Liquid-cooled AI racks move heat through a controlled fluid path instead of relying only on room air. In a common direct-to-chip design, cold plates collect heat at CPUs and GPUs, a coolant distribution unit (CDU) transfers heat between loops, and outdoor equipment rejects that heat. Understanding each boundary helps teams size the system and write a clear interface specification.
1. Cold plates collect heat at the server
A cold plate mounts to a high-heat component and transfers heat into circulating coolant. The server-side loop also depends on manifolds, quick disconnects, hoses, pumps, sensors, and leak controls. Confirm the server vendor’s allowed coolant, flow, pressure drop, inlet temperature, and connection standard before choosing a facility system.
2. The secondary loop serves IT equipment
The secondary loop distributes conditioned coolant to the racks and returns warmer fluid. Its design must support the rack’s peak and typical heat load while staying inside each server’s operating envelope. Flow balancing, filtration, water chemistry, service isolation, and air removal all affect stable operation.
3. A CDU transfers heat between loops
The CDU is the interface between the IT-side circuit and the facility-side circuit. Depending on its design, it combines a heat exchanger with pumps, controls, sensors, and filtration. A two-loop arrangement can keep the server coolant circuit separate from the outdoor heat-rejection circuit, but the exact hydraulic design and fluid requirements must be confirmed for the selected equipment.
4. A dry cooler rejects heat outdoors
Continue with planning a containerized deployment.
A dry cooler transfers heat from the facility loop to outdoor air. Capacity depends on the design ambient condition, fluid temperature, airflow, coil performance, and operating mode. Some projects consider adiabatic or evaporative assistance in hot weather; that adds water quality, water consumption, and maintenance requirements that should be included in the site plan.
Account for heat that remains in the room
Compare HY's liquid cooling reference architectures.
Direct-to-chip cooling does not automatically remove every watt generated inside a container. Power supplies, memory, storage, network switches, fans, and auxiliary equipment still need a heat path. Estimate the residual room load and specify ventilation or air-conditioning separately from liquid-loop capacity.
Liquid-cooling loop questions
What does a CDU do in a liquid-cooled data center?
A coolant distribution unit circulates coolant and can transfer heat between facility-side and IT-side loops through a heat exchanger. Its pumps, controls, sensors, filtration, and redundancy must match the rack load and server interface.
What is the difference between the primary and secondary cooling loops?
The primary loop carries heat from the CDU toward the heat-rejection equipment. The secondary loop supplies coolant to the IT equipment and returns warmer fluid to the CDU. The CDU separates their hydraulic and fluid requirements while transferring heat.
Does a dry cooler eliminate all room air conditioning?
Not necessarily. Cold plates remove heat from selected components, while memory, storage, network devices, power conversion, and other equipment may still release heat to the room. The residual heat load and ventilation or air-conditioning requirement must be calculated.
Can any glycol mix be used with a server cold plate?
No. Fluid type and concentration must meet the server supplier’s limits and be compatible with pumps, seals, metals, filters, and heat exchangers. Check viscosity, freeze point, water quality, and operating temperature for the actual site.
For a project layout, see the GB300 liquid-cooled container design checklist and PUE measurement guide. Review liquid-cooling products or send HY your rack load and site conditions.