How much water does AI use? Annual totals miss the peak-day demand that forces towns to build pipes, pumps and treatment capacity.
AI’s next water bill may arrive before a new data center processes its first ChatGPT prompt. A UC Riverside study, conducted with Caltech, estimates that U.S. community water systems could need $10 billion to $58 billion in new infrastructure by 2030, depending on the pace of data center growth and assuming new efficiencies don’t reduce demand. Most of the debate asks how much water one AI query uses, then compares the answer with a teaspoon or a bottle. Cities don’t design water systems around an average query. They build for the hottest day. That peak, and who pays to serve it, is the commercial issue hiding behind the viral numbers.
The Per-Prompt Debate Hides the Pipe
The bottle comparison has a real source. In 2024, The Washington Post worked with researchers at UC Riverside to estimate the resources needed for GPT-4 to produce a 100-word email at an average U.S. data center. Their model produced a water estimate of 519 milliliters, including both cooling and the water associated with electricity generation.
The researcher behind that number has since revised it. Shaolei Ren of UC Riverside, who produced the 2024 estimate with the Post, now puts the figure closer to 15 milliliters for a GPT-4 prompt, including about five milliliters for on-site cooling. He calls the original figure outdated for today’s systems.
Sam Altman wrote that an average ChatGPT query uses about 0.000085 gallons, or roughly 0.32 milliliters. He didn’t publish enough detail with that figure to reconcile it with the Post’s estimate. Other researchers have found a wide spread across models and workloads. A 2025 benchmarking study, “How Hungry is AI?,” estimated less than two milliliters for efficient models across its test lengths, while some reasoning models exceeded 150 milliliters per query.
Those figures don’t describe the same task or draw the same accounting boundary. They change with the model, output length, hardware, cooling system, power source and location. The spread explains why a national average tells a local utility little about the capacity a specific project will require.
How Much Water Does AI Use at Peak?
Annual totals show scale while concealing the engineering constraint. A 2024 Berkeley Lab report estimated that all U.S. data centers, not AI facilities alone, directly consumed about 17.4 billion gallons of water in 2023. It projected direct consumption from hyperscale data centers could reach 16 billion to 33 billion gallons a year by 2028.
A town still can’t size a pipe from an annual national total. The UC Riverside and Caltech team found that daily demand from evaporative cooling can rise to six to 10 times the annual average during hot weather. For some planned facilities, the multiple can exceed 30. Large projects can draw more than one million gallons on a hot day, and some facilities under construction have received allocations of up to eight million gallons per day.
Without new efficiencies, the researchers estimate that U.S. water systems could need another 697 million to 1.45 billion gallons of peak capacity per day by 2030, roughly the daily supply of New York City. Much of that capacity would sit unused outside the hottest periods. Utilities would still have to finance and maintain it.
Annual averages can look manageable even when a few hot days force a major infrastructure investment. The pipes, pumps and treatment capacity built for those peaks must be financed and maintained for decades.
Cooling Moves Cost Between Water and Power
Banning water cooling would move part of the burden to the power grid. Evaporative systems remove heat efficiently but consume water. Dry cooling and air-cooled chillers cut direct water use but generally draw more electricity. Circle of Blue also notes that the regional power mix changes the indirect water footprint because some forms of electricity generation consume far more water than others.
The industry’s shift toward waterless and closed-loop cooling can sharply reduce direct water consumption. But “closed loop” doesn’t settle the question by itself. Water may circulate repeatedly inside a data center, yet the facility still needs a final method for rejecting heat. That system may use air, evaporation or both.
These choices are made well before the servers arrive. Site selection sets the available water source and climate. The power agreement influences indirect water use. Cooling design sets the trade between water consumption and electricity demand. Changing those decisions after construction is expensive, which makes procurement and development agreements more consequential than a consumer’s prompt count.
Put Peak Water Terms in the Deal
Executives and local officials should require four numbers before approving a large data center project. First, get peak daily water demand under the hottest expected conditions, not just the annual average. Then identify the water source during drought restrictions and separate the project’s direct consumption from the water associated with its electricity supply. Finally, put a number on the developer’s contractual contribution to new treatment, storage and pipeline capacity.
That last figure matters. The UC Riverside and Caltech researchers recommend that developers help fund verifiable water-system improvements so expansion costs don’t fall entirely on local ratepayers. Water utilities should also consider a large-user tariff or capacity charge that assigns more of the dedicated infrastructure cost to the customer creating the demand. Peak reporting and enforceable funding terms belong in the agreement, not in a sustainability report published after the site opens.
The familiar question, how much water does AI use, still deserves an answer. It needs the right unit. Milliliters per prompt can compare models, but they can’t tell a mayor whether a treatment plant will handle the next heat wave or tell an executive what a site will cost to serve. Start with peak daily demand, the water source and the infrastructure agreement. Those figures reveal whether a data center is a manageable customer or a decades-long liability.
