Liquid cooling infrastructure
BOTTLENECKConcentrated CDU and cold plate suppliers command premiums during shortages while custom manifold fabricators capture elevated margins.
Equipment moving liquid coolant directly to chips and servers includes direct-to-chip cold plates, coolant distribution units, manifolds, rear-door heat exchangers and full immersion tanks. Liquid cooling handles 1000W+ per GPU in modern AI clusters where air cooling fails. CDU makers and cold plate suppliers command premiums amid supply shortages; custom manifold fabricators also capture elevated margins.
Why the concentration exists
Liquid cooling infrastructure moves coolant directly to chips and servers through cold plates, coolant distribution units, manifolds, and immersion tanks. Water has about four times the specific heat capacity of air and about 25 times the thermal conductivity, making liquid cooling approximately 3,000 times more effective than air cooling for high-performance computing infrastructure. Direct-to-chip liquid cooling can remove approximately 90% of the total heat generated within a rack by extracting thermal load directly from CPUs and GPUs via cold plates. TSMC demonstrated that direct-to-silicon liquid cooling could cool a reticle-sized die dissipating 2kW of power using 40°C water with less than 10W of pump power.[26][23][21][6]
Rack densities in data centers have rapidly surpassed 100kW, and Nvidia's 120kW, 72-GPU GB200 NVL72 rack is exclusively cooled via direct-to-chip liquid cooling. The latest Nvidia-based GPU server racks require 142 kW of power, with the next evolution scheduled for release in less than a year requiring around 240 kW per rack. Direct-to-chip liquid cooling requires multiple cooling loops and cooling distribution units to manage the thermal load. Sizing the facility water loop for megawatt-scale rows of NVL72 racks is a primary constraint in liquid-cooled site design.[14][11][10][27]
The coolant distribution unit manages the interface between the IT room loop and the heat rejection loop, and facilities cannot deploy liquid-cooled racks without CDUs. The remaining 10% of heat in direct-to-chip deployments comes from components not hydraulically connected to the liquid loop, such as memory modules and power delivery components. Direct-to-chip liquid cooling reduces energy use by 30% to 60% in AI data centers. In data centers, direct-to-silicon liquid cooling can reduce overall cooling infrastructure requirements by nearly half according to TSMC.[10][21][19][6]
What the evidence shows
CDU lead times currently run 12-18 months for quality units.
mlq.aiChiller lead times run 20 to 30 weeks under normal conditions.
thecoolingreport.comWho supplies it
Schneider Electric acquired a controlling interest in Motivair in February 2025 to combine power and liquid cooling infrastructure for AI data centers. Eaton acquired Boyd Thermal for $9.5 billion in 2025, Trane purchased Stellar Energy Digital, and Daikin bought Chilldyne in the same year. Vertiv reported net sales of about $13.5 billion and is positioned in the path of Blackwell deployments as an Ohio-based data center infrastructure provider and Nvidia partner. These acquisitions reflect how incumbents, HVAC vendors, electrical OEMs, and chemistry players are converging on integrated cooling platforms.[20][13][16][3]
CoolIT Systems partners with top OEMs to deliver direct liquid cooled servers featuring pre-installed cold plate loops for scalable, high-performance computing. LiquidStack's GigaModular CDU platform, part of Trane Technologies, has validated capacity expanded to 14 MW and supports pay-as-you-grow expansion for AI campuses. Accelsius launched its NeuCool HyperStart program at Data Center World 2026 to help hyperscale operators and neocloud providers evaluate two-phase cooling. Johnson Controls made a multi-million dollar investment in Accelsius in October 2025 for two-phase direct-to-chip liquid cooling technology.[7][12][12][22]
Envicool counts Nvidia, Intel, Alibaba, and Tencent among its customers for precision cooling, and its Shenzhen-listed shares more than tripled over the past year. Chinese companies including Sanhua Intelligent Controls, Lansi Technology, Kehua Data, Shenzhen FRD, and Goaland Energy Conservation have announced plans to expand into liquid-cooling systems. Asetek has been developing liquid cooling solutions for graphics cards for more than 10 years using patented sealed all-in-one liquid cooling technology. Asetek had more than 1.8 million liquid cooling units deployed in the field as of 2014.[15][15][17][17]
Who controls it
What it depends on, and what depends on it
