Data Centers & Utilities

The environmental debate over data centers is often framed through water. Headlines emphasize gallons consumed, local residents worry about strained aquifers and municipal systems, and technology companies respond with efficiency metrics and sustainability language. The controversy is understandable. Data centers are physically large, politically opaque, and increasingly associated with artificial intelligence (AI), a sector whose infrastructure demands are expanding faster than most communities can evaluate them. Water also carries unusual public force. It is visible, local, and morally intuitive in a way that electricity rarely is.

Yet the public discussion often suffers from poor accounting. The term "water use" can refer to withdrawal, consumption, potable municipal supply, reclaimed water, evaporation through cooling towers, closed-loop coolant inventory, or indirect water used in electricity generation. These categories describe different environmental problems. Treating them as interchangeable produces a distorted view of data center impacts. The resulting confusion allows both sides to speak past each other: critics can imply that every gallon reported by a company represents a permanent local loss, while operators can cite aggregate efficiency metrics that obscure site-specific stress.

A more rigorous analysis begins by separating three questions. First, what water is being used, where, and under what conditions? Second, how does the facility remove heat from increasingly dense computing hardware? Third, how will the data center secure the continuous electricity required by modern AI systems? The third question increasingly dominates the long-term infrastructure problem. Water should remain a serious permitting concern, especially in drought-stressed regions, but power is becoming the strategic bottleneck. Large AI data centers now resemble industrial loads more than ordinary commercial buildings. Their future depends on generation capacity, transmission, interconnection, regulatory reform, and the availability of firm low-carbon energy.

This essay argues that the data center controversy should be understood as a problem of infrastructure accounting and public capacity. Water use deserves scrutiny, but it should be measured locally and technically rather than through sensational annual totals. Cooling systems are improving, but no thermal design can avoid the basic requirement of rejecting heat. The growth of AI shifts the center of gravity toward power procurement, grid politics, and firm generation. Nuclear energy, including small modular reactors, offers one plausible answer to the demand for continuous low-carbon electricity. Its promise is substantial, although present timelines, licensing realities, and construction risks make it a medium-term infrastructure strategy rather than an immediate remedy.

1. H2O

The confusion begins with the word "use." In water accounting, withdrawal and consumption describe different processes. Withdrawal refers to water taken from a source. Consumption refers to water that does not return to the same source in usable form, frequently because it evaporates. A facility may withdraw a large volume while returning most of it. Another may withdraw less while consuming more. A third may rely on reclaimed wastewater, reducing pressure on drinking-water supplies even if its gross volume appears large.¹

Data centers interact with water in several ways. Some water is used directly on-site for cooling and humidification. Some is consumed indirectly through electricity generation. Some facilities draw potable municipal water. Others use reclaimed wastewater, rainwater, industrial water, or other non-potable sources. Some water evaporates through cooling towers. Some circulates inside closed technology loops as water-glycol coolant and remains in the system for years, with losses occurring mainly through maintenance, leakage, permeation, or replacement. These different uses carry different environmental and political meanings.

The popular mental model is often too simple. A data center is imagined as a digital factory that draws clean water, contaminates it, and discharges waste. Certain industrial processes operate closer to that model, and some data center projects may still impose unacceptable local burdens. Modern data center cooling, however, generally relies on recirculating systems, evaporative systems, air-cooled chillers, dry coolers, economizers, liquid cooling loops, or hybrid designs. The relevant question is not whether a facility has a large annual water figure. The relevant questions are more specific: what type of water is being drawn, what share is potable, what volume is consumed through evaporation, what happens during peak heat, how stressed the watershed is, and how transparent the operator is about site-level demand.

Large-scale comparisons can help reveal proportionality, although they must be handled carefully. Amazon has disclosed a 2025 data center water efficiency figure of 0.12 liters per kilowatt-hour, and recent reporting placed its global data center water use at approximately 2.5 billion gallons.² Older estimates cited by the United States Golf Association suggest that U.S. golf course irrigation used about 2.08 billion gallons per day.³ The comparison is rhetorically striking: one of the world’s largest cloud operators can appear, in crude volumetric terms, comparable to roughly a day of American golf course irrigation.

That comparison supplies context rather than absolution. National totals can distort local reality. A gallon consumed in a water-secure region has a different political meaning from a gallon consumed from a stressed municipal system during drought. A facility using reclaimed wastewater in a humid region has a different impact from a facility drawing potable water in an arid region during a heat wave. Averages across a global portfolio cannot answer whether a specific town, aquifer, or utility system can absorb a new facility.

