What a Coolant Distribution Unit Does at the Loop Boundary

Open Compute Project’s Project Deschutes specification rates a single coolant distribution unit at 2 MW of thermal capacity, a 3°C approach temperature difference, and 500 GPM at 80 to 90 psi. A Korean national R&D program targeting completion in 2029 lists a CDU rated at 100 kW for the same broad market. Both devices are called CDUs. The gap between those two numbers is not a range of options on a catalog page. It describes two different jobs.

A CDU exists because two coolant loops are kept separate on purpose

The facility water system (FWS) and the technology cooling system (TCS) are built as two closed loops, and the Open Compute Project treats keeping them apart as the coolant distribution unit’s core function, not a side benefit. OCP’s cold plate requirements document defines the CDU as the component that “facilitates heat transfer and removal between dedicated liquid loops,” with the TCS loop running from the CDU through the rack manifold, the quick disconnects, and the cold plate, then back to the CDU. The FWS side, tied to the building’s chiller or cooling tower, sits outside that document entirely. Source: Open Compute Project, “ACS Liquid Cooling Cold Plate Requirements Document,” Revision 1.0

ASHRAE’s TC9.9 committee goes further, stating in a technical alert that isolating the TCS from the FWS “is essential for effective liquid cooling implementation,” and recommending that operators connect only a single IT platform to each CDU rather than sharing one loop across mismatched equipment. Source: ASHRAE TC9.9, “The CDU’s Critical Role of TCS and FWS Isolation in Cold Plate Deployments,” as cited by JetCool (2025)

A 2021 TC9.9 reference architecture, reproduced in Lenovo’s Neptune planning documentation, lays out the same split: the facility loop and the technology loop run as separate circuits so that each side can carry a different liquid and meet a different water-quality class. Source: Lenovo Press, “Lenovo Neptune Direct Water-Cooling Standards Planning / Implementation” (2024), referencing ASHRAE TC9.9 (2021)

Isolating the loops trades a chemistry problem for a hydraulic and cost one. Running facility water straight to a cold plate would tie coolant quality to whatever the local water utility delivers; the industry has judged that a two-loop system with a dedicated heat exchanger, more pumps, and more joints that can leak is worth the added hardware.

The CDU’s job is heat exchange, flow control, and filtration, not chilling

A CDU does not produce cold water. It moves heat from the technology loop into the facility loop, holds flow and pressure inside a range the cold plates can tolerate, and keeps the technology-side coolant clean enough not to foul microchannels. OCP’s Project Deschutes v1.0 specification, released in February 2026, states the target design point as 2 MW of thermal capacity at a 3°C approach temperature difference (ATD), with 500 GPM of flow available at 80 to 90 psi. Source: Open Compute Project, Project Deschutes CDU Specification v1.0 (Feb. 2026), OCP Academy course announcement

Boyd’s showcase of a Deschutes-compliant unit at the 2025 OCP Summit added that the design includes 0.2-micron side-stream filtration and redundant pump circuits, both aimed at keeping the technology loop clean and running through a pump failure rather than at raising the headline capacity number. Source: Boyd, BusinessWire press release, “Boyd to Showcase Google’s Recently Released Two-Megawatt Project Deschutes Coolant Distribution Unit” (Oct. 2025)

The 3°C approach temperature does more design work here than the 2 MW figure. It sets how closely the technology loop’s supply temperature can trail the facility loop’s, and that gap is what decides how much rack density a given chiller plant can support before someone has to build a colder facility loop to keep up.

Checking the arithmetic behind a published CDU rating

A CDU’s heat-transfer number is not a marketing figure; it comes directly from q = ṁ · Cp · ΔT, and a published measurement should reproduce it. A 2024 peer-reviewed study tested two in-row liquid-to-air CDUs cooling three racks at a combined 128 kW, and reported CDU 1’s actual heat transfer at full power as 58.87 kW, with a supply fluid temperature of 39.15°C, a return fluid temperature of 53.8°C, and a coolant flow rate of 59.95 liters per minute of PG25 (a 25 percent propylene glycol and water mix). Source: Heydari et al., “Experimental evaluation of direct-to-chip cold plate liquid cooling for high-heat-density data centers,” Applied Thermal Engineering 239 (2024)

