How Water and Wastewater Capacity Now Decide AI Data Center Sites. 10 Min Read. Evaporative cooling can cut peak power draw but increases direct water use, shifting pressure from the grid to municipal supplies.Getty. This is the latest installment of The Breaking Points, a series that examines the critical challenges for AI infrastructure expansion.. A proposed Virginia data center campus requested up to two million gallons per day (MGD) of water capacity for its initial deployment, with future demand potential reaching as much as 8 MGD. The utility-services agreement acknowledged that the project’s projected demand exceeded existing water and wastewater planning assumptions. Buried deeper in the filing was another requirement: “continuous evaporative cooling to protect sensitive equipment required for essential operations.”. For the past two years, the AI infrastructure race has centered on electrical systems. Utilities rewrote load forecasts, grid operators struggled under waves of interconnection requests, and hyperscalers locked up substations, gas capacity, and transmission access to support increasingly power-hungry AI clusters. Now the same pressures are spilling into another system built for slower, steadier growth: municipal water infrastructure.. Related:Data Center Growth Draining Global Water Supplies. In multiple regions, water is becoming a siting constraint akin to grid capacity, dictating where large AI campuses can be built or expanded. As developers push toward larger campuses packed with dense GPU deployments, utilities and municipalities are confronting questions around cooling architectures, access to reclaimed water, wastewater treatment capacity, and long-term drought planning. In Newton County, Ga., a water authority representative responding to a proposed 6 MGD request from a data center project said bluntly: “We just don’t have the water.” The remark captured a broader reality: securing electricity does not guarantee sufficient cooling water, wastewater capacity, or municipal support.. In Texas, the draft 2027 State Water Plan projects existing statewide water supplies could decline roughly 10% by 2080 even as population rises more than 50%. The state estimates roughly $174 billion in water infrastructure projects may be required over the next 50 years to meet growing AI demand and maintain water supply. The plan frames water planning around drought-of-record conditions – precisely the conditions that intensify cooling-water stress from large AI clusters – yet it does not explicitly model AI-related data center demand as its own planning category.. The Physics Beneath the AI Boom. Water systems were not designed for dense AI clusters operating at massive thermal scale. The industry talks endlessly about compute; the underlying physics revolve around heat rejection.. Related:The Breaking Points: Power Emerges as AI’s Defining Limit. “Nearly all the server energy is converted into heat, which must then be removed from the data center server room to avoid overheating,” researchers from UC Riverside wrote in a recent paper examining AI’s water footprint.. For years, most data centers relied primarily on air cooling. AI changed the equation. Modern GPU systems pack far more thermal density into each rack than traditional enterprise hardware. Operators increasingly deploy direct-to-chip (D2C) liquid cooling, rear-door heat exchangers, chilled-water loops, and large cooling distribution systems to move heat efficiently. Pumps, cooling towers, treatment systems, chillers, reclaimed-water systems, and municipal water infrastructure now sit directly in the thermal path.. In some counties, multi‑MGD requests now exceed available capacity for peak conditions. (Image: Getty). Energy-Water Trade-offs, Made Concrete. Cooling choices force operators to balance electricity demand against water consumption. Evaporative systems can reduce electricity demand while increasing direct water consumption. Dry or adiabatic cooling reduces water dependency but can raise energy overhead during extreme heat events.. “Cooling towers are the worst water-wise; dry or adiabatic cooling is the best,” Vaibhav Bahadur, an associate professor at the University of Texas at Austin who studies thermal systems, liquid cooling, and water use in AI infrastructure, told Data Center Knowledge. “This has to be contrasted with energy use as well.”