Vertiv holds a narrow lead over LiquidStack for liquid cooling infrastructure paired with high-density GPUs. The two providers compete closely alongside immersion and direct-to-chip alternatives.
The largest cloud service providers (CSPs) design or contract custom data centers that accommodate liquid-cooled server configurations for massive AI training models:
Microsoft Azure: Has been deploying direct-to-chip liquid cooling infrastructure across its specialized AI data centers to support massive multi-node GPU clusters.
Amazon Web Services (AWS): Implements advanced liquid and hybrid cooling systems for ultra-dense EC2 instances housing next-generation machine learning accelerators.
Google Cloud: Utilizes custom cooling architectures—combining high-efficiency water and targeted liquid cooling loops—to support its Tensor Processing Unit (TPU) pods and dense external GPU deployments.
Oracle Cloud Infrastructure (OCI): Designs high-performance AI clusters that increasingly rely on specialized cooling distributions to handle dense NVIDIA GPU architectures seamlessly.
2. Specialized Colocation & Data Center Operators
Many enterprises do not build raw facilities themselves; instead, they rely on carrier-neutral colocation and wholesale data center providers equipped for liquid cooling:
Equinix: Offers tailored high-density power and liquid-cooling retrofits/deployments across key International Business Exchange (IBX) data centers globally to support AI enterprise deployments.
Digital Realty: Integrates liquid-cooling-ready data halls (supporting high kilowatts-per-rack metrics) across its global footprint, allowing enterprise AI deployments without localized thermal bottlenecks.
CoreWeave & Lambda Labs (AI-Specialized Clouds): Specialized GPU cloud providers that build or partner with facilities specifically engineered from the ground up for high-density direct-to-chip liquid cooling.
Vertiv: A dominant player providing end-to-end thermal management, including direct-to-chip liquid cooling systems and large-scale fluid management utilized by major cloud operators.
Schneider Electric (incorporating Motivair): Delivers integrated data center power alongside precision floor-mounted and rack-level liquid-to-liquid cooling loops.
LiquidStack: Specializes in hyperscale liquid immersion cooling and modular liquid cooling architectures (such as their GigaModular systems).
Submer: Engineers single-phase immersion cooling ecosystems where entire server blades are submerged in dielectric cooling fluids.
Would you like to explore direct-to-chip cooling vs. immersion cooling for a specific AI hardware stack, or are you looking for regional availability from a particular provider?
Yes. As of October 2026, liquid cooling has moved from a niche HPC capability to a mainstream requirement for the highest-density AI infrastructure. TrendForce estimates liquid cooling penetration for AI chips at about 53% in 2026, with further growth expected in 2027.
The providers fall into three useful groups:
Major cloud / AI-cloud providers
Provider
Liquid cooling support
What it means
Microsoft Azure
Production
Microsoft says it has deployed hundreds of thousands of liquid-cooled Grace Blackwell GPUs and is rolling Rubin into modern liquid-cooled Azure facilities. Its newer designs use closed-loop, direct-to-chip cooling.
Google Cloud
Production / infrastructure technology
Google operates liquid-cooled AI infrastructure and has developed Brazos, a generally available rack-level closed-loop liquid-to-air system for putting liquid-cooled AI equipment into existing air-cooled facilities.
AWS
Production / expanding
AWS has developed configurable liquid-to-chip cooling for new and existing facilities, supporting systems such as Trainium3 and NVIDIA GB200 NVL72.
Oracle Cloud Infrastructure (OCI)
Production
OCI has live liquid-cooled NVIDIA GB200 NVL72 deployments and is building large AI facilities using closed-loop cooling.
CoreWeave
Core capability
One of the clearest examples of an AI-native cloud designed around liquid cooling. CoreWeave says many of its AI data centers use closed-loop direct-to-chip cooling, including production-scale Blackwell deployments.
Lambda
Core capability
Lambda's Superintelligence Cloud uses high-density liquid-cooled AI factories and single-tenant GPU superclusters; it specifically describes liquid cooling as a baseline requirement for rack-scale AI.
Nscale
Core capability
Nscale designs/builds AI data centers with closed-loop liquid cooling; its Ward County, Texas facility uses direct-liquid cooling and its Iceland deployment is a large liquid-cooled Blackwell installation.
Nebius
Production
Nebius has installed NVIDIA GB200 NVL72 liquid-cooling infrastructure in its Finland data center and a Kansas City colocation facility.
Crusoe
Production / infrastructure
Crusoe's Abilene AI factory, supporting OCI, uses direct-to-chip liquid cooling at 1.2 GW planned scale.
Vultr
Supporting next-gen deployments
Vultr supports AMD Helios rack-scale infrastructure with direct-liquid-cooling options.
