources. Researchers described a growing mismatch between how quickly data centers are being built and how slowly water infrastructure can adapt.

That transparency gap has real consequences. US drinking water and wastewater systems already face a $744 billion backlog in deferred maintenance, according to Brookings Institution research. Adding a massive new water user to a strained local system doesn’t just add pressure — it can accelerate a crisis that was already quietly building.

For people in drought-prone areas who watch their water bills, restrict lawn watering, or rely on well water: a new data center in your county is now a real variable in your local water picture.

What Tech Companies Are Actually Doing About It

Google, Microsoft, and Amazon have all pledged to become “water positive” by 2030 — meaning they aim to return more water to communities and watersheds than they consume. Whether they’ll hit those targets is a separate question, but some concrete steps are already underway.

The Closed-Loop Fix Microsoft Is Testing

Microsoft developed a data center cooling design that breaks from the standard approach. Instead of an open evaporative system where water disappears into the air, their design circulates the same water continuously between servers and chillers. Nothing evaporates. The same water stays in the loop, used over and over.

They’re rolling this out at facilities in Wisconsin and Arizona starting in 2026. If it works at scale, it could reduce freshwater consumption by up to 70% compared to standard cooling towers.

Amazon’s Treated Wastewater Experiment

Amazon took a different approach: instead of trying to use less water, they changed the source. Twenty-four of Amazon’s data centers already cool themselves using treated sewage water — recycled from local wastewater systems — rather than drinking-quality freshwater. Amazon is working with local utilities to expand this approach significantly across its network.

Other methods being explored include immersion cooling, where server components are submerged in non-conductive liquid instead of relying on water-based cooling entirely, and simply locating new data centers in cooler, rainier regions where passive cooling requires far less water to begin with.

Can You Do Anything About It?

Realistically, not much on an individual level. Skipping a ChatGPT session won’t meaningfully change a data center’s water draw. The scale mismatch between one person’s usage and industrial consumption is too large for individual choices to move the needle.

What does matter is policy. Organizations including the Environmental Law Institute and the Environmental and Energy Study Institute are pushing for mandatory water disclosure from data centers — similar to how public companies report carbon emissions. The more public attention this issue gets, the more likely that kind of transparency requirement becomes law.

If you run a small business and you’re choosing between AI tools, it’s worth knowing that Microsoft and Amazon have been more transparent about their water use metrics than most companies in this space. That’s not an endorsement of either platform — it’s information worth having when you’re making decisions about which AI services you pay for.

Frequently Asked Questions

How much water does one ChatGPT question use?

According to researchers at UC Riverside, a conversation of 20 to 50 questions with ChatGPT uses approximately 500 milliliters of freshwater — roughly one standard water bottle. That works out to about 10 to 25 milliliters per individual question, though the exact amount varies depending on which data center handles your request and local weather conditions affecting cooling efficiency.

Does using AI on my phone use water?

Yes, indirectly. When you use a phone app that runs on AI including voice assistants, image-editing tools, or chatbots your request travels to a remote data center where the actual computing happens. That data center uses water to stay cool. Your phone itself doesn’t consume water, but the infrastructure behind every AI response does.

Which states are most affected by data center water use?

Arizona, Texas, and Nevada are among the most affected. These states are already managing water scarcity, yet they’ve attracted a significant share of new data center construction because of cheaper land, available power, and tax incentives offered by local governments. Texas data centers alone are projected to use nearly 50 billion gallons of water in 2025.

Is AI water usage worse than other industries?

It’s growing faster than almost any other industry’s water demand right now. Agriculture, mining, and manufacturing have long been large water consumers. What’s different about AI data centers is the rate of growth — consumption is projected to multiply several times over within just a few years, in a way that most established industries aren’t. That speed of expansion is what’s catching the attention of environmental researchers and regulators.

The Water Clock Is Running

The water question isn’t going away. Every month, more data centers go up. Every month, more Americans start using AI tools that didn’t exist a year ago. The technology is scaling faster than the infrastructure managing its environmental footprint and faster than the regulations designed to hold it accountable.

Half a liter per ChatGPT session may shrink as models become more efficient and cooling technology improves. But right now, in April 2026, the thirst is real. It’s showing up in towns in Arizona and Nevada that already know what water scarcity feels like, and in aquifer levels and water bills that don’t make national headlines but matter deeply to the people living there. Knowing about it is at least a start.