. Related:The Breaking Points: Cooling Struggles to Keep Pace With AI Power Density. In recent public commentary, Shaolei Ren, an associate professor at UC Riverside who studies AI infrastructure and water systems and who was a co-author of the recent paper on AI’s water footprint, argued evaporative-assist cooling can reduce peak cooling power demand by 20% to 60% during extreme summer conditions, effectively shifting part of the infrastructure burden from the electric grid onto municipal water systems. He also warned that data centers carry “two distinct water dependencies:” direct cooling demand and the large off-site water burden tied to electricity generation.. AI Density Changes the Cooling Equation. A 2025 review paper from researchers affiliated with Lawrence Berkeley National Laboratory found workload-level water use can vary by more than 10,000-fold depending on cooling architecture, electrical-grid water intensity, server utilization, climate zone, and efficiency.. The UC Riverside paper notes D2C liquid cooling systems themselves “do not evaporate or consume water.” Much of the industry now emphasizes closed-loop liquid cooling systems, where coolant circulates through sealed pipes and cold plates rather than evaporating inside the server environment. But closed-loop systems do not eliminate the broader cooling burden; the facility still must reject that heat to the outside environment.. “Getting the heat from the data hall is simply concentrating it and moving it somewhere else,” said Justin Blumling, chief marketing officer at EkkoSense, in an interview with Data Center Knowledge. “Liquid cooling changes how heat is transported, but the facility still must reject it to the outside environment.”. Blumling added that monitoring granularity becomes more important as rack densities climb and GPU clusters operate closer to sustained maximum utilization. “The blast radius of a problem with AI racks, often running near maximum levels, could be greater,” he said.. Indirect vs. Direct Water Use. The UC Riverside paper notes direct-to-chip liquid cooling systems themselves “do not evaporate or consume water.”. But large AI clusters still require facility-scale heat rejection, which can shift water demand elsewhere through cooling towers, evaporative-assist systems, or electricity generation depending on facility design and power sources.. That distinction may become more important as developers pursue behind-the-meter natural gas generation to bypass grid interconnection bottlenecks. While on-site generation systems can reduce dependence on constrained transmission infrastructure, it may also increase regional water consumption depending on generation technology and cooling design.. “While direct water use will go down with adoption of better technologies and optimization, indirect water use is likely to be a big issue,” Bahadur said.. Performance and Thermal Headroom. Even as water consumption draws greater scrutiny, operators continue pushing toward liquid cooling because dense AI workloads increasingly overwhelm traditional air-cooling systems.. A recent benchmarking study comparing liquid-cooled and air-cooled 8× Nvidia H100 systems found the liquid-cooled configuration maintained GPU temperatures between 41–50°C (106-122°F) under peak load, compared with 54–72°C (129-162°F) in the air-cooled system. The lower temperatures translated into roughly 17% higher throughput during sustained stress testing. During production AI workloads, the liquid-cooled system also consumed roughly 1–1.5 kW less node-level power under high utilization while maintaining equal or better training performance. Those gains are driving the industry toward liquid cooling despite the engineering complexity it introduces.. Blumling noted that newer D2C systems capable of operating at warmer water temperatures are also reshaping the geography of AI infrastructure deployment. “We’ve seen major data center announcements over the last year in Mississippi, Alabama, and Louisiana,” he said. “Prior to the AI boom, these locales weren’t hotbeds of data center construction; the climates are, of course, very hot/humid.”. He added that some newer cooling technologies can operate at significantly warmer temperatures than legacy chilled-water systems, making large AI deployments financially viable in regions previously viewed as difficult cooling environments.. The Problem with Peaks. Researchers from UC Riverside , Caltech, and Rochester Institute of Technology argue the industry is colliding with a less visible constraint: peak water capacity. Their paper, “Small Bottle, Big Pipe,” warns that many public water systems lack the surplus capacity necessary to absorb large AI-related cooling loads during extreme summer conditions.. The researchers estimate US data centers could require between 697 million and 1.45 billion gallons per day of new water capacity through 2030 if current water-use intensity persists — roughly comparable to New York City’s average daily water supply. The strain would fall heavily on host communities, many of which already operate aging or capacity-constrained public water systems.. Those tensions are surfacing in local permitting fights as utilities confront peak-demand scenarios many municipal systems were never designed to absorb.. Ren told Data Center Knowledge that utilities often receive maximum-demand requests during project review, but many smaller systems may not have that capacity ready today during extreme summer