Running that through the equation requires two properties the paper does not publish for its own internal calculation: a density and specific heat for PG25. Using typical values from standard glycol property tables for a 25 percent propylene glycol solution near the loop’s operating temperature, density ≈ 1020 kg/m³ and specific heat ≈ 3.8 kJ/kg·K, the flow rate converts to a mass flow of roughly 1.02 kg/s, and the temperature rise across the rack is 53.8 − 39.15 = 14.65 K. Substituting: q = 1.02 kg/s × 3.8 kJ/kg·K × 14.65 K ≈ 56.8 kW. That is within about 4 percent of the paper’s own reported 58.87 kW, a gap consistent with the approximate glycol properties used here rather than the paper’s own measured fluid data. This is a single measured lab data point at one operating condition, not a nameplate rating.

The arithmetic does more to define what a CDU is than an effectiveness column does. The unit is the accounting boundary where a mass flow rate and a coolant’s specific heat get converted into a heat-transfer number; nothing inside the box invents capacity that the equation does not already predict.

A 2 MW CDU and a 100 kW CDU describe different scopes, not different sizes of one product

Table 1. CDU capacity ratings under their stated conditionskW / MW

* Capacity figures are not directly comparable across rows; each depends on its own stated approach temperature, flow condition, and scope

Source: OCP Project Deschutes v1.0 (2026); khARN reporting on the KEIT/Ministry of Climate, Energy and Environment program (2026); Heydari et al., Applied Thermal Engineering (2024); Schneider Electric Blog (2026)

A Korean national R&D program run under the Ministry of Climate, Energy and Environment and the Korea Institute of Energy Technology Evaluation and Planning (한국에너지기술평가원), titled 「데이터센터 전력소비 절감을 위한 직접액체냉각시스템 개발」 (Development of a Direct Liquid Cooling System to Reduce Data Center Power Consumption), splits its CDU targets by construction context rather than by a single capacity ladder. The new-build track pairs two-phase cold plates with a 2 MW in-row CDU, targets 300 kW or more per rack, and aims for a PUE of 1.0x. The retrofit track pairs single-phase cold plates with a 100 kW in-rack CDU and a rear-door heat exchanger, targets 100 kW per rack in existing air-cooled facilities, and aims for a PUE of 1.1x. Source: 칸(khARN), “신성엔지니어링, 기후에너지환경부·한국에너지기술평가원 국책과제 수행” (Shinsung Engineering to carry out Ministry of Climate, Energy and Environment / KEIT national R&D project), Aug. 2, 2026, reporting on the consortium’s program announcement

That split tracks a broader industry pattern. Schneider Electric’s own guidance places rack-mounted CDUs in the 100 to 300 kW range and reserves centralized, floor-mounted units for loads above 300 kW, without naming a single fixed threshold that every vendor follows. Source: Schneider Electric Blog, “In-rack vs floor-mounted CDU for AI data centers explained” (Apr. 2026)

The 2 MW and 100 kW figures in the Korean program are not two points on the same capacity curve. One is an aggregate rating for a unit serving a bank of racks running two-phase cold plates in a new building; the other is a per-rack rating for a single-phase retrofit into a chassis that was never designed to be liquid-cooled. Reading them as a small CDU and a big CDU misses that the cold plate technology, the construction context, and the PUE target all changed along with the label. That a 20-times difference in capacity maps to a retrofit-versus-new-build split rather than a product tier is not something a spec sheet alone conveys; it comes only from reading the program document itself.

What to check before comparing two CDU spec sheets

A capacity number on its own answers nothing. Before setting two CDU ratings side by side, four things need to match or be adjusted for:

• The stated approach temperature or ΔT the capacity number assumes, since OCP’s 2 MW figure is tied to a 3°C ATD and a different ATD produces a different kW number from the same hardware
• Whether the rating covers one rack, a row, or a room, since “2 MW” for an in-row unit and “100 kW” for an in-rack unit are not measuring the same span of IT equipment
• Whether the cold plate side is single-phase or two-phase, since the Korean program pairs each CDU class with a different cold plate technology rather than treating capacity as an independent variable
• Whether the number is a shipping nameplate rating, a lab measurement like the 58.87 kW figure above, or an R&D target for a program that does not conclude until 2029

The Korean program’s 2 MW and 100 kW figures are development targets for a project running from April 2026 to December 2029 with a budget of KRW 19.24 billion, not specifications of hardware currently on a data center floor. OCP’s Deschutes v1.0 spec is further along, with vendors including Eaton, Boyd, and STULZ building to it, but it is a February 2026 release and field history is still limited. Neither status should be mistaken for the other when a sheet lands on a desk for comparison.