This is where accusations of whataboutism become relevant. Comparing data centers to golf courses, agriculture, or thermoelectric generation can be analytically useful if the comparison clarifies scale and opportunity cost. The comparison becomes evasive when it functions as a moral escape hatch. The appropriate conclusion is narrower and stronger: water controversies should be evaluated at the level of watershed, source, timing, and consumption mechanism. The annual gallon total is a starting point, not a verdict.

Corporate sustainability language can further obscure the issue. Operators increasingly discuss being "water positive," funding replenishment projects, or improving water usage effectiveness. These efforts may be beneficial. They may also fail to address the same local constraint created by a facility. Restoring wetlands or funding conservation within a broad region does not necessarily guarantee that a municipal water system can handle peak withdrawals during drought. Public concern is therefore not reducible to ignorance. It is often a rational response to incomplete disclosure.

A serious permitting process should require site-specific reporting. Communities should know annual water use, peak-day water demand, potable and non-potable shares, expected consumption rather than mere withdrawal, drought operating plans, wastewater or reclaimed-water arrangements, and the relationship between the facility and the local watershed. Without those details, both alarm and reassurance remain speculative.

2. Thermodynamics

A data center converts electricity into computation and heat. Computation is the useful output; heat is the unavoidable physical byproduct. Every cooling architecture answers the same question: how should heat be moved away from chips, racks, rooms, and finally the facility itself?

The most important technical distinction is between chip-side cooling and facility-side heat rejection. Chip-side cooling describes how heat leaves the computing hardware. Traditional servers are air cooled. Fans move air across heat sinks, and the facility manages the warm air through containment, airflow design, and mechanical cooling. High-density AI systems increasingly use direct-to-chip liquid cooling, where coolant passes through cold plates attached to processors or accelerators. Immersion systems place hardware in dielectric fluid, although these remain comparatively niche because they complicate equipment design, maintenance practices, materials compatibility, warranties, and operations.

Facility-side heat rejection describes how the building transfers collected heat to the outside environment. The facility may use cooling towers, air-cooled chillers, dry coolers, evaporative systems, economizers, hybrid wet-dry systems, or heat reuse. Direct-to-chip liquid cooling changes the first stage of heat removal, but the facility must still reject the heat. If that rejection occurs through evaporative cooling towers, the facility consumes water through evaporation. If it uses dry cooling or air-cooled chillers, direct water consumption may fall, while electricity demand may rise, especially in hot weather.

This tradeoff is central. Evaporation is effective at carrying away heat, which can reduce cooling energy. It also consumes water. Dry cooling conserves water, but often requires more fan power or mechanical cooling. Liquid cooling improves the thermal pathway from chip to facility infrastructure, enabling denser racks and potentially improving energy efficiency. It does not eliminate thermodynamic obligation. Heat removed from a processor still enters a loop, a heat exchanger, a chiller, a dry cooler, a tower, or a reused-heat system.

The phrase "water-cooled data center" therefore provides too little information. A closed technology loop may contain a water-glycol mixture that functions as inventory, remaining in the system for years with limited losses. The facility loop may nevertheless use a cooling tower and consume water continuously through evaporation. Conversely, an air-cooled server room may reject heat through an air-cooled chiller and consume little on-site water. Environmental assessment requires the full architecture: chip-side method, facility-side rejection, water source, climate, electricity intensity, and operating profile.

AI accelerators make the cooling question more urgent. Advanced GPU systems concentrate high power draw into dense racks, making conventional air cooling less attractive at the upper end of deployment. Direct liquid cooling is becoming a practical component of the AI infrastructure stack. Uptime Institute’s 2024 cooling survey found that 22 percent of respondents reported some use of direct liquid cooling, though adoption was often confined to a small share of racks.⁴ This suggests early but uneven diffusion: the technology is no longer exotic, while broad fleet-wide conversion remains incomplete.

Two-phase direct-to-chip systems may become more important as rack densities increase. These systems exploit phase change, moving heat through evaporation and condensation within a controlled loop. They can improve heat transfer and may reduce some water dependence on the technology side. They also introduce engineering and operational burdens, including refrigerant selection, containment, leak management, service procedures, and environmental risk if working fluids escape. Single-phase immersion and two-phase immersion offer additional possibilities, but their adoption remains constrained by hardware compatibility, serviceability, fluid management, and operational conservatism.

