Data Centre Liquid Cooling Market

The data centre liquid cooling market is worth USD 4,992.0 million in 2025 and reaches USD 37,126.0 million by 2035, compounding at 22.22% a year. The figure is built bottom-up: roughly 5,200 megawatts of data centre IT load deployed with liquid cooling in 2025 across direct to chip cold plate systems, coolant distribution units and secondary loops, immersion cooling systems, and rear door heat exchangers and hybrid systems, at an average realised value of USD 0.96 million per megawatt, triangulated against data centre construction, accelerator shipments and supplier disclosures. Liquid cooled capacity grows 26.0% a year as artificial intelligence servers exceed the limits of air cooling, while value per megawatt falls 3.0% a year as designs standardise and production scales. Air cooling, aisle containment covered in our separate containment coverage, and chillers and cooling towers outside the data hall are excluded. This study sits within our data centre infrastructure coverage and follows the published Douglas Insights methodology.

Why can air no longer cool AI servers?

Because artificial intelligence servers pack so much power into each rack that blowing air over them cannot remove the heat fast enough. For decades, data centres cooled servers by pushing cold air through racks and removing hot air, which worked well for racks consuming a few kilowatts to perhaps twenty or thirty kilowatts. The graphics processors and accelerators used to train and run artificial intelligence models consume far more power per chip, and they are packed densely together so they can communicate quickly. Racks built around the latest accelerator systems draw well over one hundred kilowatts, several times what air cooling can handle, and chip makers design these systems to be liquid cooled from the outset. Liquid carries heat far more effectively than air, so cooling plates attached directly to chips, fed with coolant, can remove this heat efficiently. As a result, liquid cooling has moved from a niche used in supercomputers and gaming to a requirement for new artificial intelligence data centres. The exclusive chapter of this report tracks rack power density by deployment type, since density determines whether liquid cooling is required.

What does this market include?

This study covers liquid cooling systems deployed in data centres. Direct to chip cold plate systems cover cold plates mounted on processors and accelerators, the manifolds and piping in racks, and associated hardware, the largest category. Coolant distribution units and secondary loops cover the units that circulate coolant to racks and exchange heat with the facility water system. Immersion cooling systems cover tanks in which servers are submerged in dielectric fluid, in single phase and two phase designs. Rear door heat exchangers and hybrid systems cover liquid cooled doors on racks and systems combining air and liquid cooling. Air cooling equipment, aisle containment, covered in our separate containment coverage, chillers, cooling towers and facility water systems, and the servers themselves sit outside the boundary. Value is measured at the price data centre operators and server makers pay.

Direct to chip or immersion?

Direct to chip cooling has become the mainstream choice for artificial intelligence, while immersion cooling remains a smaller, specialised option. Direct to chip systems pipe coolant to cold plates on the hottest components, removing most of the heat with liquid while fans handle the rest, and they fit into standard rack designs that server makers and data centre operators already understand. Leading accelerator systems are designed around direct to chip cooling, which has made it the default. Immersion cooling submerges entire servers in fluid, removing all heat with liquid and allowing very high densities with no fans, but it requires specialised tanks, servers adapted for immersion and new maintenance practices, and concerns about fluid costs and, for some two phase fluids, environmental regulation of fluorinated chemicals have slowed adoption. The model reflects direct to chip systems as the largest and fastest growing category in value, with immersion growing from a smaller base.

What drives demand?

The first driver is artificial intelligence. The rapid build out of data centres for training and running artificial intelligence models, with very high density accelerator racks, requires liquid cooling.

The second driver is chip power. Each generation of processors and accelerators consumes more power, pushing more workloads beyond the limits of air cooling.

The third driver is energy efficiency. Liquid cooling can reduce the energy used for cooling and allow waste heat reuse, helping operators meet efficiency and sustainability goals.

The fourth driver is density and space. Liquid cooling allows more computing in less space, which matters where land and power connections are constrained.

What restrains the market?

Three restraints are modelled. AI investment cycles are the first: demand depends on continued heavy investment in artificial intelligence infrastructure, and a slowdown would reduce new deployments. Supply chain capacity is second: rapid growth has strained supply of coolant distribution units, cold plates and skilled installers. Retrofit complexity is third: adding liquid cooling to existing air cooled data centres requires facility water systems and changes to buildings, which is costly and slows adoption in older facilities.

Which categories carry the value?

