Liquid Cooling Crosses Hyperscaler Threshold, $4.75B CoolIT Deal Proves

On 3 July 2026, Ecolab closed its acquisition of Calgary-based CoolIT Systems for $4.75 billion, folding the specialist direct-to-chip cooling vendor into a global water-treatment and industrial-services giant that now projects its high-tech division will reach $4 billion in annual revenue by 2030. The deal, confirmed in a company statement and reported by Seeking Alpha, closed months ahead of the original schedule. For anyone tracking the physical infrastructure of compute, the acceleration of the timetable was as significant as the price: CoolIT's order book had swollen faster than either party anticipated when the acquisition was first announced.

CoolIT builds coldplates. More precisely, it builds the copper-and-microchannel assemblies that bolt directly onto a GPU or CPU, pump a dielectric fluid through a closed loop, and pull heat away at the chip surface before it ever reaches the surrounding air. A single CoolIT coldplate announced on 1 June 2026 is rated for 15 kW of heat removal, according to Insider Monkey. That figure, 15 kilowatts per socket, is the number that makes the entire air-cooling paradigm look like a quaint holdover. A standard air-cooled rack might manage 15 to 20 kW total. The CoolIT unit handles that at a single chip, and a rack filled with such hardware can push past 100 kW without the data centre operator needing to redesign the airflow of an entire hall.

The thermal wall that air cooling hit was not a gradual incline. It was a step function, and the step arrived with NVIDIA's Blackwell GPU generation. Blackwell's B200 module draws roughly 1,200 watts in its maximum configuration. A rack of 72 such GPUs, the standard GB200 NVL72 configuration, pulls approximately 120 kW. Air cooling a 120 kW rack requires volumes of forced air that are physically impractical in most existing colocation halls: the floor-to-ceiling pressure differential needed to move that much air through perforated tiles exceeds what standard raised-floor designs can sustain without structural modification. The industry's engineering consensus, reflected in Omdia's mid-2026 analysis of AI infrastructure, is that the air-cooling era for high-density GPU clusters is effectively over.

The market research firm Omdia, in a report covered by Agence France-Presse on 16 July 2026, framed the shift as one of five dynamics redefining AI infrastructure this year. Omdia now classifies the emerging buildout as an "AI Factory" market, distinct from the general-purpose data centre market, and the defining physical characteristic of an AI Factory is that it is designed from the slab up for direct liquid cooling. The traditional measure of data centre capacity, square footage of raised floor, is being supplanted by a new metric: megawatts of IT load per hall, with the cooling system sized to match at a coefficient of performance that air-cooled designs cannot approach.

NVIDIA itself made this explicit in late June 2026 when it unveiled a reference data centre architecture that replaces evaporative cooling towers with a fully closed-loop liquid system. Gizmodo reported the announcement under a headline that captured the sales pitch: a "100% reduction in water use." The design routes heat from the chip-level coldplate loop to a secondary liquid loop, then to a dry cooler or adiabatic system that rejects heat to the atmosphere without consuming water for evaporation. Cooling, 24/7 Wall St noted, "consumes roughly 40% of a data centre's energy budget." Eliminating the evaporative step changes the arithmetic of where a facility can be sited and how quickly it can be permitted.

The asterisk on NVIDIA's water-use claim, as TechTimes pointed out within days, is that the closed-loop design addresses only the facility-level cooling load. It does nothing about the water consumed upstream, in the generation of the electricity that powers the GPUs. A gas-fired power plant consumes roughly 500 to 700 gallons of water per megawatt-hour for cooling its own turbines. A 100 MW data centre running at 85 percent utilisation will require something on the order of 745,000 MWh annually. The water consumed at the power plant, depending on the generation mix, can easily exceed the water the data centre would have used with evaporative cooling on site. This is not an argument against closed-loop liquid cooling; it is a reminder that the boundary drawn around a sustainability claim matters enormously.

The choice confronting hyperscaler engineering teams is not whether to adopt liquid cooling but which architecture to standardise on, and here the industry has split into two camps. Direct-to-chip, or D2C, keeps the server form factor largely intact: coldplates attach to the primary heat-producing components, a coolant distribution unit (CDU) manages the secondary loop, and the rest of the server remains air-cooled. Immersion cooling submerges the entire server in a dielectric fluid, eliminating air cooling entirely but requiring a wholly different rack design, different service procedures, and different assumptions about fluid compatibility with every component in the bill of materials.

D2C is winning the near-term deployment race for a straightforward reason: it can be retrofitted. A colocation provider with existing air-cooled halls can install CDUs at the end of each row, run secondary-loop piping overhead, and convert a portion of its floor to support 50 kW-plus racks without gutting the building. CoolIT built its business on this retrofit proposition, and the $4.75 billion Ecolab paid reflects the size of the addressable market: every air-cooled data centre hosting GPU clusters is a candidate. Immersion, by contrast, demands purpose-built facilities. Shell, the oil major, has been running immersion-cooled high-performance computing clusters for years at its technology centre in Amsterdam, but the format has not yet crossed into hyperscale mass deployment the way D2C has.

MarketBeat, in a 26 June analysis titled "Liquid Gold: The AI Cooling Retrofit Trade," framed the retrofit as a distinct investable theme, separate from new-build data centre construction. The argument is that the existing global stock of roughly 8,000 operational data centres, many of them built in the 2010s for 5 to 10 kW per rack, represents a stranded asset unless they can be upgraded for AI workloads. A D2C retrofit costs somewhere between $2 million and $5 million per megawatt of IT load, depending on the density target and the state of the existing power distribution. That is a fraction of the cost of building new, and the schedule is measured in months rather than years. For a hyperscaler trying to deploy Blackwell clusters in 2026 rather than 2029, the retrofit path is often the only path.

