By Adam Kotrba, Director of Flat Products for Copper Development Association, and Erin Smith, Deputy Director, EHS and Recycling

SUMMARY

Data centers must remove more heat from less space without treating water and electricity as separate problems. Direct-to-chip, rear-door, and immersion systems recirculate coolant in closed loops; a facility’s water footprint is determined by how the heat is ultimately rejected. The best designs match cooling architecture to rack density, climate, watershed, power supply, and heat-reuse opportunities. Copper supports that balance at every handoff, thermally from the processor to the cold plate to the CDU and ultimately to the facility heat exchangers, and electrically by carrying current with low losses in compact, durable, recyclable components.

KEY TAKEAWAYS

1. Liquid cooling is not water consumption. With a CDU containing a heat exchanger in between, the IT loop and the facility heat-rejection system are separate fluids and must be evaluated separately.
2. Cutting on-site water usually raises electricity use, so water and energy must be optimized together.
3. Cooling strategy should reflect rack density, climate, watershed stress, reliability requirements, and the local power mix.
4. Copper lowers solid-side thermal resistance where heat is most concentrated—cold-plate and heat-sink bases, heat pipes, and vapor chambers—and its formability, joinability, and corrosion resistance enable compact, durable coils, tubing, and heat exchangers downstream.

Cooling has become a system-level design problem. Some of the electricity delivered to IT equipment is converted into heat, and AI accelerators concentrate far more of it within each rack. Since about 2018, processor power has climbed steeply while allowable case temperatures have trended down, a combination that ASHRAE describes as forcing the shift from air to liquid cooling [1]. The goal is not to choose water or energy, liquid or air, in isolation; it is to move heat reliably at the lowest combined resource cost.

Start With the System Boundary

Direct-to-chip cooling circulates coolant through cold plates on CPUs and GPUs to a coolant distribution unit (CDU), which transfers the heat to a facility loop, then back to the chip. Rear-door heat exchangers do the same at the rack boundary; immersion systems submerge hardware in dielectric fluid. In every case, the IT-side fluid recirculates, and the facility still needs a way to reject the heat [1].

Heat Capture and Heat Rejection Are Different Problems

A closed IT loop captures heat well and consumes essentially no water, but the heat must still go somewhere. Water is consumed at the facility stage, chiefly by evaporative cooling towers and adiabatic systems, which use the latent heat of water because it is usually the cheapest way to reject heat. Where water is constrained, facilities use “dry” heat rejection instead—dry coolers (fans only) and air-cooled chillers (fans plus compressors) that move heat to outdoor air, consuming no water but more electricity [2, 3]. “Liquid cooled,” therefore says nothing on its own about a site’s water footprint. Microsoft’s next-generation design, which closes the facility loop and rejects heat without evaporation, avoids more than 125 million liters of evaporated water per data center each year, in exchange for a PUE increase that warmer coolant temperatures and high-efficiency economizing chillers help offset [4].

How Data Center Chip Cooling Approaches Compare

Fig. 1 | Cooling approaches, the copper contribution in each, and where water is actually consumed.

Balance Water and Energy Together

At the heat-rejection stage, water and electricity are substitutes. In a study of two Phoenix-area colocation facilities, the site using air-cooled chillers had a source water usage effectiveness that was 66% lower but a PUE that was about 13% higher than the site using water-cooled chillers with evaporative cooling [3]. Google reports that its water-cooled data centers use about 10% less energy and emit roughly 10% less carbon than many air-cooled ones, and it decides on a campus-by-campus basis by weighing carbon-free energy against responsibly sourced water [5]. There is no universal best system independent of location [6].

What to Measure Across the System

1. Direct water. Withdrawals and consumption, tracked separately [7], by source, season, and operating mode, including tower makeup, humidification, and leaks. ISO/IEC 30134-9 defines water usage effectiveness (WUE) for consistent comparison [8].
2. Cooling electricity. Pumps, fans, chillers, and controls across the full load range, not only at the design point.
3. Indirect impacts. Lawrence Berkeley National Laboratory estimates that U.S. data centers consumed about 66 billion liters of water on-site in 2023 and roughly 800 billion liters indirectly through power generation; with data-center load projected to reach 6.7–12% of U.S. electricity by 2028, small efficiency gains scale across the grid [9].
4. Local context. Watershed safe-withdrawal rates, drought risk, competing demands, and climate [6], plus discharge constraints and the reliability value of each cooling mode.

