The Water Bill of Intelligence: How AI Data Centers Tap Cities' Drinking Supplies

01 Evaporation Is the Price of Cool

A data center is, thermodynamically, a very large electric kettle that must never boil. Every watt a server draws eventually becomes heat, and that heat has to go somewhere. The cheapest way to move a lot of heat with a small amount of electricity is to let water carry it away as vapor: as droplets flash to steam in a cooling tower, they take their latent heat of vaporization with them into the atmosphere. That physics is why the industry's water problem exists at all. Evaporative cooling is not a design mistake — it is a rational engineering choice that happens to consume a substance cities also need to drink.

The mechanism deserves precision because the rhetoric around it rarely has any. In a recirculating loop, warm water from the data hall is sprayed over fill material inside the tower while fans pull air upward. A fraction of the water evaporates; the cooled remainder falls to a basin and returns to the heat exchangers. What is lost to the sky must be replaced — the industry calls this make-up water — and because the minerals left behind by evaporation concentrate in the loop, more water is periodically flushed out as blowdown. The result is that a facility can be a modest withdrawer of water but a genuine consumer of it: unlike a power plant that returns most of its withdrawal to the river, a cooling tower sends its water into the air, effectively subtracting it from the local watershed.

02 How Much Water AI Drinks — and Why Nobody Truly Knows

The honest answer is that precise, facility-level numbers are scarce, so the public conversation runs on ranges. Widely reported estimates — most influentially from academic work on AI's water footprint, including the 2023 "Making AI Less Thirsty" analysis from UC Riverside researchers — suggest that a single AI interaction can carry a water footprint on the order of tens of milliliters to hundreds of milliliters once on-site cooling is counted, with some estimates later revising downward for cooler climates and more efficient hardware. Scaled across billions of queries, those milliliters become reservoirs. Reported corporate disclosures point the same direction: annual tech-sector water consumption figures published by major operators run to billions of gallons, with sharp year-over-year increases coinciding with AI buildouts. These are magnitudes, not measurements of your chat session, and they should be read as such.

Why do estimates vary so much? Three multipliers sit between a chat prompt and a real gallon. First, location: a campus evaporating water in a hot, dry climate can consume multiples of what an identical campus uses in a cool one, because the loop rejects less heat by evaporation when the air is already cold and damp. Second, accounting boundary: some estimates count only on-site cooling, others include the water evaporated upstream at power plants to make the electricity, and off-site consumption can rival or exceed on-site use. Third, the technology mix keeps changing — a training run that happened before an operator shifted to closed-loop or reclaimed-water systems simply does not describe the current fleet. Any confident single number is a choice about those three assumptions dressed up as a fact.

03 Where the Water Comes From

The reason this is a municipal story and not just an engineering one is siting. Data centers cluster where fiber, land, power, and tax incentives are, and those corridors frequently overlap with towns whose water systems were sized for the households and farms already there. A campus that draws a million gallons a day is, in effect, a new mid-sized town arriving on the pipes of a small one. Where the utility has surplus capacity this is absorbable; where aquifers are already over-drafted or drought restrictions are in force, the same connection request becomes a political event. The BBC World Service report that occasioned this article walks through exactly that collision in communities discovering that a facility down the road shares their potable supply.

The distinction cities should care about is between withdrawal and consumption. A facility that pulls water, uses it, and returns it to the system is a different neighbor from one that pulls water and evaporates it. Cooling towers consume; closed loops mostly recirculate; reclaimed-water schemes substitute non-potable supply for potable. A city negotiating a data center contract is therefore really negotiating over three numbers: peak daily draw, consumptive loss, and what happens in a drought year — because peak demand for cooling and peak stress on a reservoir tend to arrive in the same hot summer.

