AI’s Thirst and the American City

Ibrahim Khan — July 14, 2026

Data Centers, Water, and the Limits of a Viral Statistic

Ask ten people how much water artificial intelligence uses, and you’ll get answers that don’t even sound like they’re talking about the same planet. One camp will tell you it’s basically nothing; the other will claim AI is practically drinking the Colorado River dry. People frequently struggle to reconcile these divergent viewpoints, leading to a widespread narrative conflict regarding the true environmental weight of cloud infrastructure.

Both sides can point to real numbers. It’s a battle over which fact matters most. If you’re trying to understand this debate, the goal shouldn’t be to memorize a single, shocking statistic; it’s understanding why the statistics are at war with one another.

The Battle of the Droplets

The viral number everyone quotes comes from a 2023 UC Riverside study led by researcher Shaolei Ren. It found that roughly 500 milliliters (about one standard water bottle) evaporates for every 10 to 50 mid-length ChatGPT prompts. That includes both the water evaporated on-site for cooling and the water consumed upstream at the power plants keeping the lights on. When the Washington Post repackaged this into a sleek 2024 visualization, showing that a single 100-word GPT-4 email gulps down a bottle of water, the stat exploded across TikTok and cable news.

Tech executives quickly fired back. Sam Altman and industry giants countered with a microscopic figure: roughly 0.3 mL per query. Google echoed this in its annual sustainability report, claiming a median Gemini text prompt takes about 0.26 mL, literally five drops of water.

Neither side is lying; they’re just measuring completely different rings of the same target. In the tech world, water usage is tracked in three scopes:

Scope 1: Water evaporated right on-site to keep servers from overheating.
Scope 2: Water consumed off-site to generate the electricity the data center draws.
Scope 3: “Embodied” water in the supply chain, like the thousands of gallons of ultrapure water needed just to fabricate a single silicon wafer.

Altman’s “five drops” is a hyper-narrow Scope 1 estimate on ultra-efficient, modern hardware. The “water bottle” figure is a combination of Scope 1 and Scope 2, calculated using 2023-era hardware at a standard U.S. electricity grid mix.

Both are technically true, but neither is universal. The exact same query costs virtually nothing in a Nordic data center cooled by freezing fjord water and running on hydropower, but sky-rockets in a Phoenix facility drawing ancient groundwater while running on a gas-heavy grid at 3:00 PM in the dead of July. Location, time of day, hardware generation, and cooling tech can swing the actual water cost by a factor of a thousand. To witness this stark contrast, look to facilities like Norway’s Lefdal Mine Datacenter, which achieves near-zero freshwater use using cold seawater siphon systems.

Furthermore, some of the most shocking numbers in circulation assume energy-per-query figures that independent researchers (like Epoch AI) have found to be inflated. Anyone citing “a bottle of water per email” without these caveats is likely treating a worst-case scenario as an absolute rule.

Why the National Number Doesn’t Matter

National averages are irrelevant to local political fights.

Statistically, U.S. data centers directly consumed about 17 billion gallons of water in 2023. That sounds staggering until you realize that U.S. golf courses consume an average of 312,000 gallons daily per facility, adding up to billions of gallons every week, and residential lawns soak up another 9 billion daily. On a national pie chart, AI is a tiny sliver of freshwater withdrawal.

But nobody experiences “national freshwater withdrawal.” People experience their own kitchen tap.

The AI water crisis doesn’t live in macro-percentages; it lives in specific counties where a massive new tech campus can, almost overnight, become the single largest industrial water user in the region. A 2025 Bloomberg investigation found that over 70% of new data center projects since 2022 were built or proposed in communities already facing severe water stress. Tech companies aren’t chasing water; they are chasing cheap land, tax breaks, and proximity to power grids.

The real debate isn’t “how much water does AI use?” The real question is: Who decides which communities absorb that burden, and under what terms?

Flashpoints Across the Map

Look at the local cities, and you’ll see the exact same friction points repeating themselves:

Fayetteville, Georgia: In a suburban subdivision south of Atlanta, residents suddenly noticed their water pressure plummeting. Investigators traced the drop to a massive QTS data center campus. The facility had drawn nearly 30 million gallons through industrial hookups, one of which was installed without the utility’s knowledge, without paying a dime for it. This happened while the county was begging residents to stop watering their lawns during a historic statewide drought. To those who advocate against data centers, grievances like this capture the asymmetric water usage of everyday citizens and tech giants.