Comparisons to golf courses or households convey scale [6, 7] but are no substitute for a facility-specific water balance that uses consistent geography and time periods, distinguishes withdrawals from consumption, and shows demand in the hottest, driest periods.

How Copper Supports the Cooling Path

Thermal performance depends on low resistance at every handoff: silicon to conductor, conductor to air or liquid, rack to facility loop, facility loop to final heat sink. Copper’s thermal conductivity, about 400 W/m·K compared with roughly 235 W/m·K for pure aluminum, limits the temperature drop along each solid-conduction path [10], and its formability enables thin walls, microchannels, small-diameter tubing, and compact heat exchangers. The result is a lower-resistance thermal path that supports higher density and warmer setpoints with less fan, pump, and compressor work.

Conventional Air Cooling

Air remains essential for memory, storage, power electronics, and many servers, and even here copper handles the most concentrated portion of the load: copper bases, heat pipes, and vapor chambers, pull heat off small package surfaces and spread it where airflow can remove it. At the room level, copper-tube coils in CRAC, CRAH, economizer, and chiller equipment move that heat out of the facility air and into water, glycol, or refrigerant circuits or simply outside. Copper matters twice: directly at the chip and in and out of the facility.

Direct-to-Chip Liquid Cooling

A copper cold-plate base conducts heat away from intense local hot spots while coolant flowing through internal channels carries it off. Channel designs multiply wetted area, shorten the conduction path, and strengthen convective transfer, especially with microchannels, a concept first demonstrated in silicon by Tuckerman and Pease in 1981 and now standard practice in copper [11]. Single-phase cold plates typically use a copper base, often with skived copper fins, and a brazed copper or stainless-steel cover; two-phase designs likewise put copper where heat enters and use the cover primarily for pressure containment.

Lower chip-to-fluid resistance permits higher chip power, higher rack density, and warmer coolant, which extends economizer hours and, in suitable climates, can eliminate chillers for the liquid loop, though the residual air-cooled share of the load still needs conventional cooling [1]. That advantage is under pressure because a fixed case-temperature limit must accommodate a larger chip-to-fluid rise; supply water must drop to about 30°C above roughly 500 W per chip and to about 27°C at 700 W, at which point most sites lose their free-cooling window [12]. Every degree recovered at the cold plate matters.

Rear-Door Heat Exchangers

A rear-door heat exchanger (RDHX) mounts a liquid-to-air coil on the back of the rack; fans push cool air across the coil and return warm air to the room. The copper tube wall is the bridge between the airstream and the coolant, and small-diameter tubing packs more airside surface area into the limited depth of a door. The Open Compute Project treats RDHXs as a means of scaling air-cooled facilities to higher densities and supporting hybrid air-liquid deployments [13].

In a 2025 study by OTS R&D (Optimized Thermal Systems), RDHX coils optimized around 5 mm copper tubes delivered 7% higher cooling capacity with 25–30% lower mass and 28% less air pressure drop than the baseline coil, saving up to 28% in fan power; a 1 MW data-center model showed 23% net energy savings after compressor and fan effects. Raising the inlet water temperature from 17°C to 27°C reduced compressor power by 32% and improved PUE by 7% [14]. Tube diameter, fin spacing, circuit geometry, and operating temperature should be optimized together.

Immersion Liquid Cooling

Immersion removes the air interface at the IT side, not the need for an efficient solid-to-fluid path. Copper heat sinks conduct heat from concentrated sources into the bath, and copper tubes and fittings carry it from the tank loop to the facility loop. Because every component is wetted, designers must qualify compatibility among fluid, copper alloys, finishes, seals, and dissimilar metals; for hydrocarbon fluids, copper-strip corrosion testing (ASTM D130) is the standard check. OCP’s immersion requirements list copper as an acceptable wetted material for single-phase loop components, and its design guidelines warn that fluorochemical two-phase fluids can extract contaminants from elastomers and dissolve thermal pastes, deferring in-tank material choices to the fluid manufacturer [15].