Direct to chip cold plate systems lead with 48% of 2025 value, USD 2,396.2 million, the mainstream choice for artificial intelligence racks. Coolant distribution units and secondary loops hold 26%, USD 1,297.9 million, required alongside direct to chip systems. Rear door heat exchangers and hybrid systems account for 14%, USD 698.9 million, used to bridge air and liquid cooling. Immersion cooling systems contribute 12%, USD 599.0 million. Each category is modelled through 2035 by region.

Where is liquid cooling deployed?

North America leads with 52% of 2025 value, USD 2,595.8 million, growing 21.07% a year, reflecting the concentration of hyperscale artificial intelligence data centre construction in the United States. Asia Pacific holds 26%, USD 1,297.9 million, at 23.6%, driven by China, Japan, South Korea, Singapore, India and Australia. Europe holds 16%, USD 798.7 million, at 22.4%, with growing artificial intelligence capacity and strong efficiency regulation. The Middle East contributes USD 199.7 million and grows fastest at 26.0%, reflecting large artificial intelligence data centre investments in the Gulf, Latin America USD 69.9 million at 22.0% and Africa USD 30.0 million at 22.0%. Six regional models sum to the global figure, with country tables in the Excel model.

Who supplies data centre liquid cooling?

Vertiv and Schneider Electric, which acquired liquid cooling specialist Motivair, supply coolant distribution units and complete cooling systems, and nVent, CoolIT Systems and Boyd supply cold plates and liquid cooling components. Server makers integrate liquid cooling into their systems, working closely with accelerator designers. Immersion specialists include LiquidStack, GRC and Submer. Asian manufacturers supply cold plates and components at scale. The competitive chapter profiles each supplier’s product range, capacity and customer positions.

How is liquid cooling priced?

Average realised value is USD 0.96 million per megawatt of IT load in 2025, covering cold plates, manifolds, coolant distribution units and associated hardware within the data hall. Costs vary with system design and redundancy, and immersion systems are priced differently from direct to chip systems. Prices per megawatt are expected to fall as designs standardise, production scales and competition grows, which is the reason for the negative price leg. The pricing chapter publishes price bands by category.

How do the scenarios diverge by 2035?

The base case carries 26.0% growth in liquid cooled capacity and a 3.0% annual decline in value per megawatt for a 22.22% revenue CAGR and USD 37,126.0 million in 2035. The AI-slowdown scenario, in which investment in artificial intelligence infrastructure cools, sets the legs at 16.0% and minus 4.0%, landing near USD 14,510 million. The AI-acceleration scenario, in which chip power keeps rising and liquid cooling becomes standard in most new data centres, sets them at 32.0% and minus 2.0%, carrying the market past USD 64,900 million. Each 1-point change in capacity growth moves the 2035 figure by roughly USD 2,950 million. Confidence is moderate given dependence on the pace of artificial intelligence investment.

Which rules and standards apply?

Three layers matter. Energy efficiency regulation comes first: data centre energy reporting and efficiency requirements, particularly in Europe, encourage efficient cooling and waste heat reuse. Chemical regulation is second: proposed restrictions on fluorinated substances affect some two phase immersion fluids, influencing technology choices. Industry standards are third: specifications for liquid cooling interfaces, coolants and rack designs developed by industry groups support interoperability and adoption. The regulatory chapter maps these requirements.

Can waste heat be put to use?

Liquid cooling produces warmer water than air cooling, which makes it easier to reuse data centre waste heat for district heating, greenhouses or industrial processes, turning a waste product into a resource. Several data centres in Northern Europe already supply heat to local heating networks, and regulations in some countries encourage or require waste heat reuse for new data centres. Heat reuse depends on having nearby users of heat, which many data centre locations lack, so it will remain limited to suitable sites. For this market, waste heat reuse strengthens the case for liquid cooling, particularly in Europe. The model treats it as a supporting factor rather than a main driver.

Douglas Exclusive: the rack density tracker

This report tracks, by deployment type and region, rack power densities, accelerator shipments, the share of new capacity requiring liquid cooling and technology choices, converting data centre construction into liquid cooled capacity and value by category and region. Licence holders receive it as a maintained tab in the Excel model.

Methodology and receipts

The model is built bottom-up from capacity: data centre construction by region and type, accelerator and high density server deployments, liquid cooling adoption rates, and realised values per megawatt from supplier disclosures, with air cooling, containment, facility chillers and cooling towers, and servers excluded. Every figure carries a numbered source and a confidence grade in the fact sheet above, and the working model ships with every licence. The next scheduled review of this study is September 2027.