AI Could Drink 600 Billion Gallons of Water by 2030, and Power Generation, Not Cooling, Is the Real Culprit

Data center water consumption has become one of the most contentious points in the fight over where, and whether, to build the next wave of AI infrastructure. Until recently the numbers were fuzzy enough that both sides could argue past each other. A cluster of reports published over the past year has now put harder figures on the table, and the picture they draw reframes the whole debate. The water problem is real and growing fast, but it is mostly not a cooling problem. It is a power problem.

The headline numbers

Start with what is verifiable on U.S. soil. A MostPolicyInitiative report puts direct water consumption by American data centers at 17.4 billion gallons in 2023, and that figure predates the gigawatt-scale buildouts now breaking ground. Direct consumption could rise by as much as 73 billion gallons by 2028 as those facilities come online. The word doing the heavy lifting there is "direct," meaning water that physically passes through the building's cooling systems.

The larger story sits in the indirect column. The UNU report on the environmental cost of AI's energy usage estimates that global data center electricity consumption required just under a trillion gallons of water in 2025, the bulk of it consumed at the power plants generating that electricity. AI workloads account for roughly 20% of the total, or about 200 billion gallons. That is on the order of 300,000 Olympic swimming pools. The projection has AI's share climbing to 40% by 2030, pushing the AI-attributable figure to 600 billion gallons and giving AI a global electricity appetite larger than the entire country of Nigeria.

These numbers originated in a Guardian report that cited Xylem, a water technology vendor, which is reason enough to cross-check rather than take them at face value. The independent figures do not line up exactly, but they converge on the same trajectory: the volumes are large, and they are accelerating.

Context matters

A footnote worth keeping in view: AI is not close to the top of any water-consumption ranking. U.S. agriculture used roughly 26.4 trillion gallons in 2024 per USDA figures, dwarfing every digital workload combined. AI's 200 billion gallons is, however, closing in on international oil refining. OPEC produced around 86 million barrels of crude per day in 2024, and at a rough 0.4 barrels of water per barrel of crude, that lands near 550 billion gallons annually. So AI water use is not yet an industrial outlier. The concern is the slope of the curve, not the current altitude, and the fact that buildouts concentrate demand in specific, often drought-prone, locations.

Why cooling is the wrong villain

The intuitive culprit is the cooling tower, and developers have leaned into that intuition because it is the lever they can pull. Moving from evaporative cooling, which boils off water to shed heat, to closed-loop direct-to-chip systems can slash on-site water draw dramatically. Microsoft's Satya Nadella has claimed the company's newest AI facilities run cooling systems efficient enough to "operate effectively with zero water consumption," comparing their annual water use to that of a single restaurant.

For anyone who has built a PC, the shift is familiar. Evaporative cooling is the equivalent of letting hot coolant boil away. Closed-loop direct-to-chip is the data center version of an all-in-one liquid cooler that recirculates the same fluid. Push the concept further and you get immersion cooling, fanless liquid loops, and the more speculative undersea and in-orbit proposals. All of them point toward facilities that consume very little water at the rack.

Here is the catch. Closed-loop systems trade water for electricity; they draw more power than the evaporative designs they replace. And the indirect water cost of that electricity, the water evaporated at thermal power plants, swamps the direct savings. Solve cooling perfectly and you have addressed the smaller half of the problem while nudging the larger half in the wrong direction.

The power curve is the story

The reason power dominates is sitting in Nvidia's product roadmap. The accelerator power envelope has gone vertical over four generations:

• A100 (Ampere): 300 to 400W TDP
• H200 (Hopper): up to 700W
• GB200 (Blackwell): as much as 1,200W
• Vera Rubin (next gen): up to 2,300W per chip

Scale those parts into full racks and the density becomes the headline. Traditional server racks drew 10 to 15KW. The GB300 NVL72 design can pull upwards of 150KW per rack. Vera Rubin racks, more efficient per unit of compute, may still draw upwards of 230KW each. Legacy data centers were simply never engineered for this thermal and electrical density. Feeding it is the central challenge, and every watt generated by an evaporatively-cooled thermal plant carries a water tax.

This is also why the comparison to heavier industries holds up. Steel and iron, chemicals and petrochemicals, cement, and glass each consume several times more power than all data centers combined, and therefore more indirect water through generation. AI is not unique in this mechanism. It is just the fastest-growing claimant on a constrained grid.

Where the fix actually lives

If power generation is the lever, then the supply mix is the answer, and the near-term reality is ugly. Hyperscalers racing to energize sites have leaned on mobile methane jet turbines, sometimes in violation of local permitting. Those turbines are light on water but heavy on carbon, trading one externality for another.

The durable solutions are renewables, water recovery, and nuclear. The Switch Tahoe Reno exascale facility, a 650MW site built in 2017, runs on 100% renewable solar. In Portugal, the SINES Start Campus, a 1.2GW facility partially online, runs on 100% renewables and uses seawater cooling to sidestep freshwater draw entirely. As solar deployment costs keep falling, renewable-powered data centers that avoid both the carbon and the water penalty look increasingly viable rather than aspirational.

Nuclear is the other pillar. Smaller modular reactor designs promise faster, repeatable deployment, and there are proposals as exotic as repurposing decommissioned aircraft carrier and submarine reactors for near-site power. Developers clearly see this path, which is why they are scrambling to lock up uranium supply. The likely future for AI power is some blend of renewables and co-located nuclear, with the methane turbines serving as the stopgap nobody wants to defend.

What it means for the buildout

American sentiment toward data centers is souring, and water is a sharper grievance than carbon because it is local and immediate. A resident facing drought, well contamination, or rationing experiences the facility down the road very differently than a developer reading a quarterly utilization chart. For communities organizing against these projects, the leverage point may not be cooling commitments, which address the smaller share, but power-sourcing requirements and mandatory investment in local water infrastructure to offset indirect demand.

The analyst's read is straightforward. Cooling innovation is genuine and worth pursuing, but it is a distraction from the binding constraint. As long as the marginal megawatt comes from a water-intensive thermal plant, every efficiency gain at the rack is partially refunded at the power station. The facilities that resolve this, the ones pairing renewables and near-site nuclear with closed-loop cooling, will be the ones that actually get permitted and finished. The rest will keep colliding with the communities whose water they quietly depend on.