Direct Liquid Cooling Cold Plate Path to Facility Water

A cold plate mounted on a chip package removes heat through roughly six hardware stages before that heat leaves the building: thermal interface material, cold plate, quick disconnect, rack manifold, coolant distribution unit, and facility water system. Cold plates capture 60 to 90 percent of a rack’s total heat load, depending on which source is cited; the rest still crosses the server as air. A flow rate rule of thumb published across coolant distribution unit vendor literature, 1.2 to 1.5 liters per minute per kilowatt, is not an arbitrary convention: running it through the sensible heat equation reproduces the 10 to 12 degree Celsius coolant rise that vendors also state directly.

The cold plate removes heat through a stack of resistances, not a single contact point

The distance from silicon junction to coolant is not one thermal interface. It is a chain of them, and each stage adds resistance the one before it cannot recover.

The Open Compute Project’s OAI System Liquid Cooling Guidelines describe junction-to-coolant thermal resistance as the sum of several elements, including the thermal interface material, the cold plate itself, and contact resistance between the heat source and the cold plate, which the guidelines note can significantly affect core temperature at high power. The same document states that cooling limits often trace back to package design and quality control as much as to the cooling solution, and that cold plate structure, whether straight fin, pin fin, or offset fin, is chosen against the coolant flow rate and pressure drop the loop can supply. Source: OCP, “OAI System Liquid Cooling Guidelines” (rev. March 2023)

OCP’s separate cold plate qualification white paper requires thermal resistance to be measured across multiple coolant flow rates rather than at one point, producing a performance curve rather than a single number, and requires hydrostatic pressure testing to a defined leak standard. Source: OCP, “Cold Plate Development and Qualification” white paper

A published research design for a jet impingement cold plate reports a thermal resistance of 0.019 degrees Celsius per watt at a flow rate of 2 liters per minute, dissipating a thermal design power above 1,000 watts in testing. That figure describes a lab prototype, not a shipping product. Source: ScienceDirect, “Thermal management of data centers: chip-scale cooling using novel distributed inlet-outlet jet impingement liquid cold plate” (2025)

A cold plate thermal resistance number without a paired flow rate is not comparable to another cold plate’s number. Resistance is a curve, not a constant, and a spec sheet that reports one figure without the flow rate it was measured at has dropped the variable that determines whether it applies to your loop.

Quick disconnects and the rack manifold move coolant, not heat

Once coolant leaves the cold plate, the job changes from heat transfer to containment. A quick disconnect and a rack manifold exist to move fluid without spilling it, not to cool it further.

OCP’s Universal Quick Disconnect (UQD) and Blind-Mate (UQDB) specifications define a hand-mate or blind-mate, drip-free, hot-pluggable fluid connector for use in the technology cooling system. The blind-mate specification requires the socket to withstand 5,000 make-and-break cycles while still meeting stated fluid loss limits at end of service life, and blind-mate couplings built to the spec add a self-alignment feature with roughly 1 millimeter of radial misalignment tolerance. Source: OCP, “Universal Quick Disconnect Blind-Mate (UQDB) Specification,” Revision 1.0

At rack scale, one published direct-to-chip design for a 500 kilowatt rack routes coolant through a rack manifold built from 75 millimeter square stainless steel tube, carrying 600 liters per minute with a fluid connector spaced every 48 millimeters and a manifold pressure drop of roughly 1.7 kilopascals. Source: Chilldyne, “High-Power Liquid Cooling Design: Direct-to-Chip Solution Requirements for 500 kW Racks” (2025)

Every connector in this chain is also a place fluid can escape under pressure. Counting stages here is counting leak points as much as it is counting hardware: a design with several coolant connections per node is not equivalent to a design with two, even when both hit the same flow rate.

The CDU is a boundary between two loops, not a source of cooling

The coolant distribution unit does not generate cooling. It separates two loops and lets one absorb what the other cannot carry on its own.

