The Rack Became a Boiler: Why Liquid Cooling Grows Slower Than the Heat It Removes

A GB200 NVL72 rack draws 132 kW, and its coolant comes back out at 65°C. That is hotter than the flow temperature of a domestic hot-water cylinder, and hot enough that the heat leaving a data hall can be sold rather than rejected. Two years ago the industry’s practical design ceiling was 12 kW a rack, air-cooled, with cold aisles a person could work in. That is eleven times the heat in a single product cycle — and the exhaust has become a commodity.

The industry that serves this looks, in our own numbers, oddly unexcited. Pheonix Research sizes the global data center liquid cooling market at USD 4.30 billion in 2025, reaching USD 10.20 billion by 2033 at an 11.4% CAGR, a rate that recomputes exactly from those two endpoints. Over the same eight years, heat per rack rises by roughly an order of magnitude. If the heat goes up elevenfold and the equipment market goes up 2.4-fold, then revenue per kilowatt of heat removed is collapsing.

That is not an argument that AI capital spending is about to stop. It is an argument about what is actually being sold, to whom, and at what margin — and about the difference between owning a component merchant and owning the plumber with the service contract.

The 65-degree outlet is the whole story

Vertiv’s reference design for the GB200 NVL72 supports up to 132 kW per rack in a fully liquid-cooled configuration, with coolant entering at 45°C and leaving at 65°C. Those two numbers do more work than any product brochure. At a 65°C return temperature, the chiller is optional. A dry cooler can reject the heat, and a district heating network can absorb it without a heat pump having to lift it very far — or at all.

That changes the equipment content of a data hall in one move. The classic cooling chain — air handlers, chilled water loop, chillers, cooling towers — is expensive, energy-hungry and mechanically complex. What replaces it is a cold plate bolted to a die, a manifold, a coolant distribution unit, and a dry cooler. Fewer stages, less rotating machinery, lower capital cost per kilowatt removed. A technology transition that raises the thermal stakes enormously while removing the most valuable piece of equipment from the bill of materials.

Microsoft has been the most explicit about the trade. It began rolling out a new design in August 2024 that evaporates no water for cooling, recycling water in a closed loop and avoiding more than 125 million litres a year per data centre; across its fleet it reported an average water usage effectiveness of 0.30 litres per kWh, 39% better than the 0.49 it reported for 2021. Pilots in Phoenix, Arizona and Mount Pleasant, Wisconsin were scheduled for 2026, with those sites coming online in late 2027.

And in the same note Microsoft says the quiet part out loud: replacing evaporative cooling with mechanical cooling “will increase our power usage effectiveness (PUE)”, offset only because warm chip-level cooling lets it run high-efficiency economising chillers at elevated water temperatures, producing a nominal increase in annual energy use across the fleet.

Read that twice. The largest cloud buyer in the world is deliberately making one efficiency metric worse in order to make another better — and both metrics are now on the same European reporting form.

Europe has put PUE and WUE on the same return

Directive (EU) 2023/1791, Article 26, requires Member States to ensure that data centres with a total rated energy input above 1 MW utilise their waste heat, or connect to a heating or cooling system, among other conditions. Commission Delegated Regulation (EU) 2024/1364, adopted on 14 March 2024, then defined what operators of data centres with an installed IT power demand of at least 500 kW must report to the European database, and which sustainability indicators get calculated from it: Power Usage Effectiveness, Water Usage Effectiveness, the Energy Reuse Factor and the Renewable Energy Factor.

The Energy Reuse Factor is the new term in the equation, and it is calculated as energy reused divided by total data centre energy. A facility improves its ERF by exporting heat. It improves its PUE by consuming less energy per unit of IT work, which is often exactly what a zero-water mechanical design makes harder. Two indicators, one form, opposing incentives — and a rating scheme that has to reconcile them.

The commercial consequence is the part that forecasting tends to miss. When heat is exported, the cooling system acquires a second customer: the heat network operator, the municipal utility, the industrial park. That buyer has a different budget cycle, a different contracting counterparty, and a willingness to pay that is set by the alternative cost of generating that heat, not by the cost of the cold plate. Part of the value of “cooling” is now realised outside the cooling equipment market, as heat revenue. An 11.4% CAGR measures the box. It does not measure the thermodynamic service.

The integrator is compounding at roughly three times the category

The vendor layer tells a very different story from the category forecast, and it is worth being precise about why.

Vertiv closed 2025 with fourth-quarter net sales of USD 2,880 million, up 23%, organic orders up approximately 252% year on year in the quarter, trailing twelve-month organic orders up 81%, a book-to-bill of about 2.9 times and a backlog of USD 15.0 billion, more than double the prior year. By its second quarter of 2026 it had posted USD 3,274 million of net sales, up 24% with 18% organic growth, lifted adjusted operating margin to 22.6%, up 410 basis points, raised full-year revenue guidance to USD 14.0 billion at 31% organic growth, and guided the third quarter to 34% to 36% organic growth at a 24.0% to 25.0% adjusted operating margin.