Colocation / data-center infrastructure providers
These are particularly interesting if you're looking to deploy your own GPU servers rather than rent GPUs from a public cloud.
Strongest options:
Equinix — direct-to-chip liquid cooling, liquid-to-cage deployments, and AI-ready high-density facilities. Equinix says liquid cooling is being built into new facilities and supports private liquid-cooled environments.
Digital Realty — direct liquid-to-chip cooling available across a large portion of its global footprint; it announced the capability across 170 facilities and continues expanding high-density AI deployments.
DataBank — its new Universal Data Hall design includes pre-installed liquid-cooling taps, CDUs and water-cooled cabinet densities of , with up to 2 MW per row.
NVIDIA's ecosystem is a useful shortcut
NVIDIA maintains a DGX-Ready Colocation program specifically for this market. Its current directory identifies providers with advanced liquid-cooled or hybrid liquid/air capabilities. The directory includes Aligned, Colovore, DataBank, Digital Realty, Equinix, Flexential, Switch, QTS, Vantage, atNorth, Borealis, Verne Global, Sify, AirTrunk, Global Switch, Khazna, BDx, PDG, STT GDC, among others.
My practical shortlist
If you're evaluating providers for AI infrastructure procurement, I'd separate them this way:
For renting GPU compute:
CoreWeave
Lambda
Nscale
Nebius
OCI
Azure
Google Cloud
AWS
For deploying your own liquid-cooled GPU cluster:
Equinix
Digital Realty
DataBank
QTS
Flexential
Aligned
Switch
Iron Mountain
NTT
EdgeConneX
The important caveat is that "supports liquid cooling" does not mean every facility or every rack does. Availability is often site-, hall-, rack-density-, CDU-, and coolant-specification dependent. Equinix, for example, distinguishes liquid-to-cage deployments from liquid-to-air systems and has specific requirements around CDUs and coolant.
If you're doing this for an investment/market map, I can also build you a much more useful table of ~30–50 providers with company, geography, MW capacity, liquid-cooling type (D2C/immersion/rear-door), maximum kW/rack, GPU cloud vs. colocation, NVIDIA/AMD platforms, and whether the capacity is live vs. under construction.
Expanding a data center for high-density AI workloads requires shifting away from conventional air cooling toward advanced thermal architectures . Liquid cooling suppliers in this space generally split into major infrastructure providers (delivering Coolant Distribution Units and direct-to-chip manifolds) and specialized immersion/precision cooling innovators.
Top suppliers providing liquid cooling solutions for high-density GPU environments include:
Vertiv Holdings Co. – A dominant player in AI data center infrastructure, providing end-to-end thermal management, power distribution, and scalable direct-to-chip liquid cooling systems tailored for dense GPU clusters.
Schneider Electric (with Motivair) – Following its acquisition of Motivair, Schneider Electric offers a comprehensive portfolio of high-density liquid cooling tech, including Coolant Distribution Units (CDUs), rear-door heat exchangers (RDHx), and dynamic cold plates built for accelerated computing.
LiquidStack – Specializes in high-capacity single-phase and two-phase liquid immersion and direct-to-chip cooling, featuring modular CDU platforms (such as the GigaModular line) designed to scale cleanly for extreme AI workloads.
STULZ – A mission-critical thermal management vendor that expanded its dedicated liquid-cooling manufacturing and R&D footprint to deliver precision direct-to-chip and high-density rack cooling systems for AI.
Delta Electronics – Provides integrated grid-to-chip infrastructure, supplying high-efficiency server power solutions alongside advanced liquid-to-liquid and liquid-to-air CDUs explicitly designed to minimize the footprint of heavy GPU workloads.
Iceotope Technologies – Focuses on precision chassis-level and precision-targeted liquid cooling, enclosing servers in a sealed dielectric environment that dramatically cuts cooling energy and water consumption for high-density blades.
Rittal – Offers heavy-duty direct liquid cooling (DLC) infrastructure, including high-capacity in-row and in-rack CDUs capable of scaling up to megawatt-level outputs to handle extreme thermal densities from modern GPU arrays.
Airedale by Modine – Delivers custom and modular coolant distribution units (CDUs) coupled with intelligent software optimization to balance hybrid air-and-liquid environments for hyperscale and colocation clients.
To help narrow down which supplier fits your expansion, could you share:
Your target rack density (e.g., 50 kW to over 100 kW per rack)
Whether you prefer direct-to-chip (cold plate) or full dielectric immersion cooling
If you are retrofitting an existing facility or building a greenfield site
For a new AI data-center buildout, the supplier landscape is fairly broad, but the most relevant names fall into a few categories. Current evidence shows the market moving strongly toward direct-to-chip (DTC) liquid cooling with coolant distribution units (CDUs) for high-density GPU racks.