A CDU sits between the technology cooling system loop that touches the IT equipment and the facility water system loop connected to chillers, dry coolers, or cooling towers, exchanging heat between the two through a plate heat exchanger so their water chemistries never mix. Source: Introl, “Cooling Distribution Units” (2026)

Vendor literature places liquid-to-liquid CDU capacity in a roughly 50 to 200-plus kilowatt per rack band, while liquid-to-air units that skip the facility water connection entirely cover a lower 5 to 20 kilowatt in-rack range, extending to roughly 240 kilowatts for in-row liquid-to-air designs. Source: rdp.in, “CDUs Explained: Sizing Coolant Distribution for AI Racks”

CIBSE’s continuing education module on liquid cooling describes the CDU as the hydraulic decoupling point ASHRAE intends between the facility water system and the technology cooling system, which is also why the two loops can run to different water quality classes without cross-contaminating each other. Source: CIBSE Journal, Module 254, “Liquid cooling in data centre applications” (2025)

A CDU is a heat exchanger sized against a temperature difference, not against raw kilowatts on a label. A kilowatt rating printed on a spec sheet is only valid at the approach temperature it was tested against; changing the facility supply temperature moves the real capacity even though the nameplate number does not change.

Facility water temperature class decides how much of the path can skip a chiller

The single number that determines whether a facility needs a chiller at all is the supply water temperature class the IT equipment is rated to accept.

ASHRAE TC9.9’s fifth edition of its thermal guidelines renamed liquid cooling classes around their upper supply temperature limit: W17, W27, W32, a new W40, W45, and an open-ended W plus above 45 degrees Celsius, with every class sharing the same 2 degree Celsius lower limit. Source: ASHRAE TC9.9, “Emergence and Expansion of Liquid Cooling in Mainstream Data Centers” (2021)

The fifth edition also tightened what compliance with a class means: full, unthrottled operation is now required across the entire temperature band of the class, not merely somewhere inside it, closing a gap in how manufacturers previously self-classified equipment. Source: Upsite Technologies, “What You Need to Know About ASHRAE’s Fifth Edition of Thermal Guidelines” (2021)

The same ASHRAE paper models a dense, shadowed 1U configuration and finds that the W27 class supports up to 500 watts per CPU and up to 12 watts per DIMM with single-sided memory cooling, while at W45 the cooling capability for that configuration falls to something comparable to a conventional air cooled system running a low cold aisle temperature. Source: ASHRAE TC9.9, “Emergence and Expansion of Liquid Cooling in Mainstream Data Centers” (2021), Figure 6

A vendor advertising hardware as W45 capable is describing an entering water ceiling, not a guarantee that the chip’s rated power holds at that ceiling. The same ASHRAE paper that defines W45 also shows cooling capability falling as the class number climbs, for the same silicon.

Reported cold plate heat capture rates disagree because they describe different boundaries

Table 1. Reported cold plate heat capture share of total rack heat% of total heat, by source

* Ranges reflect different measurement boundaries (cold plate alone vs. whole rack); not directly comparable across rows

Source: Introl (2026); DCX Liquid Cooling Systems (2026); ScienceDirect (2026); Obelinf (2026); 2CRSi (2026); Chatsworth Products (2026) — see References

Published capture rate figures for direct liquid cooling range from 60 percent to a claimed 100 percent, and the spread is a difference in what was measured, not a factual disagreement between sources.

Sources describing the cold plate on its own put the share of rack heat it removes at 60 to 90 percent, with the remainder attributed to memory, storage, power delivery, and networking components that stay air cooled. Source: 2CRSi, “What Is Direct Liquid Cooling System?”; Chatsworth Products, “5 Misunderstood Facts About Direct-to-Chip Liquid Cooling” (2026)

A peer-reviewed study of a hybrid two-phase direct-to-chip and air cooled server found the liquid loop’s share of total heat rising toward 90 percent as coolant inlet temperature increased from 25 to 45 degrees Celsius, but that figure covers CPU heat only; memory, voltage regulators, the platform controller hub, and storage stayed air cooled throughout the test on a single instrumented rig. Source: ScienceDirect, “Performance evaluation of two-phase direct-to-chip liquid cooling combined with air cooling for data centers” (2026)

A separate source describing a hybrid rack architecture states 100 percent heat capture at the rack level. That figure is not a cold plate number at all: it counts a rear-door heat exchanger that catches the air-side heat the cold plate never touches, added on top of the cold plate’s own share. Source: Introl, “Direct-to-Chip Cooling Implementation” (2026)

The 60 to 90 percent figures describe what the cold plate alone removes. The 100 percent figures describe what the rack removes once a second liquid device, sitting behind the cold plate in the airflow path, cleans up what the cold plate left behind. Reading these as competing claims about the same thing is comparing a component spec to a system spec.

