Data Centre Cooling Technologies Compared: Air, Liquid, and Immersion

Cooling Is Now the Bottleneck

Data centre cooling has transitioned from a commodity infrastructure concern to a strategic differentiator. The shift to AI workloads - with GPU rack densities routinely exceeding 100 kW - has made traditional air cooling inadequate for the highest-density applications. Operators with liquid cooling capacity are quoting 20-40% premiums over air-cooled space in the same market, and filling faster.

Air Cooling: The Industry Standard

Air cooling has been the default data centre cooling technology for three decades. It works by circulating chilled air through the server room, absorbing heat from IT equipment, and rejecting that heat to the outside environment.

**How it works:** - Computer Room Air Conditioning (CRAC) or Computer Room Air Handlers (CRAH) units push chilled air under the raised floor - Cold air enters server racks through perforated floor tiles (hot aisle/cold aisle configuration) - Hot exhaust air is returned to the cooling units - Heat is rejected outdoors via chillers, cooling towers, or dry coolers

**Performance:** - Supports up to 15-20 kW per rack (with optimised airflow management) - PUE range: 1.3-1.8 (average facility) - Well-understood technology with a deep vendor ecosystem

**Cost:** - Lowest capital cost of any cooling technology - Moderate operating cost (20-40% of total facility power goes to cooling)

**Best for:** General enterprise workloads, colocation with mixed-density tenants, retrofits of existing facilities.

**Limitations:** Cannot efficiently cool rack densities above 25 kW without supplemental cooling. Free cooling (using outside air) is geographically dependent.

Direct Liquid Cooling (DLC)

Direct liquid cooling circulates a coolant (typically water or a water-glycol mixture) directly to heat-generating components via cold plates mounted on CPUs and GPUs. The liquid absorbs heat at the chip level and transports it to a heat rejection system.

**How it works:** - Cold plates attach directly to CPUs, GPUs, and memory modules - Chilled coolant flows through the cold plates, absorbing heat - Warm coolant returns to a Coolant Distribution Unit (CDU) - The CDU rejects heat to the building's chilled water loop or outdoor heat rejection

**Performance:** - Supports 80-150+ kW per rack - PUE range: 1.1-1.3 - Captures 70-80% of server heat directly at the chip (remainder handled by supplemental air cooling)

**Cost:** - 15-25% higher capital cost than air cooling - Significantly lower operating cost (cooling power reduced by 40-60%)

**Best for:** AI training and inference workloads, high-performance computing, GPU-dense deployments. This is now the standard for NVIDIA H100/H200 and Blackwell GPU clusters.

**Market context:** The liquid cooling market is projected to grow from $4.2 billion to $32 billion by 2028. Ecolab's $4.75 billion acquisition of CoolIT Systems validated the technology's transition to mainstream.

Immersion Cooling

Immersion cooling submerges entire servers in a dielectric (non-conductive) liquid. The liquid absorbs heat from all components simultaneously.

Two variants:

**Single-phase immersion:** Servers are submerged in a pool of dielectric fluid that remains liquid throughout the process. Heat is transferred from the fluid to a heat exchanger connected to the building's cooling system.

**Two-phase immersion:** Servers are submerged in a low-boiling-point dielectric fluid. The fluid boils at the chip surface, absorbing heat as it transitions from liquid to gas. The vapour rises, condenses on a coil, and drips back into the pool.

**Performance:** - Supports 100-250+ kW per rack equivalent - PUE range: 1.02-1.10 (near-theoretical minimum) - Eliminates fans entirely (zero airflow required)

**Cost:** - Highest capital cost (30-50% premium over air cooling) - Lowest operating cost (cooling power reduced by 90%+) - Dielectric fluid is expensive ($40-80/litre for engineered fluids)

**Best for:** Extreme-density deployments, harsh environments (dust, humidity), noise-sensitive locations, and operators prioritising PUE and sustainability metrics.

**Limitations:** Maintenance procedures are more complex (servers must be removed from fluid). Limited vendor ecosystem. Some server warranties are voided by immersion.

Comparison Matrix

The Hybrid Approach

Most modern facilities use a combination of cooling technologies. A typical 2026-era facility might use: - Air cooling for network equipment and storage (5-10 kW/rack) - Direct liquid cooling for GPU clusters (80-120 kW/rack) - Immersion cooling for specialised HPC zones (150+ kW/rack)

This hybrid approach optimises cost and performance across the full range of workload densities.

Evaluate Your Facility

Use our Score Tool to assess any property's power and cooling viability. Browse facilities by city to find data centres with the cooling technology you need, or contact our advisory team for help designing your cooling strategy.