Liquid Cooling at Hyperscaler Scale Moves From Pilot to Procurement

On September 3, 2026, LITEON Technology, the Taiwanese power component maker, said it would pay about $176 million for a 25 percent stake in DCX Liquid Cooling Systems, a Warsaw-based specialist in coolant distribution and cold plates. Unite.AI reported the investment as a move to fold liquid cooling into LITEON's integrated power management business for AI infrastructure. The number is not giant by hyperscaler standards. It is significant because a power supply vendor, not a datacenter operator, chose to own part of the thermal path between the chip and the cooling tower.

That choice has become ordinary in 2026. Vertiv reported second-quarter sales growth of 24 percent across the Americas and Asia-Pacific on AI datacenter demand, according to its Q2 2026 earnings transcript. The driver of that demand is no longer a fan wall and a CRAC unit. It is direct-to-chip cold plates, immersion baths, and coolant distribution units sized for hundreds of kilowatts per rack. A July forecast from ResearchAndMarkets listed Vertiv, CoolIT Systems, Submer Technologies, GRC, and Asetek as the vendors driving the liquid cooling infrastructure products category toward the end of the decade, Yahoo Finance reported.

The physics behind the shift is straightforward. A GPU rack that pulls 120 or 150 kilowatts of electrical power cannot reject that heat into air without an implausibly large volume of forced-air movement. Nvidia's Vera Rubin NVL72 racks, priced as high as $8.8 million per unit in some configurations, are built for liquid-cooled operation from the factory, Tom's Hardware reported in March. Once the chip's thermal design power crosses a threshold, the architecture question stops being whether to add water and becomes which water path to choose.

Direct-to-chip, or D2C, puts cold plates over the processors and pumps a dielectric or treated water-glycol mix across them. Single-phase systems keep the liquid as a liquid. Two-phase systems let it boil at the cold plate and carry heat away as vapour to a condenser. Immersion cooling goes further, submerging entire boards or sleds in a dielectric fluid. Both approaches have the same purpose: move heat away close to the silicon so the air side of the facility no longer has to do the heaviest lifting.

The vendor claims now carry test data. Accelsius, a two-phase direct-to-chip specialist, published independent benchmarks on July 21 showing its system held NVIDIA B200 junction temperatures 9 degrees Celsius lower than a single-phase alternative under warm-water conditions. The result matters because warm-water cooling promises to reduce or eliminate the energy spent on compressors and chillers, and a lower junction temperature also buys margin for higher sustained clock speeds. The company announced the figures through Business Wire. Benchmarks commissioned by a vendor cannot be read as neutral science, but the 9-degree delta is exactly the kind of number facilities engineers now want from any supplier.

The step from benchmark to deployment is the harder one. EdgeConneX, the global datacenter operator, is placing Accelsius's IR150 integrated rack plus coolant distribution unit at its global training center, Facilities Dive reported on August 27. The location matters: before two-phase systems go into live customer halls, the technicians who will maintain them need a reference platform. The IR150 is described as the industry's first fully integrated rack plus CDU. That descriptor signals the direction of procurement. Buyers increasingly want one part number that combines rack, cold plates, and coolant distribution, not a kit of parts from four vendors.

The same logic appears in closed-loop cooling experiments at the utility edge. In July, Reuters reported that Valar, a nuclear startup, had partnered with Nvidia on a Utah data center demonstration pairing a microreactor with Blackwell chip architecture. The companies said the combination of closed-loop cooling and the reactor could cut water use to near zero. It is an extreme case, but it names the constraint that hyperscalers now route around: water is no longer an afterthought in a permit application.

The water constraint arrives

In August, the permit pressure became explicit. A single appeal by a private citizen stopped Google from drawing water from a Belgian river to cool an AI data center in Wallonia, MSN reported, citing the regional decision over drought conditions. A private citizen over a drought. This is the new siting risk for a hyperscale build, and it lands most directly on the cooling system. Air-cooled halls with evaporative assistance consume water at the point where they reject heat. A direct-to-chip loop, if it runs closed, can shift that rejection to a dry cooler or a higher-temperature heat exchanger.

The scale of the problem makes the shift urgent rather than optional. Global data centre cooling-water consumption could nearly triple by 2030 as AI capacity expands, OilPrice.com reported on 29 August, citing industry forecasts. A Euronews report on September 8 noted that tech giants point to new cooling systems as the way to curb demand while local opposition in the United States grows. The contradiction is inherent: a data centre designed for AI chips can save water per computation yet still use more water in aggregate because the campus itself is larger.

Direct-to-chip is not automatically a water-saving technology. Single-phase D2C often ends in an evaporative cooling tower, which still consumes water, though the loop can run warmer than a chilled-water IT hall. Two-phase direct-to-chip moves heat through phase change and can use dry coolers at higher ambient temperatures. Immersion can do the same but brings its own chemistry, weight, and maintenance questions. The important detail is the water source and the heat rejection path, not the marketing category on the specification sheet.

Corporate capital has followed that distinction. Legrand announced in February a strategic investment in Accelsius alongside its acquisition of Kratos Industries, Yahoo Finance reported, citing a company release. Trane Technologies completed its acquisition of immersion-cooling vendor LiquidStack in March, Simply Wall St reported. LITEON's September move into DCX is the same pattern from the power side. The supply chain is being pulled together by buyers who do not want to manage seven vendor debates in the middle of a rack deployment.

The capacity question now follows the coolant distribution unit. On this date a year earlier, LiquidStack had brought out a modular CDU it said could scale to hundreds of megawatts of cooling power for hyperscale builds, Facilities Dive reported. A year later, the relevant specification is no longer the rack alone. It is whether the CDU can be installed as a modular system, whether it can be serviced without taking down the entire hall, and whether the facility's electrical switchgear can supply the pumps and controls at the same grade of reliability as the servers.

In South Korea, the question has moved from paper to site. LG CNS said in August that it would deploy next-generation liquid cooling infrastructure for Nvidia's Vera Rubin at the Samsong Data Center in Goyang, near Seoul, with Naver Cloud, MSN reported. Vera Rubin is the platform where rack power and thermal density make direct-to-chip or immersion a condition of installation, not an optional upgrade. A full rack of Rubin systems priced at $8.8 million will not be parked in an air-cooled hall and left to chance.

What to watch for

The next argument is about the fluid itself. New cooling chemicals introduced for single-phase loops and immersion have drawn scrutiny for perfluoroalkyl and polyfluoroalkyl substances, the PFAS family, WRAL reported in August. The chemistry that solved the heat-transfer problem may create a different permitting and liability problem. Vendors are now being asked to prove what is in the fluid, where it goes when the loop leaks, and what happens at decommissioning.

That question will shape the next generation of deals. The hyperscaler is not buying cooling in isolation. It is buying a path through the local water board, the chemical regulations, the substation schedule, and the insurance policy. In 2026, the cooling instrument is also a siting instrument.

The checkpoint to watch is the Vera Rubin deployment wave of 2026 and 2027, when the ratio of megawatts committed to the grid per rack will force a decision on which technologies get to be called standard in the first fully liquid-cooled generation of AI capacity. The schedule, more than any vendor roadmap, will provide the answer.