Modern data centers increasingly blend air and liquid cooling to handle AI-driven rack densities exceeding 100 kW

1. Why Data Center Cooling Matters More Than Ever

Cooling has shifted from a supporting function to the primary constraint on data center capacity. The Uptime Institute's 2025 Global Data Center Survey found that PUE improvements have stalled for the sixth consecutive year, with legacy air-cooled facilities stuck at PUE 1.55–1.67, wasting 30–40% of their total electricity on cooling alone.

The driver is unmistakable: AI. GPU power consumption has escalated beyond what air cooling can handle for dense deployments. Average data center rack power density increased 38% from 2022 to 2024, with AI clusters now pushing 80–120 kW per rack. NVIDIA's roadmap is even more aggressive:

The physics are unforgiving. Convective heat transfer through air limits cooling effectiveness regardless of fan speed or air handler capacity. Liquid offers approximately 3,000 times the heat capacity of air, enabling fundamentally superior thermal management at high densities. For a deeper look at how these power demands reshape cabling infrastructure, see AI Infrastructure: How Machine Learning Is Reshaping Data Center Cabling Requirements.

2. Air Cooling Systems: CRAC, CRAH, and Containment

Air cooling remains the most widely deployed method, leveraging mature technology and familiar operational models. Three primary architectures dominate:

2.1 CRAC Units (Computer Room Air Conditioning)

Traditional CRAC units compress refrigerant to cool air, circulating it through raised-floor plenums or overhead ducts. They provide reliable, centralized cooling for moderate heat loads but struggle with hotspots in high-density deployments. CRAC units are cost-effective for facilities under 15 kW per rack.

2.2 CRAH Units (Computer Room Air Handlers)

CRAH units use chilled water from central plant systems, passing return air over cooling coils. They handle larger thermal loads more efficiently than CRAC units and are standard in hyperscale facilities. CRAH systems scale well but depend on chiller plant capacity and water infrastructure.

2.3 In-Row Cooling and Containment

In-row cooling places precision air handlers directly between server cabinets, shortening the airflow path and improving efficiency. Combined with hot-aisle or cold-aisle containment, these systems can support 15–25 kW per rack. Containment prevents hot and cold air mixing, reducing fan energy by 15–20% compared to uncontained layouts.

Air cooling's fundamental limitation is thermodynamic: as rack density increases, the volume of air required for adequate heat removal grows exponentially. Fan power scales cubically with airflow, meaning doubling airflow requires eight times the fan energy. Above 25 kW per rack, this becomes economically unsustainable. Facilities managing high-density cable infrastructure should also review material compatibility and routing design for 30kW–120kW racks to ensure cabling withstands elevated operating temperatures.

Hot-aisle containment captures exhaust heat before it mixes with supply air, improving CRAC/CRAH efficiency by 15–20%

3. Liquid Cooling Technologies: D2C, Immersion, and RDHx

Liquid cooling is no longer experimental — it is the baseline specification for AI-era data centers. Three primary technologies address different density tiers and deployment contexts:

3.1 Direct-to-Chip (D2C) Liquid Cooling

D2C circulates coolant through cold plates mounted directly on CPUs, GPUs, and other high-power components. Heat is captured at the source — the most thermally efficient approach available. ASME-validated studies show D2C boosts GPU performance by 2.7% in Gflops per second while cutting overall power consumption by 12% and dropping chip operating temperatures by 20°C compared to air cooling.

D2C systems use Coolant Distribution Units (CDUs) to regulate flow, temperature, and pressure between facility and IT loops. Supermicro released NVIDIA Blackwell rack-scale solutions with 250 kW CDUs — double previous capacity — reflecting the escalating power demands of next-generation GPUs.

Two D2C variants exist:

• Single-phase: Coolant remains liquid throughout the cycle. Simpler to operate, widely deployed, lower risk.
• Two-phase: Coolant boils on hot chip surfaces and condenses back, achieving superior heat transfer. ZutaCore's HyperCool system uses a non-toxic dielectric fluid with ultra-low global warming potential (GWP) that safely evaporates if leaked.

