The next constraint on AI infrastructure may not be chips or electricity alone. It may be water. Data centers have always needed a way to remove heat, but the rapid deployment of AI systems is concentrating more computing power into individual racks and accelerating construction of large facilities. That makes data center cooling a more consequential question for the places that host them—especially where water supplies are constrained, contested or expensive to expand.

This does not mean every AI facility uses vast quantities of freshwater, or that water use is inherently incompatible with data centers. Cooling designs, weather, local water systems and the electricity grid all matter. But AI data center water use is becoming a public issue because it sits at the intersection of industrial growth and an essential local resource. A project that looks efficient in a corporate sustainability report can still pose difficult questions for a particular watershed, town or utility.

The durable issue is therefore not finding one universal number for AI’s water footprint. It is deciding what information communities need, what alternatives are feasible, and which industrial uses should receive scarce water when demand rises.

Why computing becomes a cooling problem

Almost all electricity used by computing equipment ultimately turns into heat. Processors, memory, networking equipment and power-conversion systems all add to the thermal load inside a data hall. That heat has to be moved away continuously if equipment is to operate reliably.

AI changes the engineering challenge because training and serving large models can rely on densely packed accelerators running at high utilization. Conventional enterprise servers may be spread across relatively low-density racks. AI-oriented deployments can put far more power into the same physical footprint, increasing the amount of heat that must be removed from a rack and making simple room-level air cooling less practical in some installations.

Higher density does not automatically translate into higher water use. It does, however, narrow the menu of workable cooling approaches. Operators may use more sophisticated air systems, chilled-water loops, direct-to-chip liquid cooling or combinations of these methods. The environmental outcome depends on how that system rejects heat outside the building.

What AI data center water use actually means

Discussion of data center water consumption often bundles together distinct impacts. Separating them is essential.

• Direct water use is water used at or for the facility itself, commonly to support cooling. In an evaporative system, a portion of water evaporates to carry heat away. That water is generally considered consumed because it is not immediately returned to the local watershed in usable form.
• Water withdrawal is the amount taken from a source, such as a municipal system, river or groundwater supply. Some withdrawn water may be returned after use, though its temperature or quality can matter.
• Indirect water use includes water associated with generating the electricity a facility buys. Its scale depends heavily on the regional power mix and on how individual power plants are cooled. Manufacturing chips and constructing buildings also have water footprints, though these are different from a facility’s ongoing operational demand.

Those distinctions explain why competing estimates can appear contradictory. A company might report water delivered to a cooling system, while a local utility focuses on withdrawals from its network. A regional assessment may include water used by power generation. Each can be useful, but they do not answer the same question.

Annual totals can obscure another important issue: peak demand. A facility’s water needs may rise during hot, dry periods, precisely when households, farms, ecosystems and other industries face the greatest pressure. For water-stressed regions, timing and source can be as important as the yearly volume.

How data center cooling systems use—or avoid—water

There is no single data center design. A facility can combine several methods, and its operating mode may change with outdoor temperature and humidity.

Evaporative cooling and cooling towers

Cooling towers reject heat by evaporating water. They can be energy-efficient in suitable conditions because evaporation is an effective way to shed heat, but that efficiency comes with direct water consumption. Towers also need water management to limit mineral buildup and maintain water quality; some water is periodically discharged as blowdown.

This is often the system at the center of local debates because it may rely on a steady municipal or other water supply. Its actual use varies by weather, design, operating load and the number of hours it runs in evaporative mode.

Air cooling, dry cooling and hybrid systems

Air-cooled and dry-cooling systems move heat to the atmosphere without relying on evaporation in normal operation. They can substantially reduce on-site water consumption, but may require more electricity during very hot weather or use more equipment and space. Hybrid systems switch between dry and water-assisted operation, seeking to reduce water use while limiting the energy penalty during peak heat.

The trade-off matters. A design that saves local water but draws more electricity could shift part of its environmental burden upstream, depending on the grid supplying that power. This is why AI power and water demand should be evaluated together rather than as separate scorecards.

Liquid cooling for AI

Direct-to-chip liquid cooling circulates a coolant through cold plates attached to high-power processors. It can collect heat more effectively than air at the rack level, making it attractive for dense AI hardware. Immersion cooling, in which components are placed in a non-conductive liquid, is another approach used in some specialized deployments.

Neither method is automatically “waterless.” In many designs, the liquid inside the server loop is recirculated in a closed loop, but the heat still must be rejected somewhere—through a dry cooler, cooling tower, chiller or another external system. Liquid cooling can improve thermal efficiency and enable greater density, yet the final water outcome depends on the facility-level heat-rejection design.

Why the same facility can be acceptable in one place and contentious in another

Water is local. A data center using a particular amount of water may be a manageable addition to a large, resilient municipal system, especially if it uses reclaimed water for data centers rather than drinking-quality supplies. The same demand can be politically and environmentally difficult in a drought-prone basin, a fast-growing suburb with limited treatment capacity, or an agricultural area dependent on the same aquifer.