The CoolIT acquisition also signals something about the cooling supply chain that was not obvious two years ago: the hyperscalers want a single throat to choke. A data centre operator commissioning a 200 MW campus does not want to negotiate coldplate procurement with one vendor, CDU procurement with a second, secondary-loop chemistry management with a third, and field-service contracts with a fourth. Ecolab's pitch, post-acquisition, is that it can deliver the coldplates, the CDUs, the water-treatment chemistry, the monitoring software, and the global field-service network as a single integrated package. Christophe Beck, Ecolab's chief executive, has described the strategy as "full-stack" cooling, though that phrase has not appeared verbatim in the publicly available announcements reviewed for this article.

The geographical implications of the shift are only beginning to surface in utility commission dockets and county planning hearings. A data centre that does not consume water for evaporative cooling can be sited in places that would have been politically impossible five years ago. Fort Meade, Florida, provides a live case study. A developer proposing a hyperscale facility there told The Ledger in June 2026 that the project would use a closed-loop system requiring only 50,000 gallons of water daily, a figure an academic expert on industrial cooling questioned as unrealistically low for a facility of its proposed size. Whether the number is accurate or aspirational, the developer's instinct to lead with the water figure tells its own story: liquid cooling, somewhat paradoxically, is being sold as the technology that makes data centres less thirsty, and municipalities are listening.

The power side of the equation is where the cooling conversation gets harder. A 120 kW rack, whether cooled by air, D2C, or immersion, still draws 120 kW. Liquid cooling improves the power usage effectiveness (PUE) ratio by eliminating fan energy, typically shaving 0.05 to 0.10 off a facility's PUE compared with an air-cooled design of equivalent density. But the absolute power draw of an AI cluster is rising faster than the efficiency gains can offset. TechRepublic, citing Omdia's research, reported in mid-June that many data centres "struggle to get enough power into the facility to run AI applications," and those that manage it face a secondary challenge: the local substation may not have the headroom to support the connected load.

The distinction between connected load and contracted load is one that utility planners in northern Virginia, Phoenix, and County Meath, Ireland, now live inside every day. Connected load is what the data centre is physically wired to draw at peak; contracted load is what the utility has agreed, in a legally binding interconnection agreement, to deliver. A hyperscaler can install liquid cooling and raise its rack density from 15 kW to 120 kW, but if the utility agreement was struck at 40 MW of contracted load for the campus, the new racks cannot all be powered simultaneously without renegotiating the interconnection. That renegotiation, in constrained markets like Dominion Energy's northern Virginia territory, now carries timelines of five to seven years.

CoolIT's 15 kW coldplate, impressive as an engineering achievement, is thus also a statement about what the market expects the power delivery chain to be able to support in the medium term. A coldplate rated higher than 15 kW is certainly buildable; the question is whether the transformers, switchgear, and feeder cables upstream can deliver the corresponding current at a cost that makes the deployment economic. The 15 kW figure appears calibrated to match what a standard 415-volt three-phase rack power distribution unit can deliver without stepping up to medium-voltage rack-level distribution, a leap that no hyperscaler has yet standardised at scale.

Immersion cooling advocates argue that their approach solves a problem D2C cannot: the thermal management of components beyond the primary chips. In a D2C system, voltage regulators, memory modules, network interface cards, and storage drives still rely on air cooling. As these components themselves grow more power-dense in AI configurations, the residual air-cooling load becomes non-trivial. Immersion eliminates the distinction by making every component a liquid-cooled component. The counterargument, pressed by D2C vendors and their hyperscaler customers, is that immersion introduces fluid compatibility risks across a supply chain of thousands of component vendors, most of whom have never tested their products for long-term dielectric fluid exposure.

The component-compatibility question is not theoretical. A single plasticiser leaching from a cable jacket into a dielectric fluid can, over months, alter the fluid's breakdown voltage enough to create a short-circuit risk. The testing regime required to certify an entire server bill of materials for immersion is extensive and expensive, and the liability for a failure is unclear when the server vendor, the fluid vendor, and the tank vendor are three different companies. D2C sidesteps nearly all of this: the coldplate loop is sealed, the dielectric fluid never contacts the electronics, and the component vendors do not need to change their qualification processes.

Ecolab's $4 billion revenue target for 2030, stated in the CoolIT acquisition announcement, implies a compound annual growth rate for its high-tech cooling business that substantially exceeds the broader data centre market's growth. That target embeds an assumption: that liquid cooling will move from a specialised solution deployed on perhaps 15 to 20 percent of new data centre IT load today to something approaching the default for any facility hosting accelerated computing. Whether the actual penetration reaches 60 percent, 80 percent, or somewhere lower depends on factors that the cooling industry does not control: the pace of GPU deployments, utility interconnection timelines, and the silicon vendors' own thermal design power roadmaps.

The next checkpoint comes in the fourth quarter of 2026, when Ecolab reports its first full quarter with CoolIT integrated and the hyperscalers disclose their 2027 capital expenditure plans. The Q4 earnings calls will be the first public forum in which the liquid-cooling adoption rate is visible in the financials of a major publicly traded vendor rather than inferred from supplier order books and consultant estimates. The number to watch is not the revenue line but the backlog: how many quarters of CoolIT production are already sold before the factory can build them. If that number stretches past four quarters, the supply constraint, not the technology, becomes the story.