3.2 Immersion Cooling

Immersion cooling submerges entire servers in dielectric fluid, eliminating fans and airflow management entirely. Every heat-producing component — not just CPUs and GPUs — contacts the cooling fluid directly. This approach achieves the highest thermal performance available:

• Single-phase immersion: Fluid remains liquid, circulating through external heat exchangers. Submer's SmartPod achieves 140 kW per rack with PUE between 1.03 and 1.1.
• Two-phase immersion: Fluid boils on hot surfaces and condenses back. Microsoft tested two-phase immersion for AI training clusters, reporting a 30% energy efficiency gain and improved hardware reliability.

The immersion cooling market reached $4.87 billion in 2025 and is forecast to hit $11.10 billion by 2030 (17.91% CAGR). However, immersion requires specialized hardware compatibility, dedicated tank infrastructure, and rethinking maintenance procedures — servers must be lifted from fluid for servicing.

3.3 Rear-Door Heat Exchangers (RDHx)

RDHx replaces standard cabinet rear doors with liquid-cooled coils. Hot server exhaust passes through the coil before entering the room, transferring heat to the liquid loop. The liquid never touches IT equipment — making RDHx the lowest-risk liquid cooling option and an ideal retrofit for existing facilities.

RDHx handles 20–40 kW per rack and serves as a bridge between air and full liquid cooling. Legrand's CL20 ProActive and CL23 High Performance Capacity models integrate with existing chilled water systems and monitoring infrastructure, enabling capacity expansion without comprehensive redesign.

For facilities planning the cabling infrastructure that accompanies these cooling technologies, in-rack cooling and rack cable architecture must be designed concurrently to avoid thermal interference between cable bundles and cooling loops.

4. Hybrid Cooling: The Pragmatic Middle Ground

Most data centers are not built for today's high-density workloads and cannot be taken offline for a complete rebuild. Hybrid cooling — combining air and liquid in the same facility — has become the dominant deployment strategy:

• Standard server racks (10–20 kW): Continue using air cooling with containment
• AI and HPC racks (60–150+ kW): Deploy direct-to-chip liquid cooling
• Transitional racks (25–40 kW): Use rear-door heat exchangers to augment air cooling

D2C is particularly well-suited for hybrid deployment because it cools only the hottest components (GPUs, CPUs) while relying on existing air infrastructure for lower-power elements (memory, storage, network cards). This additive approach — rather than wholesale replacement — limits retrofit scope and enables operational teams to gain liquid cooling experience incrementally.

Nautilus Data Technologies brings over 500,000 unit-hours of fully liquid-cooled AI facility operation, with its EcoCore FCD delivering 3.2 MW of cooling capacity — the highest among liquid-to-liquid systems in NVIDIA's DSX infrastructure catalog. Their patented vacuum-sealed closed-loop design eliminates leakage risk, addressing one of the most common operator concerns. For organizations planning phased transitions, AI retrofit projects for existing data centers provides a structured framework.

5. Key Metrics: PUE, WUE, and TCO Analysis

5.1 Power Usage Effectiveness (PUE)

PUE measures total facility power divided by IT equipment power. Lower is better. The industry is moving beyond PUE as the sole efficiency metric, but it remains the most widely understood benchmark:

For a 50 MW data center, moving from PUE 1.60 to PUE 1.10 saves 15.6 MW of cooling load — enough to power 15,000 additional servers without expanding the facility footprint. Dell's comparative testing measured a chip-to-coolant temperature difference of 17–20°C with single-phase liquid cooling at 500W processor loads, versus 60°C or higher with air cooling under identical conditions.

