Summary
- LG’s 600kW CDU passed more than 100 Nvidia assessment criteria for AI Factory infrastructure.
- The unit includes tight temperature control, virtual sensing, leak detection, and central cooling-system integration.
- Validation reduces part of the supplier-qualification burden, while site redundancy, water chemistry, controls, and heat rejection remain project-specific.
LG Electronics has received Nvidia AI Factory validation for a 600kW coolant distribution unit designed to serve high-density direct-to-chip computing systems.
The CDU passed more than 100 assessment criteria covering cooling performance, standardisation, reliability, monitoring, and failover behaviour. LG has joined the group of suppliers whose equipment has been assessed against Nvidia’s requirements for AI infrastructure.
Installed between the technology-cooling loop serving server cold plates and the facility-water system connected to chillers or other heat-rejection equipment, the 600kW unit can support a concentrated group of high-density racks, depending on server configuration and redundancy design.
Control moves closer to the processors
LG specifies coolant-temperature control within plus or minus 0.25°C under the applicable test conditions. Stable temperature, pressure, and flow become increasingly important as processors operate at high heat flux and rapid workload changes alter the thermal load across a cluster.
The unit includes virtual sensing, real-time monitoring, and leak-detection functions. Integration with LG’s Data Center Cooling Control Manager allows the CDU to exchange operating information with the wider cooling system and, through standard protocols, the building or data centre management platform.
LG tests its equipment at a chip-to-chiller validation facility in Pyeongtaek, where simulated processor loads, cold plates, CDUs, chillers, and controls can be examined as a complete thermal chain. Testing covers rapid load changes, low-load stability, temperature response, and system interactions.
The LG technical release provides the validation scope and product information.
Qualification does not replace system engineering
Direct-to-chip cooling introduces interfaces between the server manufacturer, cold-plate supplier, manifold designer, CDU manufacturer, water-treatment specialist, controls integrator, and facility cooling contractor. Each party must agree operating temperatures, fluid chemistry, materials, pressure limits, alarms, and failure responses.
Nvidia validation provides a common technical reference and can reduce part of the equipment-qualification process. It does not confirm that a specific data centre design will meet its availability, efficiency, commissioning, or maintenance requirements.
Project teams must decide whether CDUs serve individual rows, computing pods, or larger shared zones. A nominal 600kW unit may be installed in an N+1 arrangement, with dual power feeds, redundant pumps, isolation valves, and bypasses that permit maintenance without interrupting cooling.
The installed CDU capacity will therefore exceed the live IT heat load where redundancy is required. Shared units can reduce equipment count and simplify some pipework, although they increase the amount of computing affected by a single failure or maintenance intervention.
Water chemistry remains site-specific. The technology loop may use treated water or another approved coolant, while the facility side operates at different pressure and quality conditions. Heat exchangers separate the circuits, but fouling, corrosion, particulate contamination, and biological growth still require controls and maintenance.
LG is developing the 600kW unit as part of a broader chip-to-chiller portfolio that includes cold plates, distribution equipment, controls, and large cooling plant. It plans to pursue Nvidia validation for larger 1MW, 2.5MW, and 4MW CDU models.
Larger units can reduce the number of packages in a major AI hall, although they also concentrate more heat load behind each device. Isolation, redundancy, fault containment, commissioning, and pipe routing become more demanding as unit capacity increases.
The CDU still depends on the external plant. In cooler European climates, warmer facility-water temperatures can increase the hours available for dry or economised cooling. During warm weather or peak load, chillers, dry coolers, or hybrid systems must reject the full heat output while maintaining the server loop within its operating range.
Liquid systems can also provide a more concentrated heat stream for reuse. District-heating or industrial applications may still need a heat pump where the available temperature is below the customer’s requirement, adding electrical demand and capital cost.
Commissioning must cover pump changeover, sensor failure, loss of facility water, pressure excursions, control-network failure, leaks, and rapid changes in processor load. A validated CDU cannot compensate for an undersized chiller, weak hydraulic design, or poorly coordinated controls.
Local service capability and spare-part availability will influence procurement alongside technical validation. A CDU failure can affect several racks at once, making response time, trained technicians, replacement pumps, sensors, controllers, and heat-exchanger support part of the resilience assessment.
LG’s approval adds another qualified option to a market expanding quickly as air cooling reaches its practical limits in AI halls. Final selection will depend on delivery time, redundancy, efficiency, maintenance, facility integration, and compatibility with the chosen computing platform.