Recently qualified a 2.3 MW CoolChip CDU as NVIDIA DSX Ready; strong fit for hyperscale/colo deployments.
250 kW+
QTS — explicitly offers liquid-cooling capability for AI and other cooling-intensive workloads through its Freedom design.
Flexential — has facilities with built-in liquid-cooling capability, including its Atlanta-Douglasville site.
Iron Mountain Data Centers — offers high-density facilities with liquid cooling and says its current designs support up to roughly 50 kW/rack, with higher densities planned.
Switch — its EVO AI Factories are designed around liquid-to-air and liquid-to-chip cooling, with its architecture targeting extremely high rack densities.
NTT / NTT DOCOMO BUSINESS — offers direct-liquid-cooling-compatible facilities in Japan; its Green Nexcenter service supports 20–80 kW/rack.
EdgeConneX — building AI factories using hybrid direct-to-chip liquid + air cooling; its Lambda Chicago project is a 23 MW AI facility scheduled for 2026.
Verne Global — Nordic AI/HPC colocation provider with liquid-cooled AI infrastructure, particularly attractive for renewable-energy-intensive workloads. NVIDIA lists Verne among its DGX-ready partners.
atNorth — Nordic HPC/AI data-center operator; its liquid-cooling capability was one reason Equinix highlighted it in connection with its acquisition of atNorth.
Sify — NVIDIA lists its Mumbai facility as liquid-cooled ready.
AirTrunk — liquid-cooled ready across parts of its APAC/Japan portfolio.
Global Switch — supports both high-density air and liquid cooling and is explicitly listed as liquid-cooled ready by NVIDIA.
ST Telemedia Global Data Centres (STT GDC) — NVIDIA lists Singapore and Thailand locations as liquid-cooled ready.
Princeton Digital Group (PDG) — NVIDIA lists Japan capacity as liquid-cooled ready.
BDx Data Centers — NVIDIA lists Indonesia capacity as liquid-cooled ready.
Khazna — NVIDIA lists Khazna's MENA facilities as liquid-cooled ready.
MEGA D.C. — NVIDIA lists its Israel facilities as liquid-cooled ready.
Borealis Data Center — Iceland; NVIDIA lists it as liquid-cooled ready.
Aligned Data Centers — supports its own DeltaFlow liquid-cooling architecture and says it can exceed 300 kW/rack.
CoolIT Systems
DTC cold plates, CDUs
CHx2000 CDU supports 2 MW, aimed at dense AI racks including GB300-class systems.
Featured in Schneider's GB300 reference architecture for three 1,152-GPU clusters.
Boyd Corporation
Cold plates, CDUs, liquid loops
NVIDIA identifies Boyd among its Blackwell infrastructure providers; NVIDIA also highlights Boyd's high-density AI cooling portfolio.
Delta Electronics
CDUs and complete thermal infrastructure
Its GoCool 3 MW CDU is shipping for high-density AI/HPC deployments.
LG Electronics
Chillers, CDUs, cold plates
Its 2.6 MW CDU qualified as NVIDIA DSX Ready in September 2026, following 600 kW and 1 MW qualifications.
Trane Technologies
CDUs and facility cooling
Offers direct-to-chip CDUs up to 14 MW for high-density AI/HPC environments.
GRC
Single-phase immersion cooling
An alternative when you want to move beyond conventional DTC cooling toward immersion.
Submer
Immersion cooling
Focused on high-density immersion deployments rather than conventional cold-plate/CDU architectures.
NVIDIA's own Blackwell infrastructure ecosystem is useful for supplier screening: it lists Amphenol, Auras, AVC, Boyd, CoolIT, Danfoss, Delta, Flex, Foxconn, Motivair, Parker, Schneider Electric, Vertiv, TE Connectivity, Staubli and others among infrastructure providers.
What I'd put on an RFQ shortlist
For a new high-density GPU hall, I'd separate the procurement into:
Facility-side cooling/chillers — Schneider Electric, Vertiv, Trane, LG and conventional chiller suppliers.
Quick disconnects / fluid connections — CPC, Parker Hannifin, Danfoss, Staubli. TrendForce specifically identifies these suppliers in the NVIDIA GB200 ecosystem.
Immersion — GRC or Submer if your rack/server architecture permits it.
One important procurement point: don't evaluate a CDU in isolation. For Blackwell-class systems, the cold plates, manifolds, QDs, CDU, facility-water temperatures, rack power and server OEM all need to be treated as one thermal architecture. NVIDIA's Blackwell designs are explicitly built around liquid-cooled rack architectures.
If you tell me your target GPU (H200, B200, GB200, GB300, etc.), rack density (kW/rack), and total MW of the expansion, I can narrow this to a 5–8 vendor shortlist and compare cooling capacity, CDU topology, rack compatibility, NVIDIA validation, and likely deployment model.
