AI Data Centers Are Draining America's Water Supply. The Bill Is Coming Due.
AI data centers consumed 17 billion gallons of water in 2023. By 2028, that number hits 68 billion. The infrastructure cost to keep up: up to $58 billion.
The AI industry has spent two years arguing about compute costs, token prices, and model benchmarks. It has spent almost no time talking about water.
That's a problem, because water is now one of the hardest constraints on AI infrastructure growth in the United States, and the costs are already showing up in ordinary people's utility bills.
The Scale of AI's Water Problem
Here's the baseline: data centers powering AI systems consumed approximately 17 billion gallons of water in 2023. Projections put that figure at 68 billion gallons by 2028. That's a 300% increase in five years, and it's not an outlier estimate from an environmental advocacy group. It's the number that engineers and municipal water planners are actually working with.
The mechanism is straightforward. Data centers use evaporative cooling systems to manage the heat generated by millions of servers. Those systems are the most energy-efficient cooling option available in most climates. They are also extremely water-intensive. For every kilowatt-hour of electricity a data center consumes, roughly two liters of water evaporate. As AI inference workloads scale, and inference, not training, is now the dominant energy consumer in production AI systems, the water demand scales with them.
In February 2026 alone, three major technology companies announced they had secured multi-million gallons of water per day for new projects in Virginia, Louisiana, and Indiana. The combined water infrastructure cost for those announcements alone approached $1 billion.
The Infrastructure Gap Nobody Planned For
A UC Riverside research team, working with Caltech, put specific numbers to the infrastructure shortfall. Without meaningful water efficiency improvements, data center cooling systems could require between 697 million and 1.45 billion additional gallons of peak water capacity per day within four years. For context, that range sits roughly equal to New York City's entire daily water supply.
Even the optimistic scenario, one that assumes significant efficiency gains actually materialize, would still require new water infrastructure that could supply half of New York City for most of the year.
The total infrastructure bill to upgrade community waterworks across the United States, depending on the rate of data center construction, sits between $10 billion and $58 billion. That range is wide because the rate of data center growth is genuinely uncertain. But both ends of it are enormous, and neither end of it is being paid for by the companies driving the demand.
The burden falls on municipal water systems that weren't built for this. The research team found that many community waterworks can't deliver the large bursts of water that data centers need on peak summer days, the exact days when both AI demand and cooling requirements are highest.
Who's Paying for This Already
The abstraction breaks down fast when you look at what's happening in regions with dense data center clusters.
Northern Virginia's "Data Center Alley" is the clearest example. It's one of the most concentrated data center markets in the world, and residents there have watched electricity prices jump dramatically over the past five years. One Manassas, Virginia homeowner saw his electricity bill spike from roughly $100 to $281 in a single month in January 2026. He'd lived in the same house for nearly 40 years. Areas with high concentrations of data centers have seen electricity prices rise 267% over five years.
Water is next. The same economic logic applies. When billions more gallons of water are demanded annually from systems designed for residential and commercial use, prices follow. Businesses that rely on water, manufacturing, food processing, agriculture, compete with data centers for the same municipal supply. They don't win that competition.
This connects to a pattern worth watching: AI infrastructure costs that were once invisible to the public are becoming very visible. The story of enterprise AI bills exploding even as token prices fall is partly a story about where those costs land. Some land on enterprise buyers. Some land on ratepayers.
The Efficiency Gap
The technology to reduce water consumption exists. Dry cooling systems use air instead of water. More efficient chip architectures generate less heat per computation. Better thermal management software can reduce cooling loads during off-peak hours.
None of this is moving fast enough. The current dynamic is that companies are deploying AI infrastructure at a pace set by competitive pressure, not by the pace at which cooling technology can improve. When Anthropic's annualized revenue surges to $65 billion and hyperscalers are committing to build out massive new capacity, the incentive is to build fast, not to build efficiently.
The result is a gap between what the industry could do on water efficiency and what it's actually doing. Dry cooling, for instance, is more expensive to operate in most climates. So companies default to evaporative cooling and externalize the water cost.
There's a direct parallel to what happened with electricity. Data centers were consuming enormous amounts of power for years before grid operators started flagging the systemic risk. Water is on the same trajectory, but with shorter timelines, water systems have less slack than power grids in most affected regions.
This is also worth thinking about in the context of where new data center construction is being approved. States and municipalities competing for the economic development that data centers bring are approving projects without fully accounting for the water infrastructure costs those projects will eventually generate. New York's freeze on new data center construction was driven partly by energy concerns, but water was in the background of that conversation too.
What Actually Needs to Happen
The path forward requires three things happening simultaneously, and none of them are fully underway.
First, the companies building data centers need to internalize more of the water infrastructure cost. Right now, they secure water rights and pay connection fees, but the cost of upgrading the municipal systems to actually deliver peak water capacity lands on local governments and ratepayers. That's a subsidy the AI industry doesn't talk about.
Second, water efficiency standards for data centers need to become regulatory requirements, not voluntary commitments. Some companies publish water usage effectiveness metrics. Most don't. Without mandated disclosure, there's no baseline to improve against and no accountability for falling short.
Third, municipal water planners need better data about where data center construction is heading. The February 2026 announcements from three major companies, totaling nearly $1 billion in water infrastructure costs, came faster than most water systems could respond. Better advance notice requirements would at least give planners time to act.
The AI industry is scaling faster than the physical infrastructure supporting it can follow. That's been true for power. It's now true for water. The difference with water is that the shortage dynamics are more localized and less fungible. You can route electricity around a constraint. You can't pipe water from a region with surplus to a dry data center corridor without massive capital expenditure.
For professionals building or evaluating AI systems, this is worth factoring into infrastructure decisions. Where a model runs is increasingly not just a latency question. It's a resource question that will affect cost and availability as water constraints tighten. The question of what AI infrastructure actually costs keeps getting more complicated, and water just moved further up that list.
