Liquid Cooling Market Hits $17.8B as AI Rack Densities Surge

In March 2026, Ecolab agreed to acquire CoolIT Systems for $4.75 billion in an all-cash deal, a price that rewrote the valuation benchmarks for the liquid cooling sector overnight. CoolIT, a Calgary-based maker of direct-to-chip cold plates and coolant distribution units, had been owned by KKR since 2023. The sale returned roughly eight times salary to some frontline employees through an ownership programme KKR had put in place, Inc. reported. The buyer was not a semiconductor firm or a server OEM. It was a water-treatment company with a century of experience managing industrial fluids, and its calculus was simple: the coolant running through a datacenter cold plate is now as strategic as the silicon it cools.

The numbers that frame that calculus have been arriving in quick succession. A Morningstar report published in May 2026 projected the AI datacenter liquid cooling market would reach $17.8 billion by 2036, with the steepest adoption acceleration occurring between 2026 and 2030. A separate forecast from Market Decipher, released in late April, pegged the compound annual growth rate at 17.2 percent, yielding an $18.1 billion market by the same horizon. These are not niche-HPC numbers. They describe a wholesale re-plumbing of the world's compute infrastructure.

The physics driving the shift is unambiguous. Nvidia's Blackwell-class GPUs, now shipping into hyperscale training clusters, push single-rack power densities past 120 kW. Network World reported in March that the datacentre's forty-year reliance on air cooling has struck a physical limit: air simply cannot remove heat fast enough from a rack dissipating more than roughly 30 kW without consuming more energy in fans than the compute itself justifies. At 120 kW per rack, the choice is no longer air versus liquid; it is which liquid architecture, deployed at what scale, with which chemistry.

The industry has converged on two primary architectures, each with its own supply chain and failure modes. Direct-to-chip (D2C) cooling puts a cold plate atop the processor and pumps a dielectric or treated-water fluid through it, capturing 70 to 80 percent of the rack's heat at the source. The remaining heat, from memory modules, voltage regulators, and other board-level components, still requires air handling, which means D2C deployments do not eliminate fans. They relegate them. Immersion cooling, by contrast, submerges entire servers in a dielectric fluid that bathes every component, capturing nearly all the heat in a single-phase or two-phase loop. The trade is straightforward: immersion captures more heat but demands purpose-built hardware and a facility floor rated for tanks that weigh tonnes when filled.

CoolIT built its business on the D2C side, and Ecolab's acquisition thesis rests on the premise that direct-to-chip will be the volume play. The deal, Datacenter Dynamics noted, positions Ecolab to offer an end-to-end fluid management platform: the cold plates from CoolIT, the coolant chemistry from Dow or its own industrial water portfolio, and the monitoring, filtration, and replenishment services that keep a liquid loop from corroding, fouling, or growing biofilm over a twenty-year asset life. Ecolab's management told investors the CoolIT business could reach margins above 20 percent by 2028, a premium that reflects the service-revenue stickiness of ongoing fluid maintenance contracts.

Dow, for its part, is not ceding the chemistry layer. In late May 2026 the company launched its Dow Coolant Care Network, an end-to-end service model that packages coolant supply, system analytics, and lifecycle management into a single contract. The move acknowledges an operational reality that hyperscalers have been grappling with quietly: a D2C loop is a chemical system as much as a mechanical one, and it requires the same kind of ongoing condition monitoring that a district heating network or a refinery cooling tower demands. Glycol degradation, galvanic corrosion, and microbial fouling each represent a single-point-of-failure risk for a $500 million training cluster.

The integration problem extends beyond chemistry. At Data Center World in April 2026, Delta Electronics unveiled an integrated grid-to-chip infrastructure architecture that treats power and cooling as a single design domain. The rationale is that when a rack draws 120 kW, the conversion losses between the utility feed and the GPU core are no longer a rounding error. A 48-volt rack architecture that was efficient at 15 kW becomes a thermal liability at eight times that density. The busbar that carried 300 amps cleanly now runs hot enough to require its own liquid cooling, a problem that compounds as the industry debates a shift to 800-volt rack distribution, a voltage level borrowed from electric vehicle powertrains.

The retrofit question is equally pressing. Most of the world's hyperscale datacentre square footage was designed for air-cooled racks drawing 8 to 15 kW. Retrofitting those facilities for D2C requires running coolant supply and return lines to each row, installing coolant distribution units with pumps and heat exchangers, and in many cases upgrading the facility water loop to reject heat at a higher temperature. Airsys introduced its LiquidRack system in April specifically targeting this retrofit window, promising a simpler architecture for edge and mid-density deployments that cannot justify a full facility rebuild.

Arivor Technologies, a subsidiary of AEWIN Technologies, is taking a different approach to the same problem. At COMPUTEX 2026, the company will showcase its 2P DLC rack-scale solution, a two-phase direct liquid cooling system that Electronics360 reported is designed for AI datacentre upgrades. Two-phase cooling uses a refrigerant that evaporates at the cold plate, carrying heat away as latent energy of vaporisation rather than as sensible heat in a single-phase liquid. The engineering advantage is higher heat transfer coefficients and lower flow rates. The operational challenge is managing a two-phase regime inside a datacentre rack, where vapour quality, pressure drop, and condenser performance must stay within narrow bands across variable GPU loads.