Public concern is not simply about whether a site receives enough rain. It involves the legal and physical details of supply: existing water rights, groundwater conditions, reservoir storage, treatment capacity, pipe capacity, seasonal restrictions and the needs of downstream users and ecosystems.

This is also why broad claims that data centers are either negligible or uniquely wasteful tend to mislead. Their share of a region’s total use may be small, while their demand can still be significant for a particular utility zone or during a particular season. Conversely, a large campus connected to a robust reclaimed-water network may reduce its call on potable supplies considerably.

The measurement problem: better numbers, better decisions

Data center operators increasingly publish corporate water goals and aggregate water-use metrics, often including water usage effectiveness, or WUE. Such reporting is useful, but it has limits. A company-wide average does not reveal whether a proposed site will draw drinking water during a local drought, whether it has access to non-potable supply, or how much it needs on the hottest days.

Site-level data are often difficult to obtain because operators and utilities may treat them as commercially sensitive. Yet local planning requires more specificity than global sustainability summaries can provide.

A meaningful assessment of data center environmental impact should distinguish:

• potable, reclaimed, recycled and other non-potable sources;
• withdrawal, discharge and net consumption;
• annual demand and high-temperature peak demand;
• normal operations and emergency or backup arrangements;
• on-site cooling water and electricity-related water impacts; and
• the facility’s demand relative to local supply, treatment and transmission capacity.

These are not merely technical details. They determine whether a community can evaluate a project on its actual local consequences.

AI is accelerating an old infrastructure issue

Data centers did not begin using water with generative AI. Large computing facilities have long depended on cooling systems, and cloud providers have spent years improving efficiency. What AI changes is the pace and concentration of demand. More high-density hardware, more demand for specialized facilities and shorter development timelines can force utilities and local governments to make infrastructure decisions before they have complete information about future loads.

The International Energy Agency has identified data centers as an important source of rising electricity demand in several markets, with AI among the drivers of that growth. Electricity projections do not translate directly into water projections: a facility’s cooling design and grid connection make an enormous difference. Still, the link is clear. More electricity used for computation produces more heat, and more heat requires more cooling capacity.

This helps explain why data center permitting is becoming more complicated. A project may need electricity transmission, substations, water connections, wastewater capacity, road upgrades and tax agreements. These are public systems, and the costs and risks do not automatically stay within the data center fence line.

Ways to reduce conflict—and their limits

There are practical options for lowering freshwater demand, but none is a universal fix.

• Use reclaimed or non-potable water where available. Treated wastewater can reduce competition for potable supplies. But it requires separate pipes, treatment standards, storage and reliable delivery. A reclaimed-water system cannot be assumed to exist simply because a city operates a wastewater plant.
• Choose dry or hybrid cooling. These systems can reduce direct water consumption, particularly in water-stressed regions. Their energy use, capital cost and performance in extreme heat need to be assessed locally.
• Improve chips, servers and software. More efficient hardware and models can reduce energy per task and therefore reduce heat. Efficiency gains can be offset if total AI use grows faster than efficiency improves, so absolute demand still matters.
• Reuse waste heat where conditions allow. Heat can sometimes supply district-heating networks, buildings or industrial processes. It is most feasible where there is a nearby, year-round demand and infrastructure to move low-temperature heat.
• Make location a resource decision. Siting near suitable power, resilient water systems and potential heat users can matter more than a generic corporate pledge. A project should not be treated as location-neutral simply because its digital services are global.

Permits are where the water question becomes political

Local governments and utilities decide whether a proposed facility receives water service, what infrastructure must be built, who pays for upgrades and what restrictions apply during shortages. These decisions can be buried in development agreements, utility contracts and technical permitting documents that are hard for residents to interpret.

Better governance does not require a blanket ban on data centers. It requires clear conditions: public disclosure of expected water sources and seasonal demand, enforceable drought provisions, realistic estimates of infrastructure costs, and an explanation of alternatives considered. Where public subsidies or preferential utility arrangements are involved, the case for transparency is stronger.

The central question is not whether digital services are invisible. It is whether the physical systems behind them are being planned with the same scrutiny applied to other large industrial users.

Communities also need a way to compare proposed facilities with other legitimate demands. Water allocation is not a simple contest between technological progress and environmental protection. It is a governance problem involving homes, farms, public health, ecosystems, existing employers and future growth.

AI infrastructure is becoming water governance

There is no credible single figure for AI data center water use that applies everywhere. Water intensity varies by climate, cooling architecture, server density, utilization, electricity source and the condition of the local water system. Treating every facility as identical would be as misleading as treating all AI workloads as identical.

But uncertainty is not a reason to avoid the question. It is a reason to demand more precise, site-specific answers. As AI infrastructure expands, its physical requirements will increasingly shape municipal planning and public consent. The important shift is conceptual: water is not a secondary sustainability detail attached to computing. In many places, it is part of the core decision about what kind of industrial development a community is willing and able to support.

Image by fietzfotos on Pixabay.