4 min read
A waterless data center cooling architecture for high-density AI, validated by Munters and ZutaCore at ASHRAE 2026
Every new data center now arrives with the same three questions from the community around it: how much water will it use, how much electricity will it draw, and what will it do to the neighborhood? Operators feel that pressure twice. They need to bring operating cost down, and they need to get projects approved and built in places that are increasingly wary of them.
Munters and ZutaCore set out to answer both at ASHRAE 2026. The result is a waterless data center cooling architecture: one two-phase loop from the chip to outdoor air, with no facility cooling water at all, running a full year of real weather at 26% of the annual cooling energy of a single-phase, chilled-water plant. Zero water usage for cooling (WUE = 0), materially lower energy, and a path to high-density AI infrastructure that is ready for the next generation of silicon.
Removing facility cooling water does not make every concern about a new data center disappear. It does remove one of the largest, and it does so before the first shovel goes in.
The Test
Munters and ZutaCore modeled a 625 kW reference design, split the way most AI data centers are split today: 80% of the heat load cooled by liquid and 20% cooled by air. The liquid side ran ZutaCore® HyperCool® cold plates on a single two-phase heat transfer fluid, connected straight to Munters SyCool heat rejection on the roof. No cooling distribution unit in the room. No facility water loop. No water anywhere in the cooling architecture.
The comparison was against single-phase, and against a strong version of it: an air-cooled chiller plant with magnetic-bearing compressors, fan-array air handlers, and liquid-to-liquid cooling distribution units sized at 1.5 liters per minute per kilowatt. That is the plant a well-funded operator would specify for single-phase direct-to-chip cooling today.
Both systems were run hour by hour against a full year of Dulles, Virginia weather data, so every warm afternoon and every cold night counted.
The Results
The whole-system result comes first. Annualized partial power usage effectiveness (pPUE), a measure of cooling and infrastructure overhead relative to IT load, came out at 1.030 for the waterless architecture against 1.115 for the single-phase chilled-water plant. Strip out the IT load and that is 0.030 of cooling overhead against 0.115: roughly 74% less cooling energy over the year.
Breaking that whole-system figure into its two parts shows where the gap opens. On the liquid side, which carries 80% of the heat load, the two-phase loop ran at a pPUE of 1.017 against 1.111 for single-phase. On the air side, the remaining 20%, the two systems are closer, 1.083 against 1.132, because both still have to move air.
Two more numbers from the same work explain why the loop is so light. At 500 kW of heat rejection, the two-phase fluid supplies the servers at 42.2°C rather than 30°C, and it does the job with about a quarter of the flow: 51 gallons per minute against 197. Pump head drops from 65–70 PSID to 15–20. Less fluid, moving with less force, at a temperature outdoor air can absorb for most of the year.
Following the original model, Munters updated the heat-rejection design. The updated unit can operate without compressors up to 43.5°C outdoor ambient and is projected to hold a pPUE of 1.055 at that condition before mechanical cooling is required.
What It Means On Site
-
No facility cooling water. WUE = 0 for the cooling architecture. No cooling distribution unit, no facility water loop, and none of the water treatment, dosing, passivation or flushing that comes with it. A two-inch copper supply line replaces four-inch stainless.
-
Lower energy. Roughly 74% less annual cooling energy than the single-phase chilled-water baseline. A quarter of the pump head, and compressors that stay off for most of the year. The energy a chilled-water plant would spend moving water and air is available for IT.
-
Future-ready deployment. Simpler infrastructure built for the next generation of high-density AI, designed to support processor TDPs beyond 4,000 W. Fewer components between the chip and heat rejection enable faster, lower-risk deployment.
Why This Matters Now
Silicon is not slowing down. Accelerators are already past 1,000 W and heading toward 4,000 W, and racks that once drew ten kilowatts now present more than a hundred. The question is no longer whether to move to liquid cooling. The question is which liquid, how much water and energy it commits you to, and how much infrastructure you want to build and run around it.
Munters and ZutaCore are defining a path to high-density AI infrastructure that avoids cooling water altogether, sharply reduces cooling energy and removes unnecessary cooling infrastructure between the chip and ambient air. The proof-of-concept build of this reference design is expected to be operational in the first quarter of 2027.
Frequently Asked Questions
What was compared?
A 625 kW reference design, 80% of the heat load cooled by liquid and 20% cooled by air, using ZutaCore HyperCool cold plates on one two-phase loop to Munters SyCool heat rejection, against the same load on a single-phase chilled-water plant with liquid-to-liquid cooling distribution units. Modeled hour by hour on a year of Dulles, Virginia weather.
What does WUE = 0 mean here?
Zero water usage for the cooling architecture: no facility cooling water, no evaporative heat rejection, and no water at the server. It does not describe other water uses at a site, such as domestic supply.
Where does the 74% come from?
Cooling overhead above the IT load: 0.030 for the waterless architecture against 0.115 for single-phase chilled water. Figures vary with configuration, heat rejection design and site.
Is this a physical test or a model?
The annual energy results are modeled, hour by hour, on measured weather data, using Munters' published equipment performance. They are not a completed physical end-to-end test. The 500 kW flow, supply temperature and pump head figures come from a Munters comparison. The integrated reference-design proof of concept is planned to be operational in Q1 2027.
Does the fluid touch the servers?
No water does. The cold plates run a dielectric two-phase heat transfer fluid, so a leak is a maintenance event rather than a damaged server.
Source: Munters and ZutaCore, ASHRAE 2026, “ZutaCore and Munters two-phase single loop: deployment and performance,” updated at Munters Cooling Innovation Day 2026. Annualized pPUE modeled hour by hour on Dulles, Virginia weather data; baseline is a best-in-class single-phase, air-cooled chiller plant with liquid-to-liquid cooling distribution units at 1.5 L/min per kW.
Want the results on one page? Download the ASHRAE 2026 results brief: Waterless Data Center Cooling. A Quarter of the Energy.
