In-rack liquid-cooling hardware (cold plates, manifolds, QDs, rack CDUs)
BOTTLENECKCoolIT and a few fabricators hold the only qualified capacity for rack CDUs and QDs, so AI rack rollouts can outpace second-source output.
Cold plates, manifolds, quick-disconnects, and rack CDUs supplied with compute systems. Determines whether high-density AI racks can operate within thermal limits. Cold-plate fabricators capture premium on custom work; QD and CDU supply is binding capacity constraint.
Liquid-cooling hardware supplied as part of the compute system, including cold plates on GPU/CPU packages, in-tray and in-rack manifolds, blind-mate quick-disconnects (QDs), and rack-level coolant distribution units (CDUs); distinct from the facility cooling supply chain in L3.5.
Why the concentration exists
In-rack liquid-cooling hardware comprises cold plates, manifolds, quick-disconnects, and coolant distribution units that work together to remove heat from high-density compute racks. Cold plates attach directly to CPUs and GPUs, circulating coolant through microchannels to absorb heat at the source. Manifolds distribute coolant between the CDU and individual servers, while quick-disconnects allow servers to be connected and disconnected without draining the system. The CDU pumps coolant through the loop and rejects heat to facility water.[1][2][3][6][19]
Direct-to-chip liquid cooling is approximately 3,000 times more effective at heat absorption than air cooling, which becomes necessary as rack densities approach and exceed 50 kW. Nvidia's GB200 NVL72 platform, integrating 72 Blackwell GPUs and 36 Grace CPUs, ships with liquid cooling as standard configuration. Rack-level cooling capacity has scaled rapidly, with CDUs now supporting 100 kW to 300 kW per rack depending on the vendor and configuration.[9][14][19]
Supply concentration stems from the need for tight integration across components, as manifolds, quick-disconnects, CDUs and cold plates must be carefully aligned to ensure balanced fluidics for 24/7 deployment. Patents protect key technologies, such as CoolIT's split-flow heat exchanger and JetCool's microconvective cooling, which uses small fluid jets to target hot spots on processors. Vendors must also obtain certification from chip makers, with Mitsubishi Heavy Industries' cold plates certified by Nvidia, Intel, and AMD.[4][5][13][17][18]
What the evidence shows
Standard lead times are 16-24 weeks for CDUs and 12-16 weeks for cold plates under normal production conditions.
thecoolingreport.comLead times currently run 12-18 months for quality CDUs.
mlq.aiWho supplies it
CoolIT Systems, Boyd, and Motivair ranked as the top three vendors by data center liquid cooling revenues in 2023. CoolIT operates manufacturing sites in Canada, Vietnam, and China, producing cold plate loops, rack manifolds, and CDUs with 100% leak and functional testing on every unit. Supermicro develops its own in-rack and in-row CDUs in-house, with capacity up to 250 kW for in-rack units and 1.8 MW for in-row configurations.[8][12][1]
Chinese manufacturers are expanding into liquid cooling for AI data centers, including Sanhua Intelligent Controls, Lansi Technology, Kehua Data, Shenzhen FRD, and Goaland Energy Conservation. Envicool, a precision cooling specialist serving Nvidia, Intel, Alibaba, and Tencent, saw its Shenzhen-listed shares more than triple over one year. Mitsubishi Heavy Industries offers a Direct to Chip Cooling solution supporting 70 kW per rack with cold plates capable of cooling 1,500W processors.[11][18]
Who controls it
No independently verified market-size figure is published for this node yet.
What it depends on, and what depends on it
Cold plates use internal microchannels or microconvective jet arrays to transfer heat from the chip to the circulating coolant. Quick-disconnects are available in all-metal dry-break designs from multiple suppliers, with GF's polymer Quick Connect Valve 700 offering 25% better flow performance than metal alternatives at over 50% lower weight. Manifolds can be configured for new builds with overhead infrastructure or retrofits using raised floors.[3][6][17][21]
The hardware enables deployment of high-density AI racks that would exceed air-cooling thermal limits. NTT GDC has deployed more than 200 MW of liquid-cooled capacity, primarily direct-to-chip, across six data centers in the United States. Colovore launched its first all-liquid facility in California. Edged US broke ground in 2025 on an Aurora, Illinois facility designed for densities up to 200 kW per rack using its ThermalWorks liquid-to-chip system.[10][22]
Where it sits in the stack
Takes in: Copper/aluminium cold-plate stock, manifold tubing, blind-mate QD couplings, rack CDU pumps and heat exchangers, coolant
Sends on: Direct-to-chip liquid loop integrated into the server tray and rack, transferring GPU/CPU heat to the facility cooling loop (L3.5)
What would break it
New entrants and capacity expansions are emerging in response to demand. LiquidStack inaugurated its Carrollton, Texas headquarters in March 2025, tripling production capacity. Submer ventured into data center design, construction, and services in February 2025. Asperitas partnered with Cisco in February 2025 to combine immersion cooling with Cisco's Unified Compute System.[24]
Alternative cooling architectures offer different trade-offs. Airsys's LiquidRack uses dielectric fluid sprayed onto vertically mounted waterproof servers, consuming 80% less fluid than immersion cooling. Johnson Controls announced a multi-million dollar strategic investment in Accelsius in October 2025 for two-phase direct-to-chip technology that enables 35% OpEx savings over single-phase systems. OVHcloud's fifth-generation racks feature direct-to-chip waterblocks with a cooling module 50% more compact than previous generations.[7][20][22][23]
What to watch
JetCool launched SmartPlate liquid cooling for Dell PowerEdge R770 and R7725 at Dell Technologies World on June 2, 2026. JetCool had previously launched the SmartPlate System, a sealed closed-loop direct-to-chip cooling solution, on May 15, 2026. The company has additional products in development for 2026.[15][16]
Flex launched a Modular Rack-Level Cooling Distribution Unit developed by JetCool on September 23, 2025. The Flex and JetCool partnership delivers racks supporting up to 120 kW with an upgrade path to 300 kW. JetCool's 6U in-rack CDU, launched at the 2024 OCP Global Summit, is designed to deliver over 300 kW of cooling capacity per rack.[14][15]
Airsys introduced LiquidRack, a rack-level liquid cooling platform, on April 20, 2026, supporting up to 80 kW per rack and optimized for CPU and mid-density AI workloads. Edged US broke ground in 2025 on its second Aurora, Illinois facility designed for densities up to 200 kW per rack.[20][22]
Related nodes
Sources
- supermicro.com
- motivaircorp.com
- coolitsystems.com · 2025-06-04T21:01:47
- jaredwatkins.com · 2026-06-07T00:00:00
- jetcool.com · 2026-03-03T15:15:18
- datacenters.lbl.gov
- datacenterdynamics.com · April 24, 2026
- datacenterfrontier.com · 2023
- tomshardware.com
- datacenterdynamics.com
- scmp.com · 15 Feb 2026
- coolitsystems.com
- coolitsystems.com · 2025-04-29
- flex.com · 2024
- leadiq.com · 2026-06-02
- techarena.ai
- techtarget.com · 2024-07-10
- mhi.com
- vertiv.com
- hpcwire.com · 2026-04-20
- gfps.com
- datacentremagazine.com · February 20, 2026
- airsysnorthamerica.com
- introl.com · February 2025
Full scorecard, owner shares, supply edges and the full tracked roster are in the desk letter.
GET THE BRIEFING