TrendForce forecasts liquid-cooling penetration among AI chips at 53 percent in 2026, up from 33 percent in 2025 and representing a 47-percentage-point increase from the 6 percent baseline in 2023, with NVIDIA, AMD, and Google cited as demand drivers [S4].
Buyer-side demand is anchored by racks routinely above 60 kW, EU colocation hot-water reuse rules, and HPC retrofits, with selection resolved at three engineering gates: per-rack heat load, loop topology, and the supplier's certification chain [S1].
Tiered Market: OEM Loops, DLC Specialists, Component Makers
The 2026 supplier map is a three-tier structure rather than a single leaderboard: server-OEM integrated loops, independent direct-liquid-cooling (DLC) specialists, and ISO 9001/14001 component-tier fabricators split addressable revenue by load profile and loop topology [S1]. Lenovo Neptune 6th Gen leads the OEM-integrated tier with three sub-modes, full-system warm water, Neptune Core component-level cold plates, and Neptune Air liquid-assisted retrofits, paired with a rear-door rack water heat exchanger and over 12 years of field deployment, with claimed up to 100% heat removal on a single rack [S5][S7]. CoolIT Systems occupies the second tier as a vendor-neutral DLC specialist profiled in Data Center Dynamics' 20-year retrospective; its warm-liquid architecture lets most CRAC units and chillers be removed once rack power crosses roughly 20 kW, the documented efficiency inflection point for AI training and HPC [S1].
The component tier includes Koolance (ISO 9001 + ISO 14001 dual-certified), whose ALX-15KA is a 15 kW floor-standing ambient CDU drawing 700 W maximum on 100-240 V AC, and whose VLX-2000 is a 2 kW 110 V recirculating chiller; an optional rack-mount adapter tray for the ALX-3000 was released on 2026-06-05, signalling active 2026 SKU refresh [S5]. Related coverage in the Data Center Cooling 2026: Liquid Surge, Vendor Map, and Spec Gate article frames the broader hyperscale spending context behind these vendor moves.
Heat Load Thresholds and Rack-Scale Pressure
Rack-scale density is the load that drives every spec decision, and the 2026 numbers are blunt: NVIDIA GB200 NVL72 systems operate at 120-132 kW per rack, while Dell's rack-mounted CDU delivers up to 160 kW to support configurations with up to 96 AMD Instinct MI355X GPUs, with the 160 kW figure reflecting current product specification rather than a 2026 performance result [S4]. Microsoft GB300 NVL72 racks reach 136 kW on TrendForce data, and NVIDIA's Vera Rubin platform uses a fanless, all-liquid-cooled architecture that pulls thermal management beyond GPUs to the broader rack [S4].
Google is the most aggressive hyperscaler, with liquid cooling already deployed in more than 80% of its AI servers, well ahead of the 53% market average for 2026 [S4]. The 20-60 kW/rack band is the warm-water DLC sweet spot for CoolIT-class specialists and Neptune-class OEM loops; above 60 kW/rack, cold-plate MOQ, hydraulic compatibility with the chosen CDU, and zero-leak helium-tested qualification dominate the spec gate, and component-tier vendors with low MOQ win on prototype speed [S1].
CDU and Immersion Product Landscape
The 2026 direct-to-chip CDU field is fully shipping with published pricing and lead times: Vertiv Liebert CoolThru supports up to 100 kW/rack at $45,000-$85,000 with 14-18 week lead time, Schneider EcoBreeze DX integrates with EcoStruxure DCIM at $38,000-$72,000 and 16-20 weeks, and Airedale AirePointe targets retrofits at $28,000-$55,000 with a class-leading 10-14 week lead time for mixed-tenant colos [S3]. Immersion products have also gone mainstream: GRC ICEraQ Series 10 handles up to 200 kW per tank at $120,000-$200,000 with 20-26 week lead time and integrated fire suppression, LiquidStack two-phase immersion runs 22-30 weeks due to demand, and Submer MicroPod targets colo tenants at $85,000-$140,000 with 12-16 week lead time [S3].
