Evaporative vs closed-loop vs air-cooled — what's the difference?

Every data center has the same basic problem: server chips run hot, and heat has to leave the building. The decision about how to remove that heat is the single biggest factor in how much water and how much electricity a facility uses. There are three main approaches — evaporative, closed-loop, and air-cooled — and they produce very different numbers on the permit application.

The common language of cooling

Two acronyms help frame the rest of this article.

• PUE — Power Usage Effectiveness. Total facility power divided by IT power. A PUE of 1.5 means for every 1 watt the servers use, another 0.5 watts is spent on cooling, lights, and other overhead.^[1]
• WUE — Water Usage Effectiveness. Liters of on-site water per kilowatt-hour of IT energy delivered. A WUE of 1.0 means the facility evaporates one liter of water for every kilowatt-hour of compute.

Industry best practice on cooling is governed by ASHRAE Technical Committee 9.9, which publishes thermal guidelines for data-processing environments.^[2] ASHRAE sets recommended server-inlet temperatures and humidity ranges. Everything else — how you get there — is an engineering choice.

Option 1: Evaporative cooling

This is the most common architecture in hot climates. The basic idea is old: water absorbs heat as it evaporates. A data-center cooling tower sprays water over warm air pulled from the server hall; some of the water turns to vapor and rises out of the tower, carrying the heat with it. The remaining, now-cooler water is pumped back through the building's chilled-water loop.

A hyperscale facility using this method can evaporate between one and five million gallons of water per day at peak load, depending on size and outside conditions.^[3] PUE tends to be low (often 1.1 to 1.3). WUE is high, typically between 1.0 and 2.0 L/kWh.

Why operators pick it: cheap to build, low power cost, works well in dry heat because evaporation is most efficient when humidity is low. Texas, Arizona, and California Central Valley all fit that profile.

Option 2: Closed-loop (dry cooling)

A closed-loop system replaces the cooling tower with large radiator-like heat exchangers that reject heat to outside air using fans. No water is lost to evaporation in the cooling process itself. The same water circulates through the building for years, topped up only to replace small leaks. WUE on a true closed-loop facility is effectively zero.

The tradeoff is power. Pushing air through a dry cooler takes more fan energy than evaporating water. And when the outside air is hot, the cooler's efficiency drops. To keep server inlet temperatures within ASHRAE limits on a 105-degree Texas afternoon, a closed-loop plant typically needs supplementary mechanical chillers — which use yet more electricity. PUE on a closed-loop facility in a hot climate often runs between 1.3 and 1.5, compared with 1.1 to 1.3 for evaporative.^[4]

Translated: a closed-loop data center uses roughly 5 to 15 percent more electricity than an equivalent evaporative one. That extra power comes from the grid, and ultimately from the same set of power plants everyone else draws from.

Option 3: Air-cooled (direct outside air)

In cool climates, a building can sometimes be cooled entirely by pulling in outside air, filtering it, and exhausting it after it has picked up server heat. This is called direct air-cooled or "free cooling." Water use is near zero; power use is the lowest of the three options. Facebook's Prineville, Oregon campus is a widely cited example.

In Texas, this approach works poorly for most of the year. Summers are too hot, humidity too variable, and dust too heavy to maintain the tightly specified inlet conditions that modern AI training hardware requires. An operator proposing true direct air-cooled for a Texas site is unusual.

A newer option: direct-to-chip liquid cooling

AI accelerators produce so much heat per square inch that traditional air cooling struggles to keep up. A newer architecture, direct-to-chip or "liquid cooling," pipes coolant through cold plates attached directly to each chip. This is very efficient inside the rack, but the heat still has to leave the building — which means the facility still needs an evaporative tower, a dry cooler, or a hybrid design to reject the captured heat.

In other words, "liquid cooled" on a press release does not tell you anything definitive about a project's outdoor water use. The next sentence of the filing is what matters: how is the heat actually rejected?

Side-by-side, the numbers that show up in filings

Closed-loop does not make water use disappear. It converts it into electricity use — which shows up somewhere else in the grid.

How to tell which one the operator is proposing

Operators do not always volunteer the architecture in public-facing materials. The disclosure usually appears in one of three places:

1. The water-permit application with the local utility, the Groundwater Conservation District, or the Texas Commission on Environmental Quality. This filing typically includes a projected peak-day draw.
2. A Public Utility Commission interconnection filing, which discloses expected electrical load. A closed-loop facility on a hot site will show roughly 10 percent higher power draw per unit of IT capacity than an evaporative one.
3. The tax-abatement agreement filed with the county or school district. Sometimes, but not always, these include a cooling-architecture clause.

If none of the three answers the question clearly, a resident can reasonably ask the operator or county staff at a public meeting: "Which cooling architecture is the project using, and what is the projected peak-day water consumption?" That single question is answerable in one sentence. A proposal that cannot answer it is not yet complete.