The vendor flow rate rule of thumb reproduces a predictable coolant temperature rise

Table 2. Direct liquid cooling path, stage by stagechip to facility

* Specs shown are representative examples from cited sources, not a universal standard

Source: OCP; Chilldyne (2025); rdp.in; ASHRAE TC9.9 (2021) — see References

Running the sensible heat equation on the flow rate ratio that recurs across coolant distribution unit literature reproduces the same coolant temperature rise vendors also state directly, which means the ratio is a checkable derived number rather than a rounded convention.

The governing equation is Q = ṁ × cp × ΔT, where Q is the heat load in watts, ṁ is the coolant mass flow rate in kilograms per second, cp is the coolant’s specific heat in joules per kilogram-kelvin, and ΔT is the coolant temperature rise across the loop in kelvin. The calculation below approximates the coolant as pure water, cp ≈ 4,186 J/(kg·K), density ≈ 1.0 kg/L near room temperature; a treated water and glycol mixture, cited elsewhere as a 75/25 blend, runs 10 to 15 percent lower in specific heat, which the figures below do not account for. Source: Tark Thermal Solutions, “Advancements in CDU Pump Technology to Meet the Demands of 75-200 kW Data Center Racks”

A vendor example sizes a 60 kilowatt rack at 1.5 liters per minute per kilowatt, for a total flow of 90 liters per minute. Converting to mass flow at 1 kg/L gives ṁ = 1.5 kg/s. Substituting: ΔT = 60,000 W ÷ (1.5 kg/s × 4,186 J/(kg·K)) ≈ 9.6°C. Source: JetCool via Flex, “JetCool SmartSense CDU: Scalable Cooling for the AI Era” (2025)

A separate published design for a 500 kilowatt rack specifies 600 liters per minute of total flow. That gives ṁ = 10 kg/s, so ΔT = 500,000 W ÷ (10 kg/s × 4,186 J/(kg·K)) ≈ 12.0°C. Source: Chilldyne, “High-Power Liquid Cooling Design: Direct-to-Chip Solution Requirements for 500 kW Racks” (2025)

Both cases land in a 9.6 to 12 degree Celsius band, which is a typical value implied by the flow ratio in each case, not a rise either source measured directly. It sits within, but does not reach the top of, a separately stated 10 to 20 degree Celsius coolant rise for current direct-to-chip hardware; the 20 degree Celsius upper end would require a lower flow ratio or a design point built around a larger ΔT than either example here assumes. Source: Obelinf, “Direct to Chip vs Rear Door vs Immersion Cooling” (2026)

The 1.2 to 1.5 liter per minute per kilowatt figure that recurs across CDU vendor literature is not a rounded convention. It is what a roughly 10 degree Celsius coolant rise looks like once converted into a flow-per-kilowatt ratio, and a design that quotes a flow ratio outside that band is implicitly quoting a different target ΔT, whether or not the vendor states one.

Gaps in the source data

The sources above do not say whether the cited heat capture percentages were measured under sustained full load or against a nameplate or theoretical maximum, and only the ScienceDirect study specifies a test population, which is a single instrumented server rig rather than a fleet. The coolant composition behind most CDU capacity and flow rate figures is not stated; the water and glycol assumption used in the calculation above is drawn from one pump vendor’s typical mixture, not from the CDU or rack examples the flow numbers came from. Whether the 1.2 to 1.5 L/min per kW rule of thumb represents a design point, a worst-case margin, or a measured operating condition is not specified in the sources that cite it. Rack power figures in the 600 kilowatt to 1 megawatt range mentioned in passing in CDU market coverage describe pre-release roadmap hardware and should be read as projections, not shipping specifications. None of the sources here address the safety or code review required before a facility water system, quick disconnect selection, or CDU sizing decision is finalized; that review belongs to a mechanical engineer of record, not to a published spec sheet.

Matching DLC figures to the boundary they were measured at

• Check which boundary a quoted heat capture or capture rate figure was measured at, cold plate only versus rack total including any rear-door assist, before comparing it against a different vendor’s number.
• Check what coolant composition a flow rate or ΔT specification assumes; a water and glycol blend requires a higher flow rate than pure water for the same kilowatt load.
• Confirm a CDU’s kilowatt rating against the approach temperature it was tested at, not just the flow rate, since real capacity moves with supply and return temperature.
• Confirm the facility water system’s temperature class against the IT equipment’s rated class for full, unthrottled operation across the entire band, not only at the class’s upper limit.
• Ask whether a cold plate thermal resistance figure is paired with the flow rate it was measured at; an unpaired resistance number cannot be compared across designs.
• Treat rack power figures above current shipping generations as roadmap projections until a vendor publishes a shipping specification.

