Article

Editorial Article

Data Center Cooling: The $40 Billion Market Air Can No Longer Cool

Updated September 13, 2026

Data Center Cooling: The $40 Billion Market Air Can No Longer Cool

There is a version of the artificial-intelligence build-out in which the binding constraint is chips. There is another in which it is electricity. The one the industry actually meets first is neither, because it arrives earlier and inside the building: heat, and the fact that a modern rack of accelerators produces more of it than the room around it was ever designed to carry.

The numbers are unusually blunt. In its 2026 update to its work on energy and AI, the IEA records that the power density of AI servers rose elevenfold between 2020 and 2025, and is set to rise by a further fourfold by 2027 — at which point a single rack, the size of a large refrigerator, could carry peak power demand equivalent to 65 households. NVIDIA’s current flagship systems make the same point in product form: the GB200 NVL72 connects 72 Blackwell GPUs into one rack and is specified as a liquid-cooled design, and the GB300 NVL72 is fully liquid-cooled, with the company presenting the shift as a route to higher compute density, less floor space and lower water consumption against air-cooled H100 infrastructure. A rack like that is not a harder cooling problem. It is a different one.

The rack sets the specification now

For most of the data centre’s history, cooling could be sized against a stable assumption and bought at the end of a project. Air moved through the room at a predictable rate; the room held a predictable number of kilowatts; the mechanical plant was a line item above the servers and below the switchgear. That assumption held because rack densities barely moved for two decades.

It has stopped holding. Air has a low heat capacity and a poor thermal conductivity relative to water, so clearing a given quantity of heat with air requires moving an enormous volume of it — through a raised floor, up through the racks, and back to a plant that has to cool it again. As density rises, the air handling stops being background infrastructure and becomes one of the largest loads in the building, with a floor-space and acoustic cost attached. At some density the arithmetic simply stops working, and no amount of engineering the room fixes it.

That is why the decision has moved forward in the design process. Cooling is no longer selected after the IT load is known; it now constrains what the IT load can be.

The market, and the one next to it

Pheonix Research sizes the global data center cooling system market — precision air conditioning in its CRAC and CRAH forms, chillers and cooling towers, economisers and free-cooling systems, and the emerging liquid and immersion technologies with the controls that tie them together — at USD 18.7 billion in 2025, rising to USD 39.6 billion by 2033 at a 9.8% CAGR. That is a 2.12x expansion, or 111.8% across the eight-year window.

The comparison worth making is with the article immediately before this one. Pheonix puts the global power supply market at USD 42.6 billion in 2025 growing to USD 70.8 billion by 2033 at 6.7% — the slowest market in the electrification chain. Cooling is therefore worth 43.9% of the power supply market in 2025 and 55.9% of it by 2033: a smaller market, but a faster-growing one, closing the gap because the same load is pulling both. The relationship is not a coincidence. Conversion losses become heat. Every point of efficiency the power architecture gives back appears in the thermal budget, and every point the thermal architecture gives back appears in the power bill. It is one number approached from two ends.

Cooling stops being an efficiency line item

The clearest evidence that the relationship has changed is what has happened to Power Usage Effectiveness, the ratio of total facility power to the power that reaches the IT load. Uptime Institute’s Global Data Center Survey 2026, published in July, found that average PUE is still edging down but that the gains “remain gradual” — with legacy infrastructure and cooling constraints slowing progress — while a growing number of operators now report peak rack densities of 30 kW or above. Power, meanwhile, remains the leading cause of impactful outages in the survey, and high cost remains the industry’s single greatest management concern.

The timing is awkward. Fifteen years of engineering has gone into shaving fractions off PUE through better airflow management, containment, economisers and controls — and just as that work matures, the load inside the building changes shape. A facility optimised for 8 kW racks does not become efficient at 80 kW; it becomes a rebuild.

The IEA’s estimate of how much is at stake is worth holding onto. Cooling accounts for roughly 7% of electricity consumption in an efficient hyperscale data centre and more than 30% in a less efficient enterprise facility. In other words, cooling is somewhere between a fourteenth and a third of the entire electricity bill, depending on how the building was designed. That makes it the largest single lever an operator has when power prices move — and the hardest one to pull once the concrete is poured.