Thirty-one percent organic against our 11.4% category CAGR. I am not going to pretend these are the same measurement, because they are not: Vertiv sells power distribution, thermal management and services, it is taking share during a demand shock, and it is converting a backlog it booked eighteen months earlier. But the gap is informative. Whatever the true growth rate of the cooling category, the money in this cycle is being captured at the integration layer — packaged skids, commissioning, service contracts, multi-year capacity reservations — and not at the component layer.

The component layer is already deflating

The clearest evidence comes from the merchant that ships the most liquid cooling hardware. Shenzhen Envicool, listed on the Shenzhen exchange as 002837, reported 2025 revenue of RMB 6,067.8 million, up 32.2%, and net profit attributable to shareholders of RMB 521.9 million, up 15.3%. Its equipment-room temperature control products — the data centre business — were RMB 3,448.5 million, or 56.8% of revenue.

Net margin therefore fell from 9.86% in 2024 to 8.60% in 2025. Profit grew at less than half the rate of revenue. And this is not a marginal supplier riding someone else’s design: Envicool built the coolant distribution unit shown for Google’s Deschutes 5 design at the 2025 OCP summit, its cold plates were validated into Intel’s Xeon 6 platform, and its product set runs from cold plates and quick disconnects to manifolds, CDUs and branded long-life coolant.

So the highest-volume liquid cooling merchant, supplying the most demanding buyer in the market, grew 32% at a hardware margin and lost 1.3 points of that margin while doing it. That is what the component layer looks like once the specification stops moving and the customer starts dual-sourcing. It is also why Envicool is pushing branded fluid and leak detection: a box is a box, but a consumable is a relationship.

The real market is the air-cooled estate

Uptime Institute’s 2025 cooling systems survey, with 1,033 respondents fielded between April and June 2025, is the corrective to every new-build narrative. Twenty-two percent of operators use direct liquid cooling. Seventy-five percent still use perimeter air cooling; 32% close-coupled; 29% fresh air; 26% indirect air. Multiple answers are allowed, which is the point — most operators run a mixture.

Higher rack densities were the primary driver of direct liquid cooling adoption, cited by 69% of respondents, ahead of electricity cost at 41% and sustainability at 28%. But ask what determines whether a liquid system is actually viable and the ranking inverts into something much more mundane. Ease of retrofit into existing infrastructure came first at 46%. Lower operating costs 39%. Ease of maintenance and redundancy both 35%. Lower capital expenditure — the thing the industry talks about most — came in at 22%, ahead of only multi-sourcing and coolant availability. Respondents named lack of industry standards, cost and system-failure risk as the barriers.

Which means the binding constraint on this market is not demand from new AI halls. It is the three-quarters of the installed base that will not be rebuilt. The products that win on that basis are unglamorous: rear-door heat exchangers, in-row and side-car CDUs, manifold kits that can be bolted onto a rack someone already owns, and any supplier willing to certify a mixed air-and-liquid hall. Our own report’s segmentation already contains that bridge, labelling rear-door heat exchangers as liquid-assisted air cooling for colocation and enterprise facilities. That is not the future of the category. It is most of the current volume.

What an investor should actually be buying

Four consequences follow, and they point in different directions from the consensus.

Own the heat side of the loop. Dry coolers, heat pumps sized for the temperature lift into a district network, thermal storage, and the metering and contracts that turn exported heat into revenue. Article 26 and the Energy Reuse Factor have made waste heat a reportable output; the asset that monetises it is facility plant, not a server accessory.

Own the chemistry and the service. Long-life coolant, filtration, conductivity monitoring, leak detection, and the maintenance contract that goes with them. Recurring revenue, specified into the design, negligible capital intensity, and the one part of the chain that cannot be dual-sourced on price alone.

Be sceptical of cold plates as a standalone business. They are becoming a specified part with a standard interface, sold to integrators who are consolidating the value. Envicool’s numbers are the demonstration, not the exception.

Do not assume immersion arrives on schedule. It is in the taxonomy and it will have niches, particularly in edge and high-density clusters. But 22% liquid penetration against a 46% retrofit-ease criterion says the industry is still standardising the ordinary version of this technology. There is no volume market for the exotic one yet.

The uncertainty I would flag most loudly is inside our own forecast, not outside it. If the next platform generation takes racks to 250 or 300 kW, the kilowatt count rises faster than the box count, and revenue per kilowatt falls further. The risk to an 11.4% forecast is not that the volume disappoints. It is that the forecast is right about units and wrong about dollars.

The rack is a boiler now. Boilers are sold by people who plumb buildings, on service contracts, at low-twenties margins, to customers whose first question is whether the pipe will fit. That is a good business. It is simply not the business that the phrase “cooling technology” implies.

Note on sources: the market figures quoted above are Pheonix Research’s own, taken from the Global Data Center Liquid Cooling Market report. Company figures are from Vertiv’s fourth-quarter 2025 and second-quarter 2026 releases, Shenzhen Envicool’s 2025 annual report and Microsoft’s own datacentre design disclosures. Rack power and coolant temperatures are from Vertiv’s published reference design for the NVIDIA GB200 NVL72 platform. No third-party market-sizing estimate is used anywhere in this piece.