AI Data Center Water Usage

Every time you fire off a prompt to your favorite chatbot, a server somewhere gets hot, and something has to cool it down. A lot of the time, that something is water. Multiply that by billions of queries a day, and you start to understand why “AI data center water usage” has quietly become one of the most contested phrases in Western water policy.

A new statistics report from Axis Intelligence Research lays out the numbers in stark terms, and if you live anywhere near Arizona, New Mexico, Utah, or Colorado, you should be paying attention.

Google’s newly released 2026 Environmental Report shows the company consumed 10.9 billion gallons of water in 2025 — a 34% jump from the year before, and more than double what it used back in 2021. That’s the steepest single-year increase in the company’s disclosed history, and it lines up almost exactly with the explosion in AI model training and inference workloads.

Amazon, by contrast, made its very first water disclosure this year: 2.5 billion gallons in 2025, at a water-efficiency rate the company says is roughly seven times better than the industry average (though Axis Intelligence flags that comparison as a bit apples-to-oranges — more on that below). Microsoft sits in between at 0.30 liters per kilowatt-hour, and is betting big on “zero-water” cooling designs that don’t rely on evaporation at all.

Here’s the thing that should give everyone pause: all four hyperscalers claim they’re going “water positive” by 2030. But those pledges rest on replenishment projects scattered across watersheds that often have nothing to do with where the water is actually being pulled out of the ground. Water isn’t fungible — you can’t offset a drained aquifer in Phoenix by funding a river restoration project in Scotland. That distinction matters enormously once you zoom into the Southwest.

This isn’t an abstract global statistic. The report calls out Arizona’s West Valley specifically as a major expansion zone for Microsoft, Google, and Amazon — sitting squarely inside a multi-decade megadrought that’s already forced Arizona to cut back on Colorado River allocations. Microsoft is even piloting its new zero-water cooling technology in Phoenix, which tells you the company itself recognizes that conventional evaporative cooling is a hard sell in a region that’s rationing water for farms and families.

Widen the lens a little and the pattern holds across the broader Southwest:

• Texas, which borders the Four Corners states and shares much of the same drought exposure, is projected to see data center water consumption rocket from 49 billion to 399 billion gallons by 2030 — a 714% increase. And in a finding that should alarm anyone who trusts “self-reported sustainability,” a June 2026 Texas legislative hearing revealed that 83% of the state’s 341 data centers never even bothered to comply with mandatory water reporting. The penalty for skipping it? A maximum $500 fine — pocket change for companies spending billions on server farms.
• Nearly two-thirds of the U.S. was under drought conditions as of June 2026, per the U.S. Drought Monitor, and the states carrying most of that burden overlap heavily with the same states courting data center investment for tax revenue and jobs.

Put simply: the region with the least water to spare is the one being handed the fastest-growing industrial water demand in the country.

Here’s a number that deserves way more attention than it gets. Lawrence Berkeley National Laboratory found that U.S. data centers used 17.4 billion gallons of water directly for cooling in 2023 — but they used an estimated 211 billion gallons indirectly, through the water consumed at power plants generating their electricity. Do the math, and roughly 92.6% of the true water footprint of American data centers never shows up in a single corporate sustainability report.

Companies love touting their WUE (water usage effectiveness) scores because that number only counts what happens on-site. It conveniently ignores the fact that every kilowatt-hour of electricity a data center pulls off the grid also comes with a hidden water cost at the power plant. When a company says “our data center is water-efficient,” ask them: efficient by whose accounting?

You’ve probably seen the claim that a single ChatGPT-style query uses a fraction of a milliliter of water. That’s technically true — if you’re only counting the water evaporated on-site for cooling. A widely cited UC Riverside study found that once you include the indirect water used to generate the electricity behind that query, the real number is closer to 10–25 milliliters per query. Neither number is wrong. They’re just answering different questions, and companies tend to reach for whichever one sounds better in a press release.

Scale that up: Americans are sending an estimated 18 billion AI messages a week. At the comprehensive accounting scope, that’s over 300 million liters of water a week just keeping chatbots cool — before you even get to model training, which can burn through hundreds of thousands of liters of fresh water for a single large model.

None of this means AI has to be a water disaster. Microsoft’s zero-evaporation cooling and Amazon’s free-air cooling approach both show that meaningful reductions are possible with the right engineering and the right climate. The problem isn’t that solutions don’t exist — it’s that adoption is uneven, disclosure is voluntary, enforcement is a joke (a $500 fine, really?), and the states with the least water margin for error are absorbing the fastest growth.

If AI companies want to keep building data centers in Arizona, Texas, New Mexico, and the broader Southwest, they need to be held to a higher standard than a self-reported sustainability PDF once a year:

• Real, binding disclosure requirements with penalties that actually mean something to a trillion-dollar company — not a $500 slap on the wrist.
• Indirect water accounting, not just on-site WUE numbers, so the public actually sees the full footprint.
• Site-specific replenishment, not just fungible global offsets that let a company claim “water positive” while draining a specific, already-stressed watershed.
• Regional caution, meaning if you’re building in a megadrought zone, you don’t get to use the cheapest, most water-hungry cooling method available.

The UN Secretary-General just launched a global initiative in June 2026 demanding exactly this kind of transparency on water, carbon, and land use from every major AI company. That’s a good start. But if you live in the Four Corners states, you shouldn’t have to wait on a UN initiative or a state legislature’s five-year fight over enforcement penalties to know whether your local aquifer is being drawn down to cool somebody else’s chatbot.

AI is genuinely useful. It’s also thirsty. In the driest part of the country, those two facts are on a collision course — and the companies building this infrastructure need to start acting like it.

Source: AI Data Center Water Usage Statistics 2026, Axis Intelligence Research, updated July 23, 2026.