Every liquid-cooled rack NVIDIA ships has to hand its heat to something, and that something is a coolant distribution unit. NVIDIA decides which CDUs qualify. Few clear the bar. LG Electronics got one through on July 27. Its 600 kW CDU passed NVIDIA AI Factory validation, the first liquid cooling equipment from a Korean manufacturer to clear the program, according to Startup Fortune's coverage. LG says the unit was measured against more than 100 technical evaluation criteria.
The capacity number sets the scope. Current NVL72 racks draw between 120 and 140 kW depending on generation, so a 600 kW CDU covers four to five of them. That is a row segment. Rubin-class racks push higher still, which is why NVIDIA made liquid cooling standard on Rubin. LG's own roadmap puts the 1 MW, 2.5 MW, and 4 MW units next in the validation queue. The products hyperscalers write purchase orders against have not passed yet.
The earlier number does not match. When LG was still chasing this certification while pulling its chiller revenue target forward, the unit described as being in play was 1.4 MW. What cleared is 600 kW, and the published roadmap names 1 MW, 2.5 MW, and 4 MW with no 1.4 MW anywhere on it. Either the program scoped down or the smaller unit was the faster path to a validated part number. LG has not said which.
LG's release puts temperature control precision at plus or minus 0.25 degrees Celsius and describes virtual sensor technology handling real-time monitoring and leak detection, with a Data Center Cooling Control Manager layered on top for predictive control and centralized monitoring. Testing ran at LG's Chip-to-Chiller validation site in Pyeongtaek, which the company says evaluates transient response, temperature swings, and low-load stability. Low-load stability is the honest inclusion there, because GPU clusters idle between training runs and pumps tuned for design flow hunt badly at part load. What the release does not publish is the approach temperature, the secondary loop supply setpoint, the flow rate in liters per minute, the pump power draw, the redundancy configuration, or the filtration spec, and NVIDIA does not publish its criteria list either, so the 100-plus number counts tests without describing a single one of them.
James Lee, President of LG ES Company, framed the win around scale, saying the need for "highly reliable liquid cooling solutions has never been greater." The commercial mechanics matter more than the framing. NVIDIA's validated vendor lists determine what integrators can specify into a reference build, which turns a thermal test report into market access. LG points to in-house magnetic bearing compressors and a portfolio running from cold plates through CDUs to chillers. Deployments so far include North American hyperscaler facilities and the Sinar Mas SM+ Data Center in Jakarta. Owning the compressor and the cold plate is a cost position most CDU vendors cannot match, in a field with three frontrunners and a hundred chasers.
For operators specifying capacity this year, the practical read is narrow. A validated 600 kW CDU earns LG a line item in reference designs sized for pilot rows and edge-of-hall deployments. The hyperscale halls going to contract now get specified in megawatt CDU blocks. Until the 1 MW and 2.5 MW units carry the same stamp, LG sells chillers into liquid-cooled halls while someone else owns the manifold between the cold plate and the facility loop.