Most coverage of AI infrastructure focuses on electricity. The more revealing story is heat — specifically, that there is no cheap way to get rid of it.
Cooling a modern AI data center requires either a great deal of water or a great deal of extra electricity. Not both, and not neither. Every operator has to pick a side of that trade, and increasingly the community hosting the building has opinions about which side they picked.
The trade-off, stated plainly
GPU racks run hot enough that air alone struggles to carry the heat away. That leaves two broad approaches, and they pull in opposite directions.
Evaporative and water-based cooling is energy-efficient. It is also thirsty: roughly 80% of the water withdrawn evaporates rather than returning to the system, and the remainder comes back as warm wastewater that local treatment plants have to handle.
Air and dry cooling conserves water almost entirely. It also demands substantially more electricity, which raises both emissions and the strain on a grid that is already the industry’s tightest constraint.
Cooling typically accounts for somewhere between a fifth and two-fifths of a data center’s total energy use, so the choice is not marginal. And the right answer is genuinely local: dry cooling makes sense in a water-stressed county in Arizona, while evaporative cooling may be the better option where power capacity is scarce and the grid is carbon-heavy.
Save water and your power bill climbs. Save power and your water draw climbs. The industry calls this optimization. Communities on the receiving end tend to call it a shell game.
What the numbers actually look like
| Measure | Figure |
|---|---|
| Typical data center, daily water | ~300,000 gallons (about 1,000 households) |
| Large facility, daily water | up to 5 million gallons (a town of ~50,000) |
| US data centers, direct annual use | ~17–19 billion gallons |
| Projected direct use by 2030 | ~60–110 billion gallons |
| Indirect use via electricity (2023) | ~211 billion gallons |
| Water evaporated, not returned | ~80% of withdrawals |
The indirect figure is the one most often missed. Thermal power plants consume enormous quantities of water to generate electricity, which means a facility running “water-free” cooling may simply be moving its water footprint upstream to the generating station. By some estimates that indirect share is the majority of a data center’s true water footprint.
Chip manufacturing adds another hidden layer. Fabs require ultrapure water to rinse silicon without damaging it, and a typical plant consumes millions of gallons daily before a single chip ships.
The politics have moved fast
Public opinion has hardened. A Gallup poll conducted in March 2026 found that 70% of respondents opposed construction of new AI data centers in their own neighborhood — an unusually bipartisan result.
That sentiment is now producing outcomes. Data Center Watch, which tracks grassroots opposition, reported that at least 75 projects worth roughly $130 billion were disrupted by local opposition in the first quarter of 2026 alone — exceeding the total for all of 2025. Lawmakers in more than 30 states introduced over 300 data-center-related bills in 2026, covering moratoriums, tax incentives and energy policy, with bans under discussion in more than 20 states.
Why trust is thin
Two recurring patterns explain a lot of the local anger.
- Disclosure gaps. In the Netherlands, local reporting found a Microsoft facility had used over 22 million gallons in 2021 — several times what residents had been told to expect.
- Metering gaps. In Fayetteville, Georgia, a data center campus drew roughly 30 million gallons from the local supply before receiving a utility bill. The operator denied improper use.
Whatever the merits of any single project, communities that feel they were given numbers that later moved tend not to extend the benefit of the doubt to the next application.
The case that the panic is overblown
This is where the debate deserves a fair hearing on both sides, because the critics of the critics have real arguments.
The most cited: US suburban lawn irrigation consumes more water in two days than every data center in the country uses in a year. Golf courses and individual restaurants often out-consume a single facility, without generating comparable objections. On land, projections suggest total US data center footprint by 2030 would fit roughly within the area of Rhode Island — and the buildings themselves within a mid-sized city.
Proponents also note that evaporated water is not destroyed; it returns as precipitation within roughly nine days, though not necessarily to the watershed it left.
And the technology is moving. Closed-loop systems can cut freshwater use by as much as 70%. Meta’s El Paso facility is designed to use zero water for most of the year, and the company has committed to restoring 200% of what it consumes there. Amazon reported cutting North American water use by 946 million liters in 2024 while improving water efficiency by 17%.
Skeptics counter that “zero water” claims are partly marketing — closed loops still lose water and shift the burden onto electricity — and that voluntary corporate targets are not the same as enforceable limits.
Where this is heading
A few things look reasonably clear regardless of where you land:
- Siting will matter more than technology. Putting evaporative cooling in a drought-stressed basin is a choice, not a necessity.
- Disclosure is becoming table stakes. Projected and actual water use, reported publicly, is the least controversial ask in the entire debate.
- Cost allocation is the real fight. Who pays for the distribution lines, the treatment capacity, the grid upgrades — the developer or the ratepayer.
- The regulatory patchwork will grow. Water is regulated state by state, and Arizona’s problem is not Pennsylvania’s.
The heat has to go somewhere. The open question is not whether AI infrastructure imposes a cost, but who is asked to absorb it — and whether they were told the real number before the building went up.
Leave a Reply