conditions. Water becomes a siting constraint “when the local system cannot reliably support the project’s peak demand without affecting other users,” he said.. The real problem is volatility. Unlike most municipal users, data centers can exhibit unusually high peaking factors – the ratio between average daily water demand and maximum daily water demand. The paper estimates many data centers operate with peaking factors between 3 and 10 depending on cooling architecture and climate; some facilities exceed 30. One Wisconsin AI-focused hyperscale facility analyzed in the paper requested roughly 0.7 MGD of water capacity while averaging only about 23,000 gallons per day of actual use, implying a peaking factor above 30.. Municipal systems do not size around annual averages; they size around the hottest days of the year.. The authors describe the tradeoff explicitly: “There is a fundamental tradeoff between power and water use for facility-level cooling.” Water-efficient cooling often increases the electricity burden. Electricity-efficient cooling often increases water burden.. Reclaimed water reduces pressure on potable supplies but faces its own hydraulic limits.. Wastewater Joins the AI Stack. Public debate around data centers usually focuses on freshwater consumption, but municipal systems also face discharge burdens. Large cooling systems generate blowdown water containing concentrated minerals and chemical treatment compounds that local treatment plants must process. As AI campuses scale, wastewater systems increasingly become part of the compute stack, whether cities planned for that role or not.. In water-constrained regions, operators now pursue reclaimed-water agreements to reduce dependence on potable systems.. In Northern Virginia, water infrastructure is already expanding around data center cooling demand.. Loudoun Water is doubling treatment capacity at its Broad Run Water Reclamation Facility from 15 MGD to 30 MGD while expanding a reclaimed-water distribution network built partly to support industrial cooling loads tied to Data Center Alley. The system includes roughly 20 miles of reclaimed-water pipeline and delivered more than 745 million gallons of reclaimed water to customers in 2025.. In Douglas County, Ga., Google uses treated municipal wastewater to cool facilities before returning remaining flows to the Chattahoochee River. Even reclaimed-water systems carry limits; Loudoun Water notes hydraulic constraints and capacity allocations can restrict new reclaimed-water connections as regional demand climbs.. The shift is quietly transforming wastewater infrastructure into a strategic component of the AI buildout. Power planners, water districts, wastewater operators, and economic development officials historically operated separately. AI infrastructure is forcing those systems together.. Water-system expansion now mirrors grid expansion: expensive, slow, politically contested, and constrained by geography. Reservoir, new sourcing, treatment capacity, and wastewater upgrades can take years or decades to complete. Those timelines increasingly collide with hyperscalers racing to deploy new AI capacity.. “There is no ‘national reservoir’ that data centers can tap into,” the researchers behind “Small Bottle, Big Pipe” write.. Infrastructure Built for Another Era. Data center infrastructure debates have largely centered on electricity; AI changed the scale of the problem. Large GPU clusters concentrate enormous thermal loads into relatively small footprints, forcing operators to move much more heat through facility cooling systems and through municipal water and wastewater networks originally planned around slower industrial growth.. Utilities now must plan around peak cooling demand, wastewater-treatment capacity, drought resilience, and the trade-offs between water-intensive and power-intensive cooling architectures. At the same time, many public planning frameworks still struggle to model AI infrastructure as a distinct infrastructure category.. The Virginia utility agreement tied uninterrupted AI operations to “continuous evaporative cooling to protect sensitive equipment required for essential operations.” The same filing acknowledged the project’s projected demand already exceeded the utility’s existing long-range planning assumptions.. The pressures described in “Small Bottle, Big Pipe” suggest that water is a hard infrastructure constraint that will determine where new AI clusters can be approved – and where they cannot.. Blumling said the industry’s long-term cooling trajectory is becoming clear as rack densities climb. “If 500 kW racks are the reality, then liquid cooling is a reality mandated by physics,” he said.. All of this points to a new reality for AI infrastructure: in many markets, the next major constraint may not be generating enough electricity, but securing enough water — and enough wastewater capacity — to carry the heat away.. About the Author
