A River Runs Through It (Into the Server Room)
In the high desert of eastern Oregon, where the Columbia River cuts through basalt cliffs and semi-arid scrubland, a cluster of massive gray buildings hums around the clock. The region was chosen partly for its cheap hydroelectric power, but there is another resource these facilities depend on just as critically: water. Millions of gallons of it, drawn each day to keep processors from melting under the thermal load of constant computation.
Most people understand, vaguely, that data centers use a lot of electricity. Fewer realize they also drink from the same freshwater systems that supply farms, cities, and ecosystems. As the artificial intelligence boom accelerates, with hyperscale facilities being announced on nearly every continent, the water equation is becoming impossible to ignore. Researchers, municipal water managers, and environmental advocates are increasingly sounding alarms. The question is no longer whether data centers are water-hungry. The question is how hungry, and whether anyone is paying close enough attention.
The Physics of the Problem
To understand why data centers consume so much water, you have to start with basic thermodynamics. Every server in a data center generates heat as a byproduct of computation. Left unchecked, that heat would destroy the hardware within minutes. Cooling is therefore not optional; it is a continuous, non-negotiable operational requirement.
The most common cooling method has historically been evaporative cooling, in which water absorbs heat and is then evaporated into the atmosphere. This is highly effective and relatively inexpensive, which is why it became the industry standard. A large data center using evaporative cooling towers can consume millions of gallons of water per day, depending on its size, the local climate, and how aggressively operators have optimized their systems.
The metric researchers and engineers use to measure this is called Water Usage Effectiveness, or WUE, expressed in liters of water consumed per kilowatt-hour of IT energy used. An average facility might post a WUE of around 1.8 liters per kilowatt-hour. A highly optimized one might achieve 0.5 or below. A poorly managed or older facility could exceed 3.0. These numbers sound abstract until you start multiplying them by the global fleet of data centers, which now numbers in the thousands and collectively draws somewhere between 200 and 500 terawatt-hours of electricity per year.
Shaolei Ren, a researcher at the University of California, Riverside who has published extensively on data center water use, estimated that training GPT-3 alone consumed roughly 700,000 liters of fresh water. His work also suggested that roughly 10 to 50 medium-length ChatGPT responses could require about 500 milliliters of water, a figure that sounds trivial until you scale it across hundreds of millions of daily users.
AI Is Making Everything Worse
The generative AI surge that gathered momentum in 2023 and has continued accelerating through 2025 and 2026 has fundamentally changed the calculus. Training large language models and running inference at scale generates substantially more heat per rack than traditional cloud workloads, because the GPU clusters required are far more power-dense than conventional server arrays.
A standard data center rack might consume 10 to 15 kilowatts. An AI-optimized rack running the latest generation of GPU hardware can consume 60 to 100 kilowatts or more. This density problem means that air cooling, already strained in conventional facilities, becomes physically inadequate. The heat simply cannot be moved fast enough through air alone.
This is driving a rapid shift toward liquid cooling technologies, including direct liquid cooling (DLC), immersion cooling, and rear-door heat exchangers. These systems circulate water or a dielectric fluid directly against or around the chips, extracting heat far more efficiently than air. Microsoft, Google, and Meta have all announced or deployed liquid-cooled AI infrastructure at scale. In some configurations, direct liquid cooling can significantly reduce evaporative water loss, because the heat exchangers transfer energy to building cooling loops rather than directly evaporating water to the atmosphere.
However, there is a complication. While liquid cooling often reduces water consumption at the rack level, the sheer growth in total computational capacity means overall water withdrawal continues to rise. You can optimize the per-unit efficiency metric and still consume more total water if you build twice as many racks. That is precisely what the industry is doing.
Microsoft disclosed in its 2023 Environmental Sustainability Report that its global water consumption had increased 34 percent between 2021 and 2022, reaching 6.4 million cubic meters. Google reported in its 2023 environmental report that it had consumed approximately 5.6 billion gallons of water in 2022 across its operations, with data centers accounting for the vast majority. Meta similarly publishes water withdrawal and consumption figures in its sustainability reports. These figures represent only the direct operational water use. They do not include the water consumed upstream to generate the electricity that powers these facilities, a category known as indirect or “gray” water use.
Where the Water Comes From Matters Enormously
Not all water use carries equal consequences. A data center drawing from an abundant water source in a rainy region of the Pacific Northwest, replenishing a water table that receives ample annual precipitation, poses a fundamentally different ecological and social risk than an identical facility drawing from a drought-stressed aquifer in Arizona or a river system already over-allocated across multiple states and agricultural users.
This geographic dimension of the water problem is where local conflicts are already emerging. In 2023, a Microsoft data center development in Goodyear, Arizona drew scrutiny over its water use and the strain on local water infrastructure in one of the most water-stressed metros in the United States. In the Netherlands, Amsterdam and the neighboring Haarlemmermeer municipality paused new large data center construction in 2019, citing limited space and strain on the electricity grid. Similar concerns have been raised in Chile, Uruguay, and parts of Spain, where major facilities have been proposed or built in regions experiencing long-term precipitation decline.
Virginia’s Loudoun County, home to what is often described as the largest concentration of data center capacity in the world, has seen its local water utility grapple with the cumulative demands of the region’s extraordinary build-out. The county’s data center corridor now draws substantial volumes of water, and local planning officials have had to revise infrastructure projections multiple times as the footprint continues to grow.