CDUs, Facility Heat Exchangers, and Power Delivery

The CDU isolates the IT loop from the building loop, allowing each to run at its own temperature, chemistry, and materials [1]. CDUs typically use copper-brazed stainless-steel plate heat exchangers, whose brazed joints provide a compact, high-pressure construction; copper tubing and fittings provide corrosion-resistant, serviceable connections; and copper motor windings reduce pump electrical losses. Together with adequately sized exchangers, this keeps the approach temperature low, preserving temperature difference for the dry cooler, chiller, tower, or heat-reuse system, and copper coils in that final equipment allow compact circuits that balance capacity against fan power, pump head, and size. Copper does not decide whether a facility consumes water; the choice among dry, evaporative, and hybrid rejection does. But efficient heat exchangers help each architecture meet its target with less auxiliary energy.

High-density cooling and high-density power now share the same rack. Once the liquid is available at the IT load, emerging designs extend it to busbars and other high-current components, where copper’s electrical conductivity reduces the heat generated and its thermal conductivity helps remove the remainder.

Design the Material and Fluid System Together

Water chemistry, coolant formulation, filtration, flow velocity, dissolved oxygen, corrosion control, and compatibility among metals, finishes, seals, and joints all affect reliability, so materials and treatments should be specified loop by loop [1, 15]. Done well, copper delivers high thermal performance, serviceable connections, and an established recycling pathway: recovered scrap supplied about 35% of U.S. copper in 2024 [16], and copper is recyclable without loss of performance.

SIX PRINCIPLES FOR SYSTEM DESIGN

1. Match the architecture to the heat load. Air where it remains effective; rear-door exchangers to extend air cooling; retrofits; direct-to-chip or immersion, where chip heat flux demands closer coupling.
2. Closed, separate loops. Isolate the IT loop from the facility loop with a CDU or heat exchanger so each can use appropriate materials and chemistry [1].
3. Run as warm as reliability allows. Higher coolant and air temperatures increase economizer hours and reduce chiller load; setpoints should follow component limits, server fan-power response, redundancy, and ambient extremes rather than a generic minimum [1, 12].
4. Tune the entire flow path. Cold-plate channels, tube diameter, fin spacing, pump head, fan power, pressure drop, and controls interact; a locally efficient component can bottleneck the system [14].
5. Plan for heat reuse before construction. Reclaimed water and heat reuse pay off only if temperature, distance, ownership, and year-round demand are settled early; reviews of district-heating projects point to technical, economic, policy, and infrastructure barriers [18]. The IEA estimates 70–80% of data-center heat is recoverable with heat pumps, with business models and tariff structures rather than technology as the remaining hurdles [17].
6. Meter across the boundary. Instrument the IT, CDU, facility, and heat-rejection loops for temperature, flow, pressure, makeup water, conductivity, pump and fan energy, and rejected-heat destination [7, 8].

Design for Place, Then Design for Change

A data center in a cool, water-rich region should not be modeled after one in a hot, water-stressed basin. Siting and design should reflect seasonal conditions, competing local demands, permitting, electricity supply, heat-reuse partners, and reliability tier, and the preferred system may change as densities rise, water tightens, or the grid gets cleaner. Hybrid designs preserve that flexibility: dry rejection most of the year with evaporative assist at peak, or direct-to-chip cooling for GPUs paired with rear-door exchangers where liquid infrastructure is limited. Only about 22% of respondents to Uptime Institute’s 2025 cooling survey use direct liquid cooling, and they still rank energy efficiency far ahead of water in sustainability decisions, leaving room to grow [19].

