Liquid Cooling Racks Explained: Direct-to-Chip,...

In a nutshell: Liquid cooling moves heat with fluid rather than relying only on room air. It is becoming mainstream for dense AI and HPC racks, but it changes facility design, service workflows, leak management and vendor responsibility.

Why liquid cooling is growing

High-end CPUs and GPUs concentrate more heat into smaller packages. Air cooling can still work for many enterprise racks, but dense AI systems may require airflow, fan power and temperature margins that become inefficient or impractical.

Liquid has much higher heat-carrying capacity than air. Moving heat closer to the chip can reduce fan energy and allow higher rack densities, especially when paired with warm-water loops and efficient heat rejection.

The main liquid cooling models

Model	How it works	Operational note
Rear-door heat exchanger	A liquid-cooled door captures heat from server exhaust air.	Good retrofit option; servers remain air-cooled.
Direct-to-chip cold plates	Coolant flows through plates mounted to CPUs/GPUs.	Common for AI racks; still needs some air cooling for other components.
Single-phase immersion	Servers are submerged in dielectric fluid that stays liquid.	Changes service process and hardware compatibility.
Two-phase immersion	Fluid boils at component surfaces and condenses back.	High heat-transfer potential; stricter fluid and containment controls.

CDUs and facility loops

A cooling distribution unit, or CDU, separates the technology cooling loop from the facility water loop. It controls temperature, pressure, flow, filtration and leak detection. For direct-to-chip systems, the CDU is often the operational boundary between IT equipment vendors and facilities teams.

Designers need to define supply temperature, flow rate, pressure drop, water quality, redundancy, quick-disconnect types and maintenance access before racks arrive on site.

Risks and trade-offs

Leak management. Use dripless connectors, sensors, containment and tested response procedures.
Serviceability. Technicians need safe workflows for swapping nodes, hoses, manifolds and cold plates.
Mixed cooling. Many racks still need air cooling for memory, storage, NICs and power supplies.
Vendor boundaries. Clarify who owns the rack manifold, CDU, coolant chemistry and warranty claims.
Retrofit limits. Floor loading, pipe routes, water loops and heat rejection may be the real blockers.

Questions before ordering AI racks

What is the maximum rack kW the hall can support continuously? Is there a facility water loop at the required temperature and flow? Is the CDU redundant and maintainable under load? Are quick disconnects, hoses and manifolds standardised across vendors? Who monitors leaks and coolant quality? What is the procedure for replacing a failed GPU node?

Sources and further reading

Find vendors: use the TechDirectory company directory to compare telecom providers, system integrators, data-centre operators, IoT specialists and managed service providers in Singapore.

Liquid Cooling Racks Explained: Direct-to-Chip, Rear-Door and Immersion