The concentration risk is significant. When a single metropolitan region hosts many hundreds of megawatts of data center capacity, all drawing from the same watershed, the cumulative withdrawal can exceed what individual permitting reviews would ever flag as problematic for any single facility.
The Transparency Gap and Why It Persists
One of the most persistent frustrations for researchers and regulators is how difficult it remains to get accurate, comparable water use data from the industry. Unlike electricity consumption, which is metered precisely and reported in financial disclosures as an operational cost, water use has historically been treated as a minor operational footnote.
Large operators like Google, Microsoft, Meta, and Amazon Web Services now publish environmental sustainability reports that include water consumption figures, a significant improvement from a decade ago. But these disclosures vary in methodology, scope, and granularity. Some report only direct operational consumption. Some include cooling tower makeup water but not water used in construction or in manufacturing the hardware. Smaller colocation providers and enterprise data centers often report nothing at all.
The International Energy Agency estimated in 2025 that global data centers consumed roughly 560 billion liters of water in 2023, including indirect water use, but acknowledged that such estimates carry substantial uncertainty due to reporting gaps. In the absence of mandatory, standardized disclosure requirements, independent researchers are often left triangulating from electricity consumption data, published WUE figures from a handful of facilities, and satellite imagery.
Legislation is beginning to catch up, if slowly. The European Union’s Energy Efficiency Directive, updated in 2023, requires data centers above a certain size operating in member states to report energy and water consumption data to national authorities, which will then be compiled into a public European database. In the United States, there is no equivalent federal requirement, though some state-level initiatives have begun exploring mandatory reporting frameworks.
Alternatives, Innovations, and Hard Trade-offs
The industry is not standing still. A range of technologies and approaches are being deployed or piloted that could materially reduce water consumption, though each comes with its own set of trade-offs.
Air-side economization uses outside air to cool facilities during cooler months or in cooler climates, eliminating or drastically reducing water evaporation during those periods. Google has also pursued other approaches to cooling: its Hamina, Finland data center is cooled with seawater from the Bay of Finland, and its St. Ghislain, Belgium site uses chilled water from an industrial canal.
Closed-loop cooling systems recirculate the same water rather than continuously drawing fresh water and evaporating it, dramatically reducing consumption. These require more sophisticated heat rejection equipment, typically large dry coolers, but can reduce water use substantially compared to open evaporative systems.
Reuse of non-potable water, including reclaimed municipal wastewater, is gaining traction in some jurisdictions. Meta’s data center in DeKalb, Illinois uses cooling technology more efficient than the industry standard and reuses its water numerous times before discharging it, a model that other operators have cited as aspirational. Microsoft has committed to becoming “water positive” by 2030, meaning it aims to replenish more water than it consumes across all its global operations.
Immersion cooling, in which servers are submerged in a non-conductive dielectric fluid, eliminates the need for water-based cooling entirely for the server itself, though the thermal energy still needs to be ultimately rejected somewhere. Companies including Green Revolution Cooling, LiquidStack, and others have been deploying immersion systems at scale. For environments running extremely dense AI compute, immersion is increasingly practical rather than exotic.
However, each of these approaches involves real costs. Closed-loop systems are more expensive to build and maintain. Reclaimed water supply infrastructure requires coordination with municipal utilities that may not have the capacity or regulatory framework in place. Air-side economization is geography-dependent and less effective in hot or humid climates. And none of these solutions eliminates the fundamental tension between explosive growth in compute demand and finite water resources.
Looking Forward: Growth, Stress, and the Accountability Reckoning
The global pipeline of announced data center capacity as of mid-2026 is staggering. Industry analysts tracking construction and permitting activity have documented very large amounts of new capacity in various stages of planning or development across North America, Europe, Southeast Asia, and the Middle East. Much of this is explicitly AI-oriented infrastructure, designed from the ground up for the kind of power-dense, heat-intensive GPU clusters that modern AI training and inference demands.
If that capacity comes online over the next three to five years roughly as planned, and if water efficiency improvements do not keep pace with capacity growth, the total water footprint of the global data center industry could roughly double from where it stands today. In a world where the United Nations, the World Resources Institute, and dozens of regional water authorities are already flagging acute freshwater stress affecting billions of people, that trajectory is not sustainable without deliberate, structural intervention.
The most credible path forward likely involves several converging forces. Regulation will need to establish mandatory water disclosure and, in stressed regions, consumption limits or offset requirements. Operators will face mounting pressure from investors and insurers who are beginning to treat water risk as a material financial exposure. And municipalities that have aggressively courted data center investment for its tax revenue and employment will need to more rigorously weigh those benefits against long-term water infrastructure costs.
There is also a demand-side conversation that has barely begun. The efficiency gains in AI model architecture over the past several years have been remarkable: newer models achieve equivalent performance at a fraction of the compute cost of their predecessors. If that trajectory continues, the water intensity per unit of useful AI output could fall substantially. But those efficiency gains will only translate into reduced absolute water consumption if the industry does not simply redeploy the savings into ever-larger models and more expansive services, which is precisely what has happened with electricity efficiency gains for the past two decades.
The data centers powering the AI era are, in a very real sense, infrastructure. They are as fundamental to the digital economy as roads, electrical grids, and water systems themselves. The uncomfortable irony is that in building and running them, we are placing growing pressure on the very water systems that sustain everything else. Getting that balance right is not a technical problem with a clean engineering solution. It is a governance problem, a political problem, and ultimately a question of what we collectively decide is worth the cost. The water bill, one way or another, is coming due.