Cold-plate material choice and brazing process are the first hard engineering trade-offs because alloy windows differ: copper-brazed cold plates from ToneCooling hit 0.04 degC/W thermal resistance at standard flow rates, against a buyer benchmark of sub-0.05 degC/W, with a 5-piece prototype MOQ and 7-15 day turnaround for AI server builds [S2]. Winshare Thermal specifies liquid cold plates in copper or aluminum with internal fluid channels sized to maximize convection, while Mstirling lists both vacuum brazing and controlled atmosphere brazing for brazed plate-fin heat exchangers [S1].
Vendor Capability Matrix Across Decision Criteria
Four decision criteria separate the 2026 short list: cold-plate MOQ, platform coverage (international vs Chinese GPU support), dielectric vs deionised-water loop handling, and stated annual capacity. ToneCooling publishes the lowest cold-plate MOQ at 5 pieces with explicit support for NVIDIA GB200 plus Chinese platforms (MetaX, Biren, Sunway, Hygon) and a 500,000+ units/year capacity [S2]. Boyd Corporation (USA) and CoolIT Systems (Canada) list 100+ piece cold-plate MOQs, with CoolIT holding Dell/HPE OEM partnerships and Boyd serving Fortune 500 clients at global scale [S2]. LiquidStack (Netherlands) and ZutaCore (Israel) cover the two-phase immersion and dielectric waterless niches respectively, with ZutaCore offering a waterless zero-leak guarantee, and Schneider Electric (France) supplies CDU and infrastructure at the systems tier [S2].
For buyers building on Chinese accelerator stacks, only ToneCooling advertises simultaneous support for MetaX, Biren, Sunway, and Hygon alongside NVIDIA GB200, a combination that matters for hyperscale and HPC projects that span both ecosystems [S2]. Loop topology drives fluid selection, which in turn drives materials and leak-detection discipline: buyers should pin dielectric vs deionised-water choices, define allowable coolant conductivity, and require materials compatible with the chosen coolant, since dielectric fluids pull different elastomer and connector specs than deionised water loops [S1].
Standards, Certification, and Cross-Discipline Linkage
ISO 9001 and ISO 14001 dual certification is the public, auditable quality/environmental signal on the component tier; OEM-integrated vendors like Lenovo inherit server-platform safety/EMC certifications on the rack PDU but do not advertise a standalone ISO 14001 on the cooling subsystem [S5]. Helium leak testing is the dominant zero-leak qualification on the specialist tier, paired with stated thermal resistance below 0.05 degC/W at standard flow rates as the engineering floor for new cold-plate designs [S2].
The same thermal-management logic that resolves AI rack cooling is now migrating into commercial and industrial battery storage, where Huntkey's 215 kWh LFP cabinet specifies a cell-to-cell temperature delta under 3 degC, a constraint that pulls the same flow-meter and pressure-transmitter instrumentation discipline as a coolant distribution loop [S7]. Buyers familiar with the underlying fluid and control instrumentation will recognize that DLC, immersion, and LFP ESS cooling all hinge on the same loop-side measurement chain, from coolant flow and differential pressure to rack-inlet temperature, covered in the pressure sensor reference. The 2026 demand for component-tier fabricators with low MOQ, hydraulic-compatibility data, and free CFD simulation is, in short, the industrial valve and quick-disconnect supply chain under a different marketing label.
Selection Discipline and Trackable Signals
The engineering recommendation is to resolve the 2026 build at three gates before vendor selection: per-rack heat load (below 20 kW still belongs to air, 20-60 kW is warm-water DLC sweet spot, above 60 kW requires helium-tested qualification), loop topology (warm-water vs cold-water vs two-phase immersion, with dielectric vs deionised-water pinned on the RFQ), and the supplier's certification chain (ISO 9001 minimum, ISO 14001 preferred on the component tier) [S1][S5]. NVIDIA's Vera Rubin fanless all-liquid architecture and AMD's Helios platform, with MI500 fully on liquid cooling, are the two rack-level signals to track through 2027 shipments [S4].
Two trackable nodes for the next 90 days: the 2026-06-05 Koolance ALX-3000 rack-mount adapter tray release signals continued component-tier SKU refresh, and the LiquidStack two-phase immersion 22-30 week lead time is the supply-side bottleneck for hyperscale two-phase builds [S3][S5].