Data Center Cooling Systems: From CRAC Units to Immersion

Data center cooling removes the heat that IT equipment generates so servers stay within their operating range, and it typically consumes 30-40% of a facility’s total energy bill — second only to the IT load itself. The technology stack spans a spectrum: CRAC and CRAH air handlers for conventional loads, free cooling and evaporative systems to cut mechanical refrigeration, and direct-to-chip or immersion liquid cooling for the 60-130+ kW racks that AI training now requires. Which one a facility needs depends on rack density, climate, and how much retrofit budget is available.

Key takeaways

• Cooling is 30-40% of facility energy use in a conventional air-cooled data center, falling to 7-10% in the most efficient hyperscale facilities, per IEA and industry benchmarking.
• Industry-average PUE is 1.52 (Uptime Institute Global Data Center Survey 2026), barely moving in recent years because older facilities dilute a long tail of new, highly efficient builds running 1.2-1.3.
• CRAC units cost $15,000-50,000 per replacement; CRAH units cost $25,000-75,000 per unit plus a shared chiller plant — CRAH scales more efficiently but needs more upfront infrastructure.
• Free cooling cuts cooling energy 30-50% and chilled-water production cost up to 70% when ambient conditions allow it — but it barely works in tropical colocation hubs like Singapore or Bangkok.
• Air cooling tops out near 30-40 kW per rack. Above that, direct-to-chip or immersion liquid cooling becomes standard; NVIDIA’s GB200 NVL72 draws 120-132 kW and ships liquid-cooled by default.
• Retrofitting for liquid cooling runs about $2M/MW in an existing colocation facility versus $11M+/MW for a new liquid-ready greenfield build, per STL Partners.
• The global cooling market is roughly $13-31 billion in 2026, growing 12-17% annually; the liquid-cooling slice alone is forecast to grow at 31.5% CAGR through 2033 as AI density outpaces air.

For live pricing by market, see the colocation price index and browse cooling infrastructure across 2,900+ cataloged facilities.

How data center cooling works, in three layers

Every cooling system does the same job at a different point in the chain: pull heat off the chip, move it out of the room, and reject it outside the building.

1. Chip-to-air or chip-to-liquid. Heat leaves the processor either into ambient air pulled through a heatsink and fan, or into a coolant loop via a cold plate (direct-to-chip) or a dielectric bath (immersion).
2. Room-level handling. CRAC and CRAH units pull hot air from the room (or hot aisle), cool it, and return it to the cold aisle or underfloor plenum.
3. Heat rejection. The facility dumps the heat outside via a chiller plant, cooling tower, dry cooler, or — during free-cooling hours — directly into cool outside air.

Most facilities layer several of these: air-cooled racks feeding a CRAH system backed by a chiller plant with a free-cooling economizer bypass, plus an isolated liquid loop for the highest-density racks.

CRAC vs. CRAH: the room-level choice

Feature	CRAC (Computer Room Air Conditioner)	CRAH (Computer Room Air Handler)
Cooling source	Built-in compressor, direct refrigerant expansion	Chilled water from a central chiller plant
Typical unit cost	$15,000-50,000 (replacement, 30-200 kW capacity)	$25,000-75,000, plus shared chiller infrastructure
Efficiency at scale	Lower — each unit runs its own compressor	Higher — central chiller plants scale better
Best fit	Small/mid facilities, retrofits, standalone rooms	Larger facilities already running a chilled-water loop
Redundancy model	Unit-level N+1	Shared plant redundancy, often more cost-efficient at scale

CRAC and CRAH units together still cover about half of the installed cooling base worldwide — roughly 52% of the data center cooling market by technology share in 2025, per market research cited in vendor cost breakdowns. Neither is obsolete; they remain the standard for the large majority of enterprise and retail colocation halls running standard 5-15 kW racks. See our guide on power density from 5 kW to 130 kW per rack for where each technology stops working.

Free cooling and economization

Free cooling (economization) bypasses mechanical refrigeration whenever outside air or water is cool enough to do the job directly:

• Air-side economizers filter and bring in outside air directly, sometimes mixed with return air, cutting compressor run time.
• Water-side economizers use a cooling tower to chill water via evaporation instead of a compressor, cutting chilled-water production cost by up to 70% during active hours.

Economizer operation can lower precision cooling energy use by 30-50%, depending on local temperature and humidity, per ENERGY STAR and vendor field data. The catch is climate: free cooling is a major lever in the Nordics, Northern Europe, Canada, and the US Pacific Northwest, where ambient wet-bulb temperatures stay low most of the year. It is close to unusable in the humid tropical markets where Coloprice tracks a growing share of new capacity — see our Singapore and Johor market guides — which is one reason those markets post structurally higher PUE and pay more for mechanical cooling.