Efficiency metrics help discipline design choices. Power Usage Effectiveness, or PUE, compares total facility energy to IT equipment energy. A lower PUE indicates less overhead for cooling, power conversion, lighting, and support systems, with 1.0 representing the theoretical ideal.⁵ Water Usage Effectiveness, or WUE, is typically reported in liters per kilowatt-hour of IT energy, and lower values indicate greater water efficiency.⁶ WUE should not be interpreted as a metric that improves as it approaches 1.0. A facility with a WUE of 0.2 L/kWh is more water efficient than one at 1.0 L/kWh, other conditions equal.

PUE and WUE remain incomplete on their own. PUE does not reveal the carbon intensity of electricity. WUE does not reveal whether the water is potable, reclaimed, locally scarce, or consumed during peak drought conditions. A facility may have excellent aggregate metrics and still impose unacceptable burdens at a specific site. The next generation of data center reporting should therefore combine efficiency metrics with local resource accounting: water source, peak-day demand, grid mix, emissions profile, waste-heat strategy, noise impact, and infrastructure cost allocation.

Waste heat itself deserves more attention. Rejecting 100 megawatts or more of thermal energy into the surrounding environment is a meaningful local impact, even if it is dispersed through air rather than discharged into water. Recycling heat can be technically feasible, especially for district heating, greenhouses, industrial processes, or nearby buildings, but implementation depends on temperature levels, distance to useful loads, seasonal demand, financing, and local planning. A credible sustainability policy should ask whether waste heat can be productively used before treating atmospheric rejection as the default end state.

3. Power Demand

Water controversy attracts attention because it is tangible. Power demand will likely determine the long-term trajectory of the sector. The International Energy Agency projects that global data center electricity consumption could more than double to around 945 terawatt-hours by 2030, slightly more than Japan’s current total electricity consumption.⁷ That projection reframes the cloud as a major electricity-planning problem.

AI is central to this shift. Training and serving large models require dense clusters of specialized accelerators operating under high utilization. The resulting facilities can demand hundreds of megawatts. Multiple campuses within a region can become a grid-planning event, affecting transmission queues, substation construction, utility resource plans, local rates, and generation mix. Northern Virginia, Ireland, parts of Texas, Oregon, and other data center hubs already illustrate the pressure created when concentrated computing demand meets constrained grids.

Power and water policy are linked through cooling design. Dry cooling may reduce direct water use while increasing electricity demand. Evaporative cooling may reduce electricity demand while consuming water. Hybrid systems can shift between modes depending on temperature, humidity, water availability, and power prices. A design that appears environmentally superior in one region may be inferior in another. Climate, watershed stress, grid carbon intensity, and community priorities determine the relevant tradeoff.

The political economy of electricity is therefore unavoidable. If a data center relies on natural gas generation, AI expansion becomes a carbon and air-quality problem. If it relies on intermittent renewables without adequate firming, storage, or transmission, reliability remains unresolved. If it absorbs existing grid capacity, other customers may face higher costs or delayed interconnection. If it depends on diesel backup beyond emergency conditions, local emissions become significant. If it constructs private power with little public benefit, communities may see infrastructure scarcity privatized by the firms most able to pay.

Federal cooling programs recognize part of this challenge. ARPA-E’s COOLERCHIPS program targets total cooling energy expenditure below 5 percent of IT load for high-density compute at any time and in any U.S. location.⁸ That target is primarily an energy target, although it can also reduce water pressure depending on the cooling design. The program’s premise is revealing: at AI densities, cooling efficiency becomes a national energy problem rather than a narrow facilities-management issue.

The phrase "data center" may now understate the scale of new AI campuses. Many facilities resemble industrial loads that happen to produce computation rather than aluminum, steel, or chemicals. They require continuous electricity, substantial capital expenditure, dedicated utility planning, and political permission. Treating them as ordinary commercial buildings creates distorted permitting and infrastructure debates.

A more disciplined policy framework would require data centers to identify the origin and additionality of their power. "Additionality" matters because a facility that claims clean energy certificates while drawing from an already constrained grid may have weaker public value than a facility that finances new firm generation or pays for grid upgrades. Communities should distinguish between facilities that bring new capacity and those that merely purchase priority access to existing capacity.

This distinction also affects the claim that data centers could become net positive for local power systems. In principle, a large data center could finance new generation, support transmission upgrades, provide grid services, offer demand response for flexible workloads, or oversize generation so neighboring communities benefit. These outcomes require contractual and regulatory design. A data center is a constant load by default. Public benefit emerges when permitting, interconnection, utility regulation, and local agreements require durable infrastructure contributions.

