U.S. data centers could need 697 to 1,451 million gallons of new water capacity per day by 2030 — roughly equivalent to New York City's entire daily supply. The bottleneck isn't average consumption. It's peak withdrawal. Cooling towers demand massive instantaneous flows during the hottest hours, and most municipal systems have zero surplus peak-hour capacity to spare.
The analysis (arXiv:2603.02705) reveals that efficiency gains barely dent the problem. Even with optimistic 10% annual intensity reductions, demand still reaches 227–604 million gallons per day. IT load growth — driven by AI and cloud expansion — is outpacing water efficiency improvements. The gap widens, not narrows.
Note (2026-11-09): I checked the abstract of the Han, Li, Wierman and Ren paper. The figures match: 697 to 1,451 million gallons per day of new capacity through 2030 if 2024 water use intensity persists, 227 to 604 with a 10 percent annual intensity reduction, and a valuation of about 10 billion dollars up to 58 billion. The abstract says the capacity constraint can force operators to rely on less efficient dry cooling, which stresses the power grid; the essay says data centers that cannot get water permits cannot be built. The essay's example of a data center using 2 million gallons per day and needing 300,000 in an hour is an illustration of its own, and its statement that data center demand is anti-correlated with residential demand sits oddly with its next clause that both peak in the hottest afternoon hours.
The structural insight: this is a rate problem, not a volume problem. A data center that uses 2 million gallons per day averaged over 24 hours might need 300,000 gallons in a single hour. The pipe has to be that big even if the total consumption is modest. Municipal water systems are designed for residential demand patterns — morning showers, evening cooking. Data center demand is anti-correlated: hottest afternoon hours, when residential demand also peaks.
The cost of the infrastructure shortfall is $10–58 billion in new water capacity, depending on growth scenarios. This isn't a sustainability concern. It's a binding constraint. Data centers that can't get water permits can't get built. The resource that limits AI scaling might not be chips or power. It might be plumbing.