CDU versus chiller, ratings and later upgrades

Is a CDU the same thing as a chiller

No. A CDU moves heat between the technology loop and the facility loop through a liquid-to-liquid or liquid-to-air heat exchanger and a set of pumps; the chiller or dry cooler that ultimately rejects that heat to outside air sits further upstream on the facility water side and is outside the CDU’s own boundary. Source: Open Compute Project, “ACS Liquid Cooling Cold Plate Requirements Document,” Revision 1.0

Does a higher kW rating always mean a CDU can cool more racks

Not on its own. OCP’s 2 MW Deschutes figure is stated at a specific 3°C approach temperature difference and a specific flow and pressure condition; changing the coolant supply temperature or the flow rate changes the achievable heat transfer even on identical hardware. Source: Open Compute Project, Project Deschutes CDU Specification v1.0 (Feb. 2026)

Can a data center swap an in-rack CDU for an in-row CDU later as density grows

The sources reviewed here do not document a standard retrofit path between the two, and the two Korean program tracks pair each CDU type with a different cold plate technology rather than treating the CDU as an interchangeable module. Moving from one architecture to the other would touch facility-side piping, not just the rack. This is a system design question that needs engineering review rather than a general answer.

References

Open Compute Project, “ACS Liquid Cooling Cold Plate Requirements Document,” Revision 1.0, https://www.opencompute.org/documents/ocp-acs-liquid-cooling-cold-plate-requirements-pdf Open Compute Project, Project Deschutes CDU Specification v1.0 announcement (Feb. 2026), https://x.com/OpenComputePrj/status/2072321488242257963 Boyd / BusinessWire, “Boyd to Showcase Google’s Recently Released Two-Megawatt Project Deschutes Coolant Distribution Unit at the 2025 OCP Summit” (Oct. 2025), https://www.businesswire.com/news/home/20251010498843/en/Boyd-to-Showcase-Googles-Recently-Released-Two-Megawatt-Project-Deschutes-Coolant-Distribution-Unit-at-the-2025-Open-Compute-Project-OCP-Summit Eaton, “What is the Open Compute Project Project Deschutes” (2025), https://www.eaton.com/us/en-us/markets/data-centers/data-center-cooling/cdus/what-is-the-open-compute-project-ocp-project-deschutes.html ASHRAE TC9.9 Technical Alert, “The CDU’s Critical Role of TCS and FWS Isolation in Cold Plate Deployments,” as cited by JetCool (2025), https://jetcool.com/post/coolant-distribution-units-for-ai-choosing-the-right-fit-for-your-data-center/ Lenovo Press, “Lenovo Neptune Direct Water-Cooling Standards Planning / Implementation” (2024), referencing ASHRAE TC9.9 (2021), https://lenovopress.lenovo.com/lp2018.pdf Heydari, A., Gharaibeh, A.R., Tradat, M., Soud, Q., Manaserh, Y., Radmard, V., Eslami, B., Rodriguez, J., Sammakia, B., “Experimental evaluation of direct-to-chip cold plate liquid cooling for high-heat-density data centers,” Applied Thermal Engineering 239 (2024), 122122, https://par.nsf.gov/servlets/purl/10529800 칸(khARN), “신성엔지니어링, 기후에너지환경부·한국에너지기술평가원 국책과제 수행” (Aug. 2, 2026), https://www.kharn.kr/news/article.html?no=31504 Schneider Electric Blog, “In-rack vs floor-mounted CDU for AI data centers explained” (Apr. 2026), https://blog.se.com/datacenter/2026/04/29/in-rack-cdu-floor-mounted-cdu-option-better-data-center/