Tucson, Arizona: In a desert region that sees less than 10 inches of rain a year, the city council unanimously blocked Amazon’s “Project Blue” data center complex. Tensions boiled over when a contractor was caught hauling municipal construction water outside city limits for dust control without authorization. It wasn’t a massive volume of water, but it was symbolically explosive: the tech industry taking the exact water the city had explicitly chosen to withhold. Public concern remains incredibly high because the Project Blue development is projected to directly consume up to 31 million gallons of water annually, not including the massive hidden hydrologic footprint of the regional energy needed to power it.

Memphis, Tennessee: xAI’s Colossus supercomputer is certainly a feat of engineering. Built in just 122 days, its scale is sprawling: a peak water demand near 5 million gallons a day and an energy footprint equal to 1.5 million homes. The facility sits next to a majority-Black community already burdened by industrial pollution. When regulatory friction mounted over unpermitted gas turbines, xAI touted an $80 million water-recycling plant that would cool the servers using treated municipal sewage instead of drawing from the precious Memphis Sand Aquifer. But in early 2026, xAI abruptly paused construction on the recycling plant to “prioritize other projects,” sparking local outrage while continuing to pump millions of gallons of pristine drinking water from the aquifer. Local conservationists at Protect Our Aquifer raised concerns when xAI paused construction on the graywater plant to prioritize other expansion projects, leaving the facility drawing millions of gallons of pristine drinking water directly from the local aquifer.

Texas: The sheer scale of the coming boom seems daunting on paper. A joint university study projects data center water use in Texas to skyrocket from 49 billion gallons in 2025 to a staggering 399 billion gallons by 2030. This is the equivalent of draining Lake Mead by 16 feet in a single year, which could be critical in a state already plagued by municipal water emergencies.

Engineering the Way Out

The tech industry’s best defense isn’t that they don’t use water, it’s that water intensity is a solvable engineering problem that they are actively working to fix.

Traditional evaporative cooling is the most water-consuming method; it lets water absorb server heat and carry it away as vapor (essentially making servers sweat). But newer architectures have the potential to change the game:

Closed-loop and direct-to-chip cooling recirculate a fixed volume of fluid instead of letting it evaporate, slashing water use at the expense of higher electric bills.
Air-cooled designs in cold climates (like Nordic facilities using near-freezing outside air or seawater) can bring freshwater consumption down to practically zero.
Wastewater reuse: treating municipal sewage to an industrial standard—allows data centers to cool their systems without touching a single drop of local drinking water.

The technology is being deployed right now. The unresolved question is whether tech companies will adopt it fast enough out of corporate goodwill, or if they will only do it when a local government forces their hand. This conflict is accelerating as federal agencies push guidelines like the EPA’s Water Reuse Action Plan 2.0 to prioritize recycled municipal wastewater in industrial facilities.

The Regulatory Backlash

Public anxieties have begun to harden into binding law. Because the federal government has been slow to act, a chaotic patchwork of state and local regulations has emerged:

Roughly a dozen states have data center moratorium bills on the table.
Seattle froze new large-facility sites pending an intensive impact study.
Local Colorado counties passed emergency moratoriums after a coalition of 238 tech lobbyists successfully killed a basic water-disclosure bill at the state level.
California is pushing legislation (AB 2619) to tie business licensing directly to water-use disclosure, while states like South Carolina are considering outright mandates for zero-net-withdrawal closed-loop cooling.

Across the country, the underlying theme is clear: local governments are moving faster than tech developers are disclosing data. Simply forcing these facilities to reveal how much water they actually use has become the first major political battleground. This push for legislative disclosure highlights a critical shift as municipal networks refuse to bear the unmetered brunt of AI’s physical footprint.

The Real Takeaway

The strongest version of this argument avoids both extremes.

To say “AI is an existential threat to the global water supply” is an overstatement that ignores how small data centers are on the national tapestry. But to claim “AI’s water use is a media myth” willfully ignores documented, localized devastation to specific aquifers, taxpayers, and communities that were not given a say in the process.

The truth is conditional: AI’s water footprint is nationally modest but locally severe. It is being concentrated by economic incentives in the exact places least equipped to handle it: drought-stressed, low-income, and industrially zoned communities. Ultimately, the conflict over data centers’ water usage isn’t about the size of the statistic. It’s about who is allowed to see the numbers before the wells run dry.

Read more here:

AI’s Thirst and the American City