The coolant in direct-to-chip cooling is typically highly purified deionized water with corrosion inhibitors and chemical stabilizers. Given the risks and drawbacks of water-based cooling inside racks and servers, the industry has been moving toward more-efficient chemical alternatives such as dielectric fluid. Dielectric fluid is a non-conductive liquid that can provide cooling without the electrical risks of water-based systems. The specific heat capacity of water makes it effective for heat transfer, but dielectric fluids enable immersion cooling approaches.[21][8][8][26]
The facility water loop is the building-side cooling water circuit that receives heat from the coolant distribution unit and carries it to heat-rejection equipment such as cooling towers, dry coolers, or chillers. In liquid-cooled GPU facilities, the capacity and supply temperature of the facility water loop set the ceiling on how much GPU heat the site can reject. Direct-to-chip liquid cooling uses a water-based coolant circulated through a cold-plate heat exchanger located directly on the chip. Vertiv's reference architecture for NVIDIA GB200 NVL72 servers reduces annual energy consumption by 25%, cuts rack space requirements by 75%, and shrinks power footprint by 30%.[27][27][25][24]
Hyperscalers continue to anchor demand for liquid cooling, accounting for a substantial share of market revenue, with a significant portion of the remainder tied to deployments in colocation facilities. Around half of Google's global data center footprint has liquid cooling enabled or deployed, with about 1GW of liquid cooling capacity across 2,000 pods. Schneider Electric's liquid cooling solution supports AI racks scaling up to 227 kW per rack, co-developed with NVIDIA for the Vera Rubin NVL72 reference design. Motivair's Coolant Distribution Units scale from 105 kilowatts to 2.5 megawatts and enable thermal performance for 6 of the world's top 10 supercomputers.[1][13][19][20]
Where it sits in the stack
Takes in: Facility water or dielectric fluid; high-density rack heat loads
Sends on: Chip-level heat removed and transferred to facility cooling loop
What would break it
Risk scenarios that could change the supply picture include cold plate commoditization, CDU supply bottlenecks, immersion cooling disruption, and operational reliability failures. The liquid cooling transition co-depends on the power architecture shift toward 800 VDC convergence. Brownfield conversions of existing air-cooled facilities to support liquid-cooled racks represent a significant challenge for operators. The market is moving faster than most operators planned for, making liquid cooling investment a current competitive requirement rather than a future consideration.[2][2][2][5]
What to watch
Edged US's second Aurora, Illinois facility, breaking ground in 2025, aims to save over 277 million gallons of water annually and support densities up to 200kW per rack using closed-loop waterless cooling. Equinix became the latest top-level colocation provider to make liquid cooling a standard part of their service offerings. OVHcloud's Smart Datacenters reduce cooling electricity consumption by up to 50% and water usage by up to 30% via direct-to-chip waterblocks and a closed-loop water circuit. Schneider Electric and Motivair delivered $290 million in phased power and liquid cooling solutions for TeraWulf's Lake Mariner campus in Buffalo, New York.[22][4][22][19]
Milestones being monitored include Vera Rubin volume ramp timing, MCL qualification timing, and CDU lead-time normalization. Intel plans to squeeze one-trillion transistors onto a single chip by 2030, which will require advanced thermal management. Georgia Tech's liquid cooling technology, developed by Daniel Lorenzini, was awarded a U.S. patent in September 2024, and startup EMCOOL is actively pursuing venture capital funding to scale. Schneider Electric revealed its first complete liquid cooling portfolio for hyperscale, colocation and high-density data centre environments on September 29, 2025.[2][25][9][18]
Related nodes
Sources
- delloro.com · 2026-01-08T13:00:21
- thediligencestack.com · 2026-04-09T14:55:07
- stlpartners.com · 2026-05-14T14:12:49
- datacenterfrontier.com
- nautilusdt.com · 2026-06-16T15:10:50
- tomshardware.com · 2025-12-16T11:00:00
- coolitsystems.com · 2025-11-24T21:05:37
- datacenterfrontier.com
- gatech.edu · September 2024
- datacenterdynamics.com · November 09, 2025
- newsletter.semianalysis.com · Feb 13, 2025
- datacenterfrontier.com · 2026
- datacenterdynamics.com · 2025
- datacenterdynamics.com
- scmp.com
- tomshardware.com
- asetek.com · 1 Jul, 2014
- datacentremagazine.com · 2025-09-29
- se.com
- se.com · 29/09/2025
- dcx.eu
- datacentremagazine.com · 2025
- vertiv.com · 2024-08-21
- blogs.nvidia.com
- hdrinc.com · March 24, 2023
- tierpoint.com
- leviathansystems.co
Full scorecard, owner shares, supply edges and the full tracked roster are in the desk letter.
GET THE BRIEFING