The single-phase versus two-phase debate has not settled, and likely will not until one architecture accumulates enough fleet hours to generate meaningful reliability data. Single-phase proponents point to simpler pump and piping requirements and the absence of a vapour-compression cycle inside the rack. Two-phase advocates counter that the much higher heat transfer coefficients permit smaller cold plates and lower fluid volumes, which reduces weight and simplifies seismic bracing, a non-trivial concern in markets like Japan, Taiwan, and the western United States.

What is clear is that the cooling architecture is now being selected before the building is sited, not as an afterthought once the server purchase order is cut. POWER Magazine reported in April that the datacentre industry is having a power problem at the rack, where conversion losses that were once background noise now represent megawatts of thermal load that must be managed inside the white space. The substation, the backup generator, the uninterruptible power supply, and the coolant distribution unit are now a single engineered system. Break any one of those dependencies and the GPUs stop, regardless of how much redundancy was built into the other three.

This systems-level thinking has begun to attract control-layer investment. In March 2026, Phaidra announced a methodology for agentic liquid cooling management deployed with CoreWeave and Applied Digital on Nvidia Max-Q AI factories, using reinforcement learning to optimise coolant flow rates, pump speeds, and chiller setpoints in real time against GPU workload characteristics. The premise is that a static cooling curve, designed for a worst-case steady-state thermal load, leaves substantial efficiency on the table when workloads are bursty, which AI training and inference workloads invariably are.

The geography of cooling innovation is also shifting. CoolIT operates manufacturing in Canada, China, and Vietnam, with R&D labs in Calgary and Taipei. Arivor is a Taiwanese firm exhibiting at a Taiwanese trade show. Delta, also Taiwanese, is leveraging its power-electronics heritage to sell integrated packages. The liquid cooling supply chain is tilting toward the same Asia-Pacific axis that dominates semiconductor packaging, and for similar reasons: proximity to the server OEMs, the ODMs, and the chip fabs shortens the feedback loop between a thermal specification change at the silicon level and a cold-plate redesign on the manufacturing line.

The coolant itself has become a strategic material. Dow's entry is significant partly because the firm is the world's largest producer of silicone-based dielectric fluids, a category that competes with fluorocarbon-based fluids from 3M and Chemours and with hydrocarbon-based fluids from smaller specialty firms. The fluid choice determines the cooling loop's material compatibility matrix: what elastomers the O-rings can use, what metals the cold plate can be machined from, what pump seal materials are permissible. Get the compatibility wrong and the loop develops a slow leak that contaminates the fluid, corrodes the cold plate, and eventually shorts a GPU. The hyperscalers are acutely aware that this failure mode does not announce itself in a commissioning test; it emerges after 18 months of continuous operation.

The reliability question is the one that keeps facilities engineers awake, and it explains why the adoption curve, though steep in forecast charts, is cautious in practice. A hyperscale operator that commits to D2C across a 200-megawatt campus is making a bet not only on the cold plates and the CDUs but on the entire service infrastructure: the on-site fluid analysis lab, the spare-parts inventory for every pump and every quick-connect fitting, the training programme for the facilities staff who will swap a cold plate on a live rack because the GPU behind it cannot be decommissioned for six more months. That infrastructure does not yet exist at the scale the market forecasts imply it will need to.

The labour dimension is underappreciated. A D2C deployment requires technicians who can handle coolant under pressure, read a corrosion coupon, interpret a dissolved-gas analysis, and troubleshoot a two-phase flow instability. These are skills that overlap more with industrial process engineering than with traditional IT facilities management. The talent pool is shallow, and the hyperscalers are competing for it with the chemical processing and oil and gas industries, which offer comparable salaries and more established career paths.

There is also the question of what happens to the heat once the liquid has carried it out of the rack. A 120 kW rack in a 30-rack row rejects 3.6 megawatts of thermal energy, enough to heat a small town in a Nordic winter. In Reykjavik, where district heating networks are ubiquitous, that heat can be sold. In Ashburn, Virginia, where the world's densest concentration of datacentres sits atop a strained transmission grid, it is rejected to the atmosphere through cooling towers that consume millions of litres of water annually. The liquid cooling revolution solves the problem of getting heat out of the server; it does not solve the problem of what to do with the heat afterwards. That problem belongs to the utility commissioners, the watershed managers, and the transmission planners who are only beginning to factor 500-megawatt datacentre campuses into their load-growth forecasts.

The vendor landscape is consolidating rapidly around this stack. Ecolab bought CoolIT. KKR took its $4.75 billion and its returns. Dow is building a service network. Delta is integrating power and cooling into a single SKU. And at COMPUTEX 2026, outfits like Arivor are showing rack-scale solutions that treat the cooling architecture as a product, not a project. The implication is that within five years, a datacentre operator will not specify air versus liquid cooling any more than a car buyer specifies whether the vehicle has fuel injection. Liquid will be the default, and the only question will be whose fluid, whose cold plate, and whose service contract keeps the loop running until the next GPU generation arrives.

The checkpoint to watch is the third quarter of 2026, when Ecolab expects to close the CoolIT acquisition and the industry will see whether a water-treatment company can integrate a precision cooling manufacturer at the speed the market demands. If it works, expect chemical firms, pump manufacturers, and industrial service providers to begin acquiring their way into the rack. If it stalls, expect the hyperscalers to bring more of the cooling stack in-house, the way they have with power distribution, server design, and increasingly, silicon. Either way, the coolant floor is rising. The only remaining question is who pays for the pumps.