Data Center Cooling and Water Use: Options Compared

Data Center Cooling Options and Their Water Tradeoffs

No data-center cooling system is best everywhere. The right design depends on rack heat, climate, water conditions, reliability, energy supply, maintenance, and the final method used to reject heat outside the facility.

A claim such as "liquid cooled," "water free," or "high efficiency" describes only part of that system. To compare designs fairly, follow the heat from the chip all the way to the environment.

The decision in one sentence

Use dry or low-evaporation designs where water is the binding constraint, consider evaporative or hybrid operation where water is responsibly available and the energy benefit is material, and use direct liquid cooling when heat density makes air movement inefficient. Then test the complete heat-rejection path under the site's hottest, driest, and most stressed operating conditions.

Every design has to move the same heat

Computing equipment turns almost all of the electricity it consumes into heat. A one-megawatt information-technology load therefore creates roughly one megawatt of heat that must leave the servers and, eventually, the site.

The first cooling stage collects heat from a chip, board, server, or rack. The second transports it through air, water, refrigerant, or dielectric fluid. The last rejects it to outdoor air, water, or another useful heat sink. A design can be liquid cooled at the chip and still use an evaporative cooling tower outside.

flowchart LR
    A["Chip and memory"] --> B["Air, cold plate, or immersion fluid"]
    B --> C["Rack or coolant distribution"]
    C --> D["Facility water or refrigerant loop"]
    D --> E["Dry, evaporative, or hybrid heat rejection"]
    E --> F["Outdoor environment or useful heat recovery"]

The Department of Energy's cooling-water guidance uses this full heat path to explain why water use depends on the computing load, each transfer stage, the chiller, and the cooling tower. That boundary is more useful than a product label.

How the main options compare

These are architecture families, not guaranteed performance bands. Two facilities using the same label can have different controls, climates, loads, water quality, equipment, and outcomes.

Air cooling is familiar but density changes the equation

Conventional servers use fans to move air across components. Computer-room systems then collect the warmed air and transfer its heat to refrigerant or chilled water. The architecture is widely understood, keeps liquid away from electronics, and can be straightforward to service.

High-density AI racks make air harder to use efficiently. Moving more heat requires more airflow, more fan power, careful containment, and sometimes lower supply temperatures. Local hot spots can become the limiting condition before average room temperature looks problematic.

Air cooling does not automatically mean low water. If room air is cooled by a chilled-water plant that rejects heat through cooling towers, the site can still consume water. If it uses dry mechanical cooling, the facility may use little operational water but more electricity during hot periods.

Evaporation trades water for heat-removal efficiency

Evaporative cooling uses the physical advantage of water changing phase. That can remove heat efficiently and reduce compressor work. It also consumes water and concentrates minerals in the remaining loop, which requires treatment and blowdown.

DOE notes that tower water performance depends on cycles of concentration, controls, heat load, and operating practice. Increasing cycles can reduce makeup water and blowdown within water-quality and equipment limits. Economizers can also reduce chiller work when weather permits.

The choice is local. Evaporation can be defensible where the source is durable, the watershed can support the consumption, discharge is managed, and the energy benefit is meaningful. The same design can be a poor fit where drought, competing uses, water rights, treatment capacity, or public trust are binding constraints.

Annual gallons are not enough. A project should disclose peak-day demand, drought mode, source-water category, consumptive use, discharge, chemical treatment, forecast climate conditions, and what happens when the preferred source is interrupted.

Direct liquid cooling solves density, not every water question

Direct-to-chip systems place a cold plate on processors and move heat into a liquid loop. Water can carry much more heat per unit volume than air, and warmer coolant can sometimes eliminate mechanical chilling. Pumps can also use less energy than the fans needed for a dense air-cooled room.

DOE's Best Practices Guide distinguishes rear-door exchangers, cold plates, and immersion systems. A recent federal high-performance computing evaluation estimated lower operating costs for one direct-water-cooled system than for its air-cooled alternative.

That result should not be copied into another project without its conditions. Direct liquid cooling still needs coolant distribution, leak control, water chemistry or dielectric fluid management, heat exchangers, residual cooling for components not served by cold plates, and final heat rejection.

The strongest public claim is therefore specific: a named design avoids evaporating water for cooling during operation at the facility. It should not be rewritten as "uses no water" unless construction, electricity, manufacturing, domestic use, and the full supply chain are actually inside the measurement boundary.