References

OCP, “OAI System Liquid Cooling Guidelines,” Open Compute Project, revised March 3, 2023, https://www.opencompute.org/documents/oai-system-liquid-cooling-guidelines-in-ocp-template-mar-3-2023-update-pdf OCP, “Cold Plate Development and Qualification,” Open Compute Project white paper, https://www.opencompute.org/documents/ocp-cold-plate-development-and-qualification-with-integrated-comments-pdf ScienceDirect, “Thermal management of data centers: Chip-scale cooling using novel distributed inlet-outlet jet impingement liquid cold plate,” 2025, https://www.sciencedirect.com/science/article/abs/pii/S1359431125009524 OCP, “Universal Quick Disconnect Blind-Mate (UQDB) Specification,” Revision 1.0, Open Compute Project, https://www.opencompute.org/documents/uqdb-spec-1-0-pdf Chilldyne, “High-Power Liquid Cooling Design: Direct-to-Chip Solution Requirements for 500 kW Racks,” 2025, https://chilldyne.com/high-power-liquid-cooling-design-direct-to-chip-solution-requirements-for-500-kw-racks/ Introl, “Cooling Distribution Units,” Introl Blog, 2026, https://introl.com/blog/cooling-distribution-units-cdu-liquid-cooling-ai-data-center-2025 rdp.in, “CDUs Explained: Sizing Coolant Distribution for AI Racks,” https://rdp.in/gpu-mart/knowledge-base/cdu-coolant-distribution-unit-sizing-ai-racks/ CIBSE Journal, “Module 254: Liquid cooling in data centre applications,” September 2025, https://www.cibsejournal.com/cpd/modules/2025-09-lcdca/ ASHRAE TC9.9, “Emergence and Expansion of Liquid Cooling in Mainstream Data Centers,” 2021 (hosted copy), https://www.vertiv.com/49db1c/globalassets/documents/white-papers/ashrae_tc0909_emergence_and_expansion_of_liquid_cooling_in_mainstream_data_centers_5_may_2021_332771_0.pdf Upsite Technologies, “What You Need to Know About ASHRAE’s Fifth Edition of Thermal Guidelines,” 2021, https://www.upsite.com/blog/what-you-need-to-know-about-ashraes-fifth-edition-of-thermal-guidelines/ 2CRSi, “What Is Direct Liquid Cooling System?,” https://2crsi.com/direct-liquid-cooling Chatsworth Products, “5 Misunderstood Facts About Direct-to-Chip Liquid Cooling,” 2026, https://www.chatsworth.com/en-us/resources/blogs/2026/5-misunderstood-facts-about-direct-to-chip-liquid-cooling/ ScienceDirect, “Performance evaluation of two-phase direct-to-chip liquid cooling combined with air cooling for data centers,” 2026, https://www.sciencedirect.com/science/article/pii/S2214157X26000122 DCX Liquid Cooling Systems, “Direct-to-Chip Cooling,” 2026, https://dcx.eu/direct-to-chip-cooling/ Introl, “Direct-to-Chip Cooling Implementation,” Introl Blog, 2026, https://introl.com/blog/direct-to-chip-cooling-pue-below-12-implementation Tark Thermal Solutions, “Advancements in CDU Pump Technology to Meet the Demands of 75-200 kW Data Center Racks,” https://tark-solutions.com/CDU_Pump_Technology_to_Meet_Demands_of_75-200kW_Data_Center_Racks JetCool via Flex, “JetCool SmartSense CDU: Scalable Cooling for the AI Era,” 2025, https://flex.com/resources/jetcool-smartsense-cdu-scalable-cooling-for-the-ai-era Obelinf, “Direct to Chip vs Rear Door vs Immersion Cooling,” 2026, https://obelinf.com/blog/direct-to-chip-vs-rear-door-vs-immersion-cooling/ Attom Technology, “Direct-to-Chip Liquid Cooling: Working Principle, Architecture and Engineering Guide,” 2026, https://attom.tech/2026/08/17/direct-to-chip-liquid-cooling/