5.2 Total Cost of Ownership (TCO)

Liquid cooling systems require 15–30% higher upfront capital investment — including CDUs, cold plates, manifolds, piping, and leak detection. However, the operational savings are substantial:

• Energy cost reduction: 30–50% lower cooling energy compared to air
• Deferred facility expansion: Higher rack density avoids new white-space construction
• Hardware lifespan extension: Lower operating temperatures reduce thermal stress and failure rates
• Free cooling enablement: Higher supply temperatures allow year-round economizer operation in many climates
• Waste heat reuse: Liquid loops deliver higher-grade heat suitable for district heating (operational in Finland and Sweden)

ROI typically materializes within 2–3 years through energy savings alone, before accounting for the revenue enabled by supporting AI workloads that air cooling physically cannot host. For broader infrastructure cost optimization, see power flexibility strategies for AI data centers.

6. Best Practices for Deployment and Operations

6.1 Pre-Deployment Assessment

6.2 Airflow Management First

Even with liquid cooling deployed, airflow management remains critical. Seal all unused rack U-spaces with blanking panels, install grommets and brush strips in cable cutouts, and use vertical exhaust ducts to prevent hot-air recirculation. Facilities without complete hot-aisle containment should prioritize this before any liquid cooling deployment — it's the highest ROI improvement available.

6.3 Leak Detection and Safety Protocols

Modern liquid cooling systems incorporate multiple safety layers: quick-disconnect fittings that minimize fluid loss during maintenance, pressure sensors that detect anomalies in real time, and automatic shutoff valves. Nautilus's vacuum-sealed closed-loop design represents the gold standard — negative pressure means any breach draws air inward rather than leaking fluid outward. Train operations teams on coolant handling, emergency procedures, and routine inspection protocols before deployment.

6.4 Phased Migration Strategy

Proper cable management becomes even more critical in liquid-cooled environments where space is constrained by cooling infrastructure. Review patch panel cable management best practices for data centers and cabling strategies for high-density data centers to ensure thermal and cable pathways don't conflict.

7. Regulatory Landscape and Sustainability

Governments and regulators across Europe and Asia are drafting increasingly rigorous cooling efficiency requirements. New and high-density data centers face mandates to implement liquid cooling and publicly report PUE, WUE (Water Usage Effectiveness), and other ESG indicators:

• EU Energy Efficiency Directive: Requires data centers over 500 kW to report PUE, WUE, and REF (renewable energy factor) annually
• Germany: Amended Energy Industry Act mandates liquid cooling readiness for new facilities above 100 kW IT load
• Singapore: Green Data Centre Roadmap sets PUE targets of 1.3 for new tropical facilities
• China: East Data West Computing project requires PUE below 1.25 for new hyperscale facilities in western zones

Beyond compliance, sustainability drives operational decisions. Liquid cooling enables waste heat reuse — transferring captured heat to district heating networks. Facilities in Finland and Sweden already deliver server waste heat to municipal heating systems, reducing both carbon footprint and heating costs for surrounding communities. Thermal energy storage using ice batteries can time-shift cooling loads to off-peak hours, reducing demand charges.

The industry is also moving beyond PUE as the sole efficiency metric. System-level energy recovery, carbon intensity per compute unit, and water consumption per transaction are emerging as more holistic measures of data center sustainability. For facilities navigating European compliance requirements, Europe's AI data center infrastructure priorities provides regional guidance. Operators should also ensure cable jacket materials meet fire safety codes in these regulated environments — see PVC vs LSZH cable jacket comparison for compliance details.

8. Decision Framework: Matching Cooling to Workload

There is no one-size-fits-all solution. The optimal cooling strategy aligns technology with business goals, workload requirements, and facility constraints:

Cooling technology selection follows rack density: air (below 20 kW) transitions through RDHx (20–40 kW) to D2C (50–250 kW) and immersion (100 kW+)

The choice is rarely binary. Most modern deployments balance multiple cooling technologies within a single facility — air for legacy workloads, RDHx for transitional racks, D2C for AI clusters, and potentially immersion for extreme-density research computing. Organizations planning AI infrastructure without liquid cooling capability risk deployment limitations as GPU power continues its inexorable climb. For comprehensive infrastructure planning, review how structured cabling is changing for AI data centers and common data center cabling pitfalls to avoid.

Key Questions (FAQ)

AMPCOM Technical Team

Industry experts with 17+ years in enterprise network infrastructure and structured cabling systems