04 The Transparency Fight

Most of the public numbers that exist exist because someone demanded them. Water consumption is a sensitive disclosure for operators for competitive and commercial reasons, and reporting standards have historically let facilities lump water into sustainability summaries without local granularity. As a result, some communities have learned the scale of a neighbor's draw only after the fact, through utility records, permit applications, or journalism. That gap has produced an emerging pattern of local disputes: town hearings over permit expansions, records requests about summer draw during restrictions, and activist campaigns that turn water into the axis of opposition to AI buildouts. Where operators publish utility-level consumption voluntarily — as several major companies now do — the friction drops noticeably.

The BBC's framing is a fair one: this is a story about visibility as much as volume. The total water used by data centers remains small next to agriculture, but nobody's irrigation system sits inside a city's drinking-water utility the way a data center campus does. Precise, timely, facility-level disclosure is the single cheapest intervention available — it lets a utility plan, lets a city negotiate, and lets the public argue about real numbers instead of magnitudes.

05 What Operators Are Changing

The engineering responses are real, though slower than the headlines. Three approaches dominate. Air-side and dry cooling rejects heat through radiators and fans instead of evaporation, cutting water consumption dramatically at the cost of more electricity and reduced efficiency on the hottest days — a water-for-energy trade that makes sense in some climates and not others. Closed-loop liquid cooling, increasingly practical as servers run hotter, moves heat into sealed coolant loops that consume very little make-up water at the facility itself. Reclaimed and non-potable sources — treated wastewater, greywater, even seawater in some coastal designs — keep the cooling physics while removing the campus from the potable supply entirely. None of these is exotic; each is a redeployment of industrial cooling practice older than the industry adopting it.

The catch is that each fix relocates the problem rather than deleting it. Dry cooling raises the energy bill and, where the grid is thermal, can increase off-site water use at power plants — an accounting shell game if the boundary is drawn too narrowly. Reclaimed-water retrofits depend on municipal treatment capacity that a small town may not have. And the newest constraint is the silicon itself: modern accelerator racks run far hotter than the air-cooled enclosures a 2015 facility was designed around, which is why liquid cooling is spreading for reasons that are at least as much about chips as about water. The direction of travel is toward cooling designs chosen per climate, with water consumption as an explicit siting criterion alongside power and fiber.

06 An Honest Accounting of the Uncertainty

It is worth stating plainly what this article does not know. We do not know the water consumption of any individual AI facility beyond what operators disclose, and disclosure granularity varies from excellent to absent. We do not know how much of the widely cited per-query water figures applies to 2026 hardware, because the studies behind them measured earlier generations, and the numbers are strongly climate-dependent. We do not know the off-site water cost of the additional electricity in dry-cooled designs without knowing each grid's generation mix. And we do not know how the sector total is growing in real time, because the fleet is expanding faster than the reporting that measures it.

What can be said with confidence is narrower and still matters. Evaporative cooling genuinely consumes potable water in many facilities today. The consumption is locally significant — material to utilities, not to national hydrology. The engineering alternatives exist, are being deployed unevenly, and each moves the cost somewhere else. And the binding constraint on the next five years of this argument is disclosure: without facility-level numbers, every participant in the debate — operator, utility, regulator, resident — is negotiating with ranges instead of meters.

07 What to Watch

Three developments will decide whether AI's water question matures into a solved engineering discipline or a chronic source of local conflict. First, disclosure: watch for facility-level water reporting becoming standard in environmental disclosures, and for utilities in data center corridors publishing draw data rather than waiting for records requests. Second, siting criteria: when reclaimed-water availability starts appearing in site selection press releases alongside megawatts and latency, the market will have internalized the problem. Third, silicon: as rack densities keep rising, watch whether liquid cooling adoption — driven by physics — quietly retires the cooling tower debate faster than any regulation could.

The fair summary is that AI does not drink the world dry, but it can drink a town's margin of safety, and in a warming decade that margin is exactly what is scarce. The industry has the tools to take itself off the potable supply where it matters most; what it has historically lacked is the obligation to show anyone the numbers. That obligation is now being written, permit by permit, hearing by hearing — mostly by the residents who never expected to be the ones paying the water bill of intelligence.