Behind it, a load that is still doubling

None of this would matter if the underlying demand were flat. It is not. The IEA’s updated assessment puts global data centre electricity consumption at 485 TWh in 2025, roughly doubling to around 950 TWh by 2030 — about 3% of world electricity. Data centre demand grew 17% in 2025, in line with the agency’s projections, and consumption at AI-focused facilities grew 50%, while capital expenditure by the largest technology companies exceeded USD 400 billion in 2025 and is expected to rise by another 75% in 2026.

It is worth noting how the agency’s own numbers moved. The 2025 Energy and AI report put data centres at about 415 TWh in 2024 heading for some 945 TWh by 2030. The 2026 update lifts that to 485 TWh in 2025 and 950 TWh in 2030. The starting point moved considerably faster than the destination — which is another way of saying the industry has been arriving at its 2030 position earlier than expected, not that the destination has changed.

The liquid transition is a component transition

If density sets the specification, the specification is now liquid. Single-phase direct-to-chip cooling — cold plates attached to the hottest components, with a pumped coolant loop carrying heat to a facility-level heat rejection system — has become the default architecture for new AI clusters, with two-phase systems and immersion deployments following selectively as chip-level thermal densities rise.

The important consequence is that the growth sits in a different supply chain from the one that built the market. Precision air conditioning is a mature, standardised equipment business sold as installed capacity. Direct-to-chip and immersion are a components business — cold plates, coolant distribution units, quick-disconnects, manifolds, pumps, fluids and the controls that schedule them — with shorter qualification cycles, a much smaller installed base, and a design that is still being settled. Pheonix sizes the data center liquid cooling market separately at USD 4.3 billion in 2025 rising to USD 10.2 billion by 2033 at an 11.4% CAGR — a 2.37x expansion against 2.12x for cooling as a whole. That puts liquid at roughly 23% of the total cooling market today and about 26% by 2033: a quarter of the market growing faster than the three-quarters around it, and the quarter where the architecture is not yet fixed. Direct-to-chip leads that sub-market, on the pull of AI workloads and GPU clusters.

Where the spend lands, and who takes it

The supplier list explains how this market is structured, and it is not a list of cooling companies. Pheonix identifies Schneider Electric and Vertiv Group as the global leaders, alongside STULZ, Rittal, Johnson Controls, Daikin Industries, Mitsubishi Electric, Huawei Digital Power, and — from the liquid side — Asetek and Submer Technologies, with CoolIT Systems and Green Revolution Cooling among the immersion and direct-to-chip specialists. Read that list by origin and the convergence becomes obvious: Daikin, Mitsubishi Electric and Johnson Controls come from building HVAC, Schneider Electric from electrical distribution and power infrastructure, Vertiv from data-centre IT infrastructure, Asetek and Submer from electronics and server cooling. Four industries have arrived at the same rack.

The segment split is the other thing to hold onto. Precision air conditioning, in its CRAC and CRAH forms, still leads the market on the strength of the installed base across hyperscale and enterprise facilities — accumulated floor area that must be cooled whether or not it was designed for the density now arriving. Geographically, North America is the largest market on hyperscale cloud and AI investment; Asia-Pacific is the fastest-growing, on data-centre construction in China, India and Southeast Asia, and is also the fastest-growing region for liquid.

The two facts sit awkwardly together. The revenue is in air, sold as capacity. The growth is in liquid, sold as components. And the incumbents that dominate the first are, for now, the same names carrying the second.

A high risk rating, for unremarkable reasons

Pheonix rates data centre cooling’s overall risk high, with geopolitical exposure and substitution risk both moderate — a flatter profile than stationary storage, and flatter than the liquid sub-market beside it. The technology is rated mature, with high innovation intensity and high patent activity; the market is fragmented with low competitive intensity, high capital intensity and rising M&A activity; the operational model is vertically integrated, and supply-chain complexity is high.

The regulation line is the one to read carefully. Complexity is rated high, even though the approval pathway is a standardised commercial one. That is unusual: across the electrification chain, complexity and pathway tend to move together. Here they do not, because the regulatory surface is not market access but the two things mechanical cooling consumes — refrigerant and water. That is a slower, more local form of regulatory risk than an export control, and it lands on operating cost rather than on a supply chain.