Innovation should improve the whole system. Finer microchannels, new joining methods, and additive manufacturing increase surface area and reduce material use, but each must be tested for pressure drop, fouling, serviceability, and lifecycle performance. The most useful innovation may be integration: thermal components, power delivery, water treatment, controls, and reuse designed as one system, because judging a technology by chip temperature, PUE, or site water alone simply pushes impacts outside the chosen boundary.

The Better Question

Data centers do not need a single answer to how to cool; they need a repeatable way to make tradeoffs. Start at the chip, follow the heat through every interface and loop, and measure the water and electricity used to move and reject it. Then apply local watershed and grid conditions, reliability requirements, and reuse opportunities. Copper lowers thermal resistance at the chip, rack, CDU, and facility boundaries and reduces electrical losses in the power path; these gains are largest when the entire path is engineered as a single system. The goal is not maximum thermal performance at any cost or minimum water use in isolation. It is reliable computing at the needed capacity with the lowest practical combined resource burden.

Sources

1. ASHRAE Technical Committee 9.9. Emergence and Expansion of Liquid Cooling in Mainstream Data Centers. White paper, 2021.

2. Orr, R., and K. Klesner. Ignore Data Center Water Consumption at Your Own Peril. Uptime Institute Journal, June 17, 2016.

3. Karimi, L., et al. Water-energy tradeoffs in data centers: A case study in hot-arid climates. Resources, Conservation and Recycling 181 (2022): 106194.

4. Solomon, S. Sustainable by design: Next-generation datacenters consume zero water for cooling. Microsoft Cloud Blog, December 9, 2024.

5. Hölzle, U. Our commitment to climate-conscious data center cooling. Google, November 21, 2022.

6. Dietrich, J. Water is local: generalities do not apply. Uptime Institute Journal, May 28, 2025.

7. Mytton, D. Data centre water consumption. npj Clean Water 4, 11 (2021).

8. ISO/IEC 30134-9:2022. Information technology — Data centres key performance indicators — Part 9: Water usage effectiveness (WUE). International Organization for Standardization, 2022.

9. Shehabi, A., et al. 2024 United States Data Center Energy Usage Report. Lawrence Berkeley National Laboratory, LBNL-2001637, December 2024.

10. Hust, J. G., and A. B. Lankford. Thermal Conductivity of Aluminum, Copper, Iron, and Tungsten for Temperatures from 1 K to the Melting Point. NBSIR 84-3007, National Bureau of Standards, 1984. Values for pure metals tabulated in The Engineering ToolBox, Thermal Conductivity of Metals and Alloys.

11. Tuckerman, D. B., and R. F. W. Pease. High-performance heat sinking for VLSI. IEEE Electron Device Letters 2, no. 5 (1981): 126–129.

12. Swinhoe, D. Hot water, cold water: What’s the right temperature for water in liquid cooled systems? Data Center Dynamics, July 31, 2024.

13. Open Compute Project. Door Heat Exchanger sub-project (Cooling Environments).

14. Nasuta, D., and S. Deshpande (OTS R&D, Optimized Thermal Systems). Optimizing Rear Door Heat Exchangers (RDHXs) for an Evolving Data Center Cooling Landscape. Webinar, October 2025. Figures as reported in Elmer, M., Next-Gen Data Centers: Advancing Air Cooling with Optimized RDHX, Copper Development Association, November 17, 2025.

15. Open Compute Project. OCP ACS Immersion Requirements, Rev. 2.10 (August 18, 2023); Design Guidelines for Immersion-Cooled IT Equipment, Rev. 1.01 (December 7, 2020).

16. U.S. Geological Survey. Mineral Commodity Summaries 2025: Copper. January 2025.

17. Voswinkel, F., C. Delmastro, B. Reidenbach, and O. Kvarnström. Opportunities for district heating in the changing energy landscape. International Energy Agency, December 8, 2025.

18. Yuan, X., et al. Data center waste heat for district heating networks: A review. Renewable and Sustainable Energy Reviews 219 (2025): 115863.

19. Weinschenk, R., and D. Donnellan. Uptime Institute Cooling Systems Survey 2025. Uptime Institute Intelligence, UII Data Report 181, July 2025.

Frequently Asked Questions