4. Conditions for Public Benefit

Nuclear energy has returned to the data center debate because it addresses the sector’s hardest power requirement: continuous low-carbon electricity at large scale. Nuclear plants provide dense, firm generation with high capacity factors. These characteristics match the load profile of AI facilities more closely than intermittent generation alone. Renewables, storage, geothermal, transmission expansion, demand flexibility, and grid-enhancing technologies all matter. Nuclear has distinctive value where the central requirement is large quantities of around-the-clock clean power.

Major technology companies are already moving in this direction. Microsoft and Constellation announced a 20-year power purchase agreement tied to restarting Three Mile Island Unit 1, now named the Crane Clean Energy Center, with approximately 835 megawatts of carbon-free energy expected to return to the grid.⁹ Google and Kairos Power announced an agreement for up to 500 megawatts of advanced nuclear generation, with first deployment targeted by 2030 and additional deployments through 2035.¹⁰ These arrangements demonstrate that hyperscale operators increasingly view nuclear power as a strategic component of compute infrastructure.

Small modular reactors (SMR) are especially attractive in theory. Their proposed advantages include modular deployment, smaller unit size, high reliability, dense power output, and potential proximity to load. NuScale has received U.S. Nuclear Regulatory Commission approval for its uprated 77 megawatt-electric SMR design.¹¹ Rolls-Royce SMR has secured a contract with Great British Energy Nuclear connected to delivery of the United Kingdom’s first SMR fleet at Wylfa in North Wales.¹² These developments show meaningful institutional momentum.

The practical limitations remain substantial. A small modular reactor remains a nuclear power plant. Siting, licensing, financing, nuclear-grade manufacturing, fuel supply, security, emergency planning, waste management, workforce capacity, and public legitimacy all remain binding constraints. Smaller scale may improve modularity and deployment logic, but it does not remove the institutional burden of nuclear energy.

Regulatory reform is necessary. Nuclear licensing in the United States should become faster, more predictable, and more repeatable for standardized designs. A society pursuing electrification, AI leadership, industrial reshoring, decarbonization, and grid reliability cannot treat firm clean power as procedurally exceptional at every repetition. Safety regulation remains essential, but repeatable reactor designs need a regulatory pathway that rewards learning, standardization, and serial construction.

Even with reform, near-term nuclear deployment for data centers will likely rely on existing assets. Uprates, restarts, power purchase agreements, and life extensions offer more immediate credibility than privately operated SMRs behind every AI campus. Colocated SMRs may become viable in the 2030s, especially if first deployments demonstrate cost control and regulatory predictability. Presently, the vision is plausible rather than imminent.

The public-policy question is how to convert data center growth into infrastructure benefit. A data center can function as an extractive load, absorbing grid capacity, consuming water, producing noise, rejecting heat, and receiving tax incentives while creating relatively few permanent jobs. It can also function as an anchor customer for new clean firm generation, paid transmission upgrades, reclaimed-water systems, heat reuse, and local tax capacity. The distinction depends on regulation and bargaining power.

A serious permitting bargain would include several requirements. Potable water use should be restricted in stressed watersheds. Reclaimed water should be preferred where feasible. Peak-day water reporting should be mandatory. Developers should disclose power demand, power source, backup generation, emissions profile, cooling architecture, heat-rejection strategy, and noise impacts. Facilities that require grid upgrades should pay for them. Projects that bring genuinely new clean firm capacity should receive favorable treatment relative to projects that only consume existing capacity. Heat reuse should be evaluated where plausible. Community benefit should be defined in infrastructure terms rather than limited to construction jobs and tax abatements.

Such a framework supports technology without surrendering public authority. Data centers are part of the material economy. They occupy land, require electricity, reject heat, draw water, and depend on public permission. The central policy task is to make their resource claims legible and conditional. Water accounting must be local. Cooling claims must distinguish chip-side systems from facility-side heat rejection. Power procurement must distinguish existing grid consumption from new firm capacity. Nuclear policy must distinguish serious deployment pathways from speculative branding.

The future of data centers should be governed by a simple principle: critical digital infrastructure should carry critical infrastructure obligations. Facilities that consume civilization-scale resources should provide civilization-scale accountability. With strong disclosure, water restrictions in stressed regions, rigorous power planning, and accelerated firm clean generation, data centers can contribute to a more capable energy system. Without those conditions, private compute growth will continue to outrun public infrastructure, producing conflict over water, power, land, heat, noise, and trust.

Footnotes