PUE: the industry’s cooling scorecard

Power Usage Effectiveness (PUE) divides total facility power by IT equipment power. A PUE of 1.0 is the theoretical floor — zero overhead. The industry-wide annual average was 1.52 in Uptime Institute’s 16th annual Global Data Center Survey (2026), a figure that has moved only slowly for years.

PUE tier	Typical facility type
1.0-1.15	State-of-the-art hyperscale with liquid cooling and aggressive free cooling
1.2-1.3	New-build colocation and hyperscale facilities, industry-leading design
1.36	Uptime’s size-weighted global average (large facilities counted proportionally)
1.5-1.6	Typical air-cooled enterprise or mid-size colocation hall
1.52	Uptime’s simple (unweighted) global industry average, 2026
1.8+	Older, poorly optimized, or high-redundancy legacy facilities

The gap between the 1.52 headline average and 1.2-1.3 new-build performance exists because thousands of older facilities are still running and dilute the average — new efficient capacity is being added faster than old capacity is retired, not replacing it outright. For water-side tradeoffs of these same cooling choices, see our guide on water usage in data centers.

Air’s ceiling and liquid’s floor

Air cooling has a practical density limit. Past roughly 30-40 kW per rack, airflow volume and fan power needed to remove heat become impractical, and hot-spot risk rises sharply.

Rack density	Standard cooling approach
Under 10 kW	CRAC/CRAH air cooling, standard hot/cold aisle
10-30 kW	Air cooling with contained aisles or in-row cooling
30-40 kW	Air cooling at its practical ceiling; some operators begin hybrid liquid-assist
40-120 kW	Direct-to-chip (DTC) liquid cooling becomes standard
120-250+ kW	Immersion cooling, or DTC with maximum coolant flow

NVIDIA’s GB200 NVL72 draws roughly 120-132 kW as deployed and ships with direct-to-chip cooling built in. Uptime Institute reports 22% of operators had adopted direct liquid cooling as of 2026, mostly for AI and HPC racks rather than facility-wide. For the detailed comparison of direct-to-chip against immersion — cost per rack, retrofit complexity, and which fits which workload — see our dedicated guide on liquid cooling for AI data centers.

What cooling actually costs

Cooling investment	Typical cost	Notes
CRAC unit replacement	$15,000-50,000	Per unit, 30-200 kW capacity
CRAH unit	$25,000-75,000	Plus shared chiller plant cost
Direct-to-chip retrofit	$28,000-42,000 per rack	Cold plates, manifolds, CDU
Immersion cooling (single-phase)	$4,500-6,800 per kW of IT load	Installed cost
Colocation facility retrofit for liquid cooling	~$2 million per MW	Existing building, electrical headroom permitting
New liquid-ready greenfield build	$11 million+ per MW	Purpose-built

The binding constraint on any cooling upgrade is usually electrical headroom, not the cooling equipment itself — a facility without spare power capacity cannot support higher-density racks regardless of how good its cooling plant is. That is the same grid bottleneck covered in our guide on grid connection queues.

Cooling market size and growth

2026 market-size estimates for global data center cooling vary by research firm and scope — from roughly $13 billion to $31 billion for the current year — but the growth rate is consistently high, clustering around 12-17% CAGR into the early 2030s. The liquid-cooling segment specifically is forecast to grow much faster: from about $4.07 billion in 2026 to $27.65 billion by 2033, a 31.5% CAGR, reflecting how far AI rack density has outrun air cooling’s practical ceiling.

Regional considerations

Cooling technology choice is inseparable from climate and water availability:

• Temperate/cold markets (Nordics, Northern Europe, US Pacific Northwest, parts of Canada) get extensive free-cooling hours, pushing PUE toward the low end of the range and reducing mechanical cooling opex.
• Tropical/humid markets (Singapore, Bangkok, Johor, most of Southeast Asia) get little to no free cooling and rely on mechanical chilling or evaporative towers year-round, which is one factor behind the region’s higher colocation pricing.
• Water-constrained markets increasingly favor air-cooled chillers or closed-loop liquid systems over evaporative cooling towers to avoid water-use scrutiny — see our guide on water usage benchmarks.

What to do with this

For colocation buyers evaluating a facility, ask for the trailing-12-month PUE (not a design PUE), the split between CRAC and CRAH/chilled-water plant, and whether free cooling is climate-viable at the site — these three numbers predict both energy cost pass-through and headroom for future density increases. For AI/HPC buyers needing 40 kW+ per rack, confirm liquid-cooling readiness (CDU capacity, piping, coolant loop isolation) before signing, since retrofitting mid-contract is disruptive and expensive. Compare live facility specs and get cooling-aware quotes at Coloprice’s data center directory or request pricing through our quote form.