AI Data Centers' Water Bill Is Mostly a Power Bill — and the New City Bans Are Pricing That In

On the evening of August 27, 2026, Austin's City Council voted unanimously to fast-track rules that could keep large new data centers out of the city, directing staff to return with code amendments by December. Austin's zoning code currently treats data centers as communications services and accommodates facilities up to roughly 20 megawatts; hyperscale projects can draw more than 100 megawatts, about the load of 25,000 homes. The city is the latest in a widening list: San Marcos, Texas, barred data centers by zoning in June; Durham, North Carolina, and Cave City, Kentucky, imposed moratoriums; Jersey City banned data centers as the primary use of industrial property.

What turned local complaints into an organized push is a study. On August 25, the nonprofit Ceres published Water Behind the Watts, which estimates that data centers in the seven states with the most of them — Virginia, Texas, California, Illinois, Georgia, Ohio, and Arizona, home to roughly half of all U.S. data centers — depend on power generation that withdraws about 3.4 trillion gallons of freshwater a year, over 12 times the combined annual water use of Los Angeles, Phoenix, and Washington, D.C., and more than California's average annual urban water demand. The report also found that two-thirds of the water-using power plants sit in places under medium-high to extreme water stress.

Now the decomposition, because the headline number is doing work the study does not support. That 3.4 trillion gallons is not water splashed onto cooling towers at the facility. It is water drawn at the power plants that make the electricity the data center buys — the report's title says it, the water is behind the watt. The International Energy Agency estimates that roughly 60% of a data center's water consumption is indirect, embedded in power generation, versus direct on-site use. The visible fight over server cooling is the minority share of the bill, and it is the share with an engineering fix.

That fix is where a real third path between "don't build" and "build anyway" lives, and it is a trade rather than a solution. Standard evaporative cooling towers evaporate about 80% of the water they draw, and a large facility can use up to 5 million gallons a day, about the daily water of a town of 50,000. Closed-loop and liquid cooling that pipe coolant to the chips can cut direct freshwater use by as much as 70%. But those systems draw more electricity, so every gallon saved on-site is converted into extra kilowatt-hours — and if those kilowatt-hours come from a thermal plant, the water reappears upstream. The fuel mix matters: coal-fired generation withdraws roughly 19,000 gallons per megawatt-hour, natural gas about 2,800, and wind and solar essentially none. Solve the water problem by changing the cooling and you have imported a power problem; change the power source and most of the water leaves with it. For investment purposes, water and power are the same constraint billed in different units.

That is why the geography matters more than the gallons. A Guardian analysis in June found that 517 of 809 planned U.S. data centers — about two-thirds — sit in areas that were under drought over the prior year, and Bloomberg and World Resources Institute data found a similar share of new centers built since 2022 went into high-water-stress areas. In Texas, data centers could approach 9% of the state's water use by 2040. Deployments are landing in exactly the counties with the worst scarcity, which is why city-level resistance is spreading — and why it reads, structurally, as siting scarcity: the same constraint that grid interconnects already impose on power, now with water attached. It is a growing cap on new supply in the most important regions.

So where does the money move? Three places, and they overlap. First, cooling equipment: water limits switch capex away from cheap evaporative towers toward liquid and closed-loop systems — the demand backdrop behind Vertiv's economics, roughly 26% revenue growth year over year on a ~26% return on invested capital and a thin fixed-asset base. Second, water-free power: in drought counties, wind and solar stop being merely "clean" and become water-savings assets, and utilities are where investor money is already rotating — the main utilities ETF drew roughly $0.7 billion of net inflows over three months through late August. Third, existing capacity: an operator holding a permitted, water-secured site holds pricing power while new supply gets slower and more expensive. The market's pocketbook is only half-committed to that last one: the data center REIT ETF has seen net redemptions this year, about $49 million against a roughly $352 million fund, even as the biggest operators spend heavily — Equinix has run about $6.5 billion in trailing capital expenditure with deeply negative free cash flow. Siting scarcity is not the market's base case yet; that gap between what the stocks price and what the water data imply is where the analysis lives.

Underneath all of it is a measurement problem. The number an investor would need — how much water each build's electricity actually costs — is barely reported. Ceres identified Meta as the only hyperscaler that reports the indirect water behind its electricity, with 2024 indirect use more than 20 times its on-site consumption, and it estimates those indirect withdrawals could rise to between 4 trillion and 8 trillion gallons by 2030. The bans are the visible part; the undisclosed water cost is the systematic gap that keeps the market in its electricity-only framing. The concrete tells going forward: Austin's code amendments land in December, which operators start splitting reported "water" into direct and indirect, and whether leases and power contracts in the stressed states start marking up water-secure sites and zero-water generation as scarcity bites. The constraint is real, and its price is still being discovered.