Immersion changes service and material choices

Immersion cooling places electronic equipment in a nonconductive fluid. Single-phase systems circulate warm liquid through a heat exchanger. Two-phase systems allow a purpose-built fluid to boil at the equipment, condense, and return.

Immersion can remove high heat loads with less server fan energy. It also changes hardware compatibility, maintenance procedures, fluid selection, sealing, fire and environmental review, and the way equipment is repaired or replaced. The facility still needs to reject heat from the immersion loop.

For a new high-density design, those tradeoffs can be considered from the beginning. Retrofitting a conventional facility can be a different decision because racks, network equipment, floors, service tools, warranties, and operating practices already exist.

Reclaimed water changes the source, not the obligation to measure

The EPA's Quincy, Washington case documents a system that treats and reuses industrial water associated with data-center operations. The case is valuable because it shows the institutional work around reuse: treatment, infrastructure, water quality, multiple parties, and a local water problem.

Reclaimed supply can preserve potable water and create a use for treated effluent. It can also require new pipes, treatment energy, chemical control, storage, backup supply, and a long-term contract. The community still needs to know how much water is withdrawn and consumed, whether the reclaimed source has other beneficial uses, and what happens during drought or treatment failure.

Reuse should be compared with demand reduction. A project that consumes less water can be preferable to one that consumes a larger volume from a different source, depending on local conditions.

Microsoft has two different cooling stories

E076 discussed Microsoft's underwater data-center work and a later water-saving design as if they were the same project. They are not.

Project Natick was an experimental sealed data center deployed underwater off Scotland. Microsoft used it to study reliability, unattended operation, and a subsea concept that did not consume water for cooling.

Microsoft's later zero-water cooling design is a land-based architecture introduced in 2024 for AI workloads. The company says it uses chip-level, closed-loop cooling without water evaporation during operation. The claim describes Microsoft's design and stated boundary. Actual results for a particular facility still require metered operating data.

The correction matters because underwater placement is not the mechanism used by the later land-based design. A useful public explanation should name the heat path rather than borrow the most memorable image.

Climate changes the recommendation

In a cool climate, air-side economization or dry heat rejection may work for more hours. A humid climate can reduce the effectiveness of some evaporative approaches. A hot, dry region can make evaporation thermodynamically attractive while making water consumption socially and hydrologically difficult.

A high-density AI facility may need direct liquid collection regardless of region, but its outdoor system can still differ. One site may use warm-water dry coolers. Another may use a hybrid system that permits evaporation only above a defined temperature. A third may have a dependable reclaimed-water network and choose tower cooling to limit peak electricity.

The recommendation should come from a joint model of hourly weather, computing load, water availability, water quality, electricity, emissions, capital cost, operating modes, and reliability. An annual average can hide the week when both the grid and cooling system are under the most stress.

What a project should disclose

A credible cooling basis of design should name the IT load and rack-density range, initial and future phases, heat-collection technology, coolant, residual air load, facility loop, final heat-rejection modes, design weather, failover state, and maintenance assumptions.

Its resource model should report annual and peak electricity, water withdrawal, water consumption, discharge, source-water category, site water-usage effectiveness, and the treatment boundary. It should also show the effect of drought, extreme heat, a failed water source, a grid emergency, and a higher-than-expected computing load.

The project should state whether performance numbers come from design calculations, vendor tests, commissioning, or metered operation. Those evidence classes are not interchangeable.

Choose the boundary before choosing the winner

The durable question is not whether water or air is better. It is which complete heat-removal system fits the site's workload and constraints without hiding resource use outside the frame.

Start by defining the heat load, climate, watershed, supply category, energy system, peak condition, and reliability requirement. Then compare designs under the same conditions. If water is a fatal constraint, eliminate designs that cannot meet the limit before comparing their efficiency. If grid capacity is the fatal constraint, quantify the electricity and demand effect of dry operation.

Use [[How to Evaluate a Data Center Site]] to establish those constraints and [[Can AI Data Centers Be Sustainable]] to connect the cooling result with power, public cost, community terms, and closure. The full E076 story, [[Can AI Data Centers Become Better Neighbors]], explains why cooling is one part of a broader governance problem.

This comparison is a Venture Step synthesis informed by E076 and current DOE, EPA, and first-party technical records reviewed on July 27, 2026. It is not an equipment recommendation, engineering design, water-right analysis, or performance guarantee. AI assistance was used for research organization, drafting, and validation. Publication remains unauthorized.