The liquid sub-market carries a different profile, and it is the profile of something early. Pheonix rates it high overall risk with high geopolitical exposure, an emerging technology stage, high competitive intensity across eight tier-one players and a fragmented structure — against the mature, low-intensity air layer it is displacing. Whatever else changes by 2033, that asymmetry is worth remembering: the part of cooling that is growing carries the risk, and the part that carries the revenue is the part being replaced.

What to watch

Three things decide whether this is a 9.8% market or a faster one.

The first is rack density crossing the air ceiling. The industry’s tail — racks at 30 kW and above — is now the part of the fleet being built, and the tools that made 8 kW racks efficient were not designed to carry 80 kW. Whether operators retrofit liquid into existing halls or build only liquid-ready shells determines how much of this market is new-build revenue and how much is rework.

The second is water and refrigerant regulation. Cooling is the one part of the AI stack with a physical externality attached — evaporative water consumption and refrigerant handling — and it is regulated locally, which means the constraint arrives market by market rather than all at once.

The third is whether the incumbents convert. The names at the top of the air-cooling market are also the names in liquid. If direct-to-chip and immersion growth is captured by the existing vendors, this stays a steadily compounding market. If it is captured by the electronics-cooling specialists arriving from server design, the market grows just as fast and its leadership changes underneath it.

That is the argument for reading cooling beside the article before this one. Conversion losses become heat; heat has to be removed; and the efficiency the power architecture gives back is the thermal budget the cooling system then has to carry. Power supply and cooling are one number approached from two ends, and the end with the physics ceiling is the end that will set the limit.

Sources and further reading

Pheonix Research market intelligence:

Primary and reference sources:

  • IEA, Energy and AI (2025) and the 2026 update — data-centre electricity consumption (about 415 TWh in 2024 heading to roughly 945 TWh by 2030; updated to 485 TWh in 2025 and about 950 TWh by 2030), AI server power density, the 65-household rack comparison, and cooling’s share of facility electricity.
  • Uptime Institute, Global Data Center Survey 2026 — average PUE trend, rack densities at 30 kW and above, outage causes and management concerns.
  • NVIDIA, GB200 NVL72 and GB300 NVL72 platform documentation — liquid-cooled rack reference designs and the density, floor-space and water arguments made for liquid over air.

 

Related Market Reports

Frequently Asked Questions

How big is the data center cooling system market?

Pheonix Research sizes it at USD 18.7 billion in 2025, rising to USD 39.6 billion by 2033 at a 9.8% CAGR — a 2.12x expansion, or 111.8%, across the eight-year forecast. The market covers precision air conditioning (CRAC/CRAH), chillers and cooling towers, economisers and free-cooling systems, and liquid and immersion technologies with the controls that tie them together.

Why is cooling growing faster than the rest of the data-centre supply chain?

Because it is being pulled by rack density rather than by unit volumes. Pheonix puts the power supply market at a 6.7% CAGR and data centre management at 9.1%, against 9.8% for cooling and 11.4% for liquid cooling. The IEA records AI server power density rising elevenfold between 2020 and 2025, with a further fourfold rise projected by 2027, and cooling is the system that has to absorb that.

What is the difference between air and liquid cooling, and why is the industry switching?

Air has low heat capacity and poor thermal conductivity relative to water, so clearing a given amount of heat with air means moving a very large volume of it, at a floor-space and acoustic cost that grows with density. Liquid — direct-to-chip cold plates, two-phase systems and immersion — removes heat at the source. Pheonix's liquid cooling report, covering the same segment, is sized at USD 4.3 billion in 2025 rising to USD 10.2 billion by 2033 at an 11.4% CAGR, about 23% of the total cooling market in 2025 and about 26% by 2033. Direct-to-chip leads that sub-market, on AI and GPU cluster deployments.

Which regions lead, and who are the main suppliers?

North America is the largest market, on hyperscale cloud and AI infrastructure investment; Asia-Pacific is the fastest-growing, on data-centre construction in China, India and Southeast Asia, and is also the fastest-growing region for liquid cooling. Pheonix names Schneider Electric and Vertiv Group as global leaders, alongside STULZ, Rittal, Johnson Controls, Daikin Industries, Mitsubishi Electric and Huawei Digital Power, with Asetek, Submer Technologies, CoolIT Systems and Green Revolution Cooling leading on immersion and direct-to-chip.