Direct-to-chip liquid cooling is not growing because the market decided to embrace a new technology. It is growing because rack densities crossed a threshold that air cooling cannot manage, and operators have no other option for the workloads they are running. A MarketsandMarkets report published April 28, 2026 puts the global direct-to-chip liquid cooling market at $3.33 billion in 2026, growing at a 26.5% compound annual growth rate through 2032 to reach $17.31 billion. Those numbers are the thermal load math expressed in dollars.
AI inference racks are running 85 to 130 kW per cabinet. Liquid cooling is no longer optional at those densities. The forecast is large, the CAGR is steep, and the real story is in what the segment breakdown reveals about which operators are driving it and which geographies are moving fastest.
The hyperscale data center segment is projected to grow at 27.1% CAGR through 2032, slightly above even the market average. Amazon Web Services, Microsoft Azure, and Google Cloud are the anchor customers, but the hyperscale buildout extends considerably beyond those three. Tier 2 cloud providers, government AI programs in the US and EU, and sovereign cloud initiatives across the Middle East and Asia Pacific are all commissioning liquid-cooled GPU clusters at a pace that keeps procurement teams at every major CDU vendor managing 16 to 24 week backlogs.
The hyperscale customer is also the most demanding specification writer. Large operators do not buy off-catalog CDUs and cold plates. They drive OEM designs, often producing custom thermal architectures that never reach the commercial market but set the engineering standards that commercial vendors follow two to three years later. The 27.1% CAGR in the hyperscale segment reflects both the scale of their buildout and the engineering investment they are making in liquid cooling infrastructure that raises the bar for the entire industry.
North America holds the largest current market share. Asia Pacific is growing fastest at a 27.8% CAGR, edging out even the global average. The composition of that growth is important. China's national AI programs and data center construction surge are the largest single driver. Singapore is running at capacity and expanding. Tokyo, Mumbai, and Jakarta are seeing hyperscale investment that was negligible three years ago. State-backed AI programs across the region are commissioning GPU clusters with timelines that make North American operators' schedules look relaxed.
The Asia Pacific growth rate also reflects a different starting point. North American hyperscalers began liquid cooling deployments earlier and at larger initial scale, meaning their year-over-year growth is compounding off a larger base. Asian markets are compounding off a smaller installed base but with comparable annual capacity additions, which produces a higher percentage growth rate in the near-term forecast window.
The single-phase cooling segment is projected to grow at 26.4% CAGR and hold the largest market share by value through 2032. That tracks exactly with where engineering risk tolerance currently sits across the industry. Water-glycol based coolants are growing fastest among coolant types, and the technical reasons are straightforward.
Single-phase direct-to-chip cooling uses water or water-glycol mixtures circulated through cold plate microchannels. Coolant enters as liquid, absorbs heat from the chip surfaces via sensible heating (temperature rise with no phase change), and exits as warmer liquid. Thermal performance is a function of flow rate, supply temperature, and the cold plate's effective heat transfer coefficient. Pure deionized water delivers the highest thermal performance per unit flow: specific heat capacity around 4.18 kJ/kg·K, low viscosity, and no dissolved solids to foul microchannel passages. The problem with pure water in large deployments is operational. Freeze protection in cold-climate facilities requires careful winterization. Corrosion at dissimilar metal interfaces in manifolds, quick disconnects, and CDU heat exchangers requires continuous inhibitor management. Water-glycol solutions solve both problems at a modest thermal penalty: ethylene glycol at 30% concentration drops specific heat capacity to roughly 3.8 kJ/kg·K, an acceptable tradeoff for the operational simplicity at facilities that run at scale year-round.
Two-phase direct-to-chip systems using refrigerants can achieve superior thermal performance at the chip level. Refrigerant absorbing heat through phase change at the cold plate exploits latent heat, carrying significantly more energy per unit mass than water in sensible heating mode. The physics are compelling. The operational complexity is not. Phase change inside cold plate microchannels creates pressure dynamics that are harder to control across a manifolded loop serving multiple racks at different hydraulic distances from the CDU. Refrigerant recovery and handling under current EPA and EU F-gas regulations adds logistics and compliance overhead. Scaling a two-phase loop to 100 racks with consistent flow balance at each cold plate is a harder engineering problem than the equivalent single-phase system. The single-phase market dominance through 2032 reflects where the industry's installation experience and workforce skills currently sit.
The competitive field in direct-to-chip cooling remains fragmented despite consolidation pressure. Vertiv, Supermicro, and Modine Manufacturing hold the largest positions. DCX Liquid Cooling Systems, Schneider Electric, Koolance, and Gigabyte are active in specific sub-segments. The supply chain for CDU components is constrained, with copper microchannel fabrication, quick-connect fitting manufacturing, and pump and valve capacity all running against demand. Vendors who control more of their own supply chain are better positioned to hit the delivery windows that determine whether operators commission on schedule.
The Ecolab-CoolIT acquisition and the Trane-LiquidStack deal are examples of the consolidation wave the $17.31 billion forecast is attracting. Major industrial companies are buying their way into the direct-to-chip space rather than building organically. The rationale is consistent: thermal management is a supply chain and manufacturing problem as much as an engineering problem, and industrial companies with existing manufacturing scale have structural advantages over startups building CDU production lines from scratch.
A 26.5% CAGR through 2032 is a function of the thermal physics, not market enthusiasm. AI racks at 130 kW have no air-cooled path. NVIDIA's watt roadmap continues upward. The direct-to-chip market is being pulled by chip requirements, not pushed by vendor sales cycles.
The more useful question than "how big will this market get" is which vendors will actually ship at the volumes the $17.31 billion forecast implies. Supply chain execution at this scale is harder than the demand side of the analysis suggests. The operators who need 500 CDUs delivered on a 90-day window are the ones discovering which vendors' backlog numbers correspond to actual manufacturing capacity and which are aspirational.