Cooling Is the Compute: How Heat Became the Data Center's Real Ceiling

AI racks now concentrate more electrical power into less floor space than the buildings housing them were designed to carry. The consequence is that the data center’s decisive technology has moved off the processor and onto the power and thermal plant beneath it — and the money is following. Pheonix Research puts the combined cooling and management pool at $35.8 billion in 2025, on a path to $75.4 billion by 2033.

A data center used to be described by how many servers it held. It is now described by how many watts it can get in, and how many it can get out again. Those two numbers — supply and rejection — have become the definition of capacity, and the technologies that move them have become the center of the industry’s spending.

The load is not subtle. The world’s data centers consumed roughly 415 terawatt-hours of electricity in 2024, about 1.5% of global consumption, a figure the International Energy Agency says grew 12% a year over the preceding five years. It then accelerated: the IEA’s follow-up analysis puts data-center electricity demand up 17% in 2025, with consumption from AI-focused facilities rising 50% in the same year. Grown at that 17%, the 2024 level implies close to 486 TWh in 2025 — a Pheonix Research estimate built from the two IEA figures (415 TWh × 1.17).

The density problem

What has changed is not only how much power the sector draws, but how tightly it draws it. Schneider Electric, the power and cooling vendor, puts average rack density at 16 kW in 2025 rising to 27 kW in 2026 — a 68.8% jump in a single year, and a Pheonix Research calculation from the two figures Schneider publishes (27 ÷ 16 − 1). Averages understate the frontier: the same source notes that the newest AI systems can reach 246 kW per rack, the number it attaches to NVIDIA’s Vera Rubin generation, and that only about one operator in five considers itself ready for the 50–70 kW racks now common in AI deployments.

Air is not a technology that scales to those numbers. Moving heat with air requires volume and distance; the denser the compute, the more of both, until the mechanical plant consumes the very capacity it was installed to protect. The fix is to shorten the path between the heat and the coolant — which is why the industry’s thermal spending is shifting from air handling to liquids plumbed directly to the silicon.

Cooling stops being an operating cost

Pheonix Research’s Global Data Center Liquid Cooling Market report sizes the liquid segment at $4.30 billion in 2025, reaching $10.20 billion by 2033 at an 11.40% CAGR — a market that multiplies 2.37 times over the window, a Pheonix Research calculation ($10.20bn ÷ $4.30bn). Direct-to-chip cold-plate systems, in which coolant is plumbed to the processor itself, lead it; immersion cooling follows in the densest AI and high-performance computing builds, where operators want thermal efficiency and less dependence on large air-handling systems.

That liquid pool sits inside a larger thermal market. Pheonix Research’s Global Data Center Cooling System Market report puts total cooling at $18.7 billion in 2025 and $39.6 billion in 2033, a 9.8% CAGR. Liquid cooling is therefore 23.0% of the cooling market today — a Pheonix Research calculation ($4.30bn ÷ $18.70bn) — and it is the faster-growing part: the liquid slice multiplies 2.37× by 2033 against 2.12× for cooling as a whole, both Pheonix Research calculations. The vendors assembling around it are a mix of incumbent thermal suppliers and specialists — Schneider Electric, Vertiv, CoolIT Systems, Asetek, Rittal and STULZ alongside Submer and Green Revolution Cooling.

The constraint that outran the cooling

Heat is the argument the industry has learned to have. Power is the harder one. Global data-center electricity demand is projected to reach roughly 132 GW in 2026 and 290 GW by 2030, on Schneider Electric’s figures, and the bottleneck is no longer the generator but the connection: in several major US markets, securing new grid capacity now takes three to four years, longer than building the facility it will feed. The IEA reaches the same wall from the other side. It finds that behind-the-meter natural gas could power 15–27 GW of data-center load by 2030, mostly in the United States, but that doing so reliably requires overbuilding on-site generation by 30% to 70% — and that most operators would still rather connect to the grid.

This is where liquid cooling stops looking like a thermal choice and starts looking like a capacity lever. Legacy air-cooled halls commonly run at a power usage effectiveness (PUE) of 1.55–1.67 on the vendor’s figures, meaning roughly a third of the power entering the building is spent on cooling and ancillary systems rather than on compute. Bring coolant to the chip and that overhead falls, which converts directly into compute delivered per megawatt of interconnection — the scarce commodity in the AI build-out.

The layer that sells the capacity back

The third part of the stack is the least physical and among the fastest to change: the software that decides where the load goes. Pheonix Research’s Global Data Center Management Market report sizes the market built around data center infrastructure management (DCIM) at $12.8 billion in 2025, rising to $25.6 billion by 2033 at a 9.1% CAGR, with asset tracking, power optimisation, cooling control and capacity planning as the leading use cases. Once power is the binding constraint, the value of a management layer is measured in the watts it recovers — by capping, shifting and packing workloads into capacity that already exists. Schneider Electric, Vertiv, Siemens, ABB, Eaton, IBM, Cisco and Huawei sit on the incumbent side of that market, with Nlyte and Sunbird among the software specialists.

One pool, and it is not compute

Add the three — liquid cooling, cooling systems and management — and the physical layer of the data center is $35.8 billion in 2025 moving to $75.4 billion by 2033. That is a 2.11× expansion at a 9.8% CAGR, both Pheonix Research calculations ($75.4bn ÷ $35.8bn across the eight-year window). Not one of those figures is the server market; they are what it costs to make a building capable of running one.

Which is the point. For two decades the data center’s technology story was told through the processor. It is now told through the cold plate, the coolant distribution unit and the substation, because those are what decide how much silicon a site can actually run. The industry built its reputation on adding compute; it is being repriced on its ability to power and cool it. The number to watch is not a chip’s throughput but the liquid share of the thermal market — 23.0% today on Pheonix Research’s figures — and whether it keeps taking share from air.

Pheonix Research covers this cluster in its Data Center Infrastructure segment, including the Global Data Center Liquid Cooling Market, the Global Data Center Cooling System Market and the Global Data Center Management Market outlooks to 2033.