An AI Data Center's Real Water Bill Isn't in the Cooling Towers. It's at the Power Plant.
Everyone worried about AI's water use pictures a cooling tower. A new seven-state analysis finds the bigger draw is one step upstream, in the plants generating the electricity — and Meta's own disclosed numbers show it can run more than 20 times larger.
The invisible water bill
A power plant that isn't wind or solar boils water to spin a turbine, then has to condense that steam back to liquid — usually by evaporating water in a cooling tower. Drag the slider to set how much of a data center's electricity comes from that kind of plant, and watch its indirect water footprint move against its fixed on-site cooling. Model: a 100 MW facility running continuously for a year (876,000 MWh); on-site cooling fixed at ~530,000 gal/day, a commonly cited average for a facility that size (≈193.4M gal/yr); steam-cycle thermal generation (coal, gas, or nuclear) assumed to consume ~500 gal per MWh, the mid-range of published closed-loop cooling estimates; wind and solar assumed ~0, since neither has a condenser step. Simplified and illustrative, not one specific facility's real bill.
Set the grid's thermal share, or pick a preset:
For comparison: Meta, the only hyperscaler that discloses this number, reported its real 2024 ratio at over 20× — well above anything this simplified model produces, because the real figure reflects Meta's actual regional grid mix, plant-level efficiency, and a fuller water-accounting scope than a two-input model like this one can capture.
The plain version
When people worry about AI data centers and water, they picture the cooling towers on the roof — the equipment keeping thousands of GPUs from overheating. That part is real, but it's not the biggest number. A new analysis from the sustainability nonprofit Ceres, published September 3, found that a data center's bigger water bill sits one step removed: it's baked into the electricity itself.
Here's why. Most power plants that aren't wind or solar work like a steam engine at heart: burn fuel or split atoms to boil water, use the steam to spin a turbine, then — critically — turn that steam back into water so it can be boiled again. That last step, condensing steam, is usually done by evaporating water in a cooling tower at the power plant. So every kilowatt-hour a data center pulls from a coal, gas, or nuclear plant carries a hidden water cost that never touches the data center's own property line.
Ceres found that data centers in just seven states — Virginia, Texas, California, Illinois, Georgia, Ohio, and Arizona, which host roughly half the country's data centers — indirectly draw about 3.4 trillion gallons of freshwater a year this way, more than they use directly on-site. Meta is the only major tech company that discloses this number at all, and its 2024 figure was more than 20 times its direct water use. Worse, 66% of the electricity behind these data centers comes from regions already short on water.
The fix isn't a better cooling tower. It's what's plugged into the wall — because wind and solar need almost no water to make electricity in the first place.
The expert version
The mechanism is thermodynamic, not incidental. Coal, natural-gas steam and combined-cycle, and nuclear plants generate electricity via a Rankine cycle: fuel or fission heat boils water into high-pressure steam that expands through a turbine, and the exhausted steam must then be condensed back to liquid before it can be reboiled. That condensation step is the water-intensive one — typically handled by a wet cooling tower, which rejects waste heat by evaporating a portion of its circulating water into the atmosphere rather than returning it all to the source. Published estimates for closed-loop (recirculating) systems put coal and nuclear plants at roughly 500-1,100 gallons of water consumed per MWh generated; natural-gas combined-cycle plants run lower, near 200 gal/MWh, since more of their output comes from the gas turbine's direct combustion stage rather than a steam bottoming cycle. Wind turbines and solar PV panels have no condenser step at all and consume water only incidentally, for occasional panel washing.
Ceres' "Water Behind the Watts" report (published Sept. 3, 2026) quantifies the resulting exposure: data centers across seven states — Virginia, Texas, California, Illinois, Georgia, Ohio, and Arizona, which together host roughly half of US data center capacity — draw an estimated 3.4 trillion gallons of freshwater annually through the electricity they consume, exceeding their direct operational water use. Sixty-six percent of the electricity generated in those states comes from grid regions Ceres classifies as facing medium-high to extremely high baseline water stress. Meta is currently the sole hyperscaler disclosing embedded, indirect water consumption tied to its purchased electricity; its reported 2024 figure was 19 billion gallons — more than 20 times its direct operational consumption, and it has grown every year since 2021. Ceres projects total indirect consumption across the sector could increase by a further 4.1 to 7.6 trillion gallons by 2030, absent a shift toward lower-water generation (renewables, combined-cycle gas) or broader adoption of air-cooled (dry) condensers, which cut water use at the cost of reduced thermal efficiency and higher capital expense. Separately, a Guardian analysis of NOAA drought data found roughly two-thirds of 809 planned US data centers sit on land classified as drought-affected in the past year — meaning the sites drawing hardest on this indirect water math are disproportionately the ones with the least water margin to spare.