Folks looking at the NVIDIA GB200 NVL72 and GB300 NVL72 racks might have seen that the power and liquid cooling have been standardized between various components. NVIDIA took an OCP ORv3 Rack and enhanced it for its NVL72 designs. At NVIDIA GTC 2025, we saw a number of liquid cooling manifolds out of the MGX racks. This was neat to see both the standardization of the components and the variations between eight different vendors.
This is the NVIDIA MGX 44RU Manifold for Liquid Cooling
As a quick bit of context, here is the ASUS AI POD for the NVIDIA GB200 NVL72 (we also covered the GB300 NVL72 version at GTC 2025.) As you can see, the eighteen compute trays and nine NVLink Switch trays are at the heart of the system. These are all liquid cooled.
The manifolds in the rear of the rack has hot and cold sides. The connections to the components are designed to be blind-mating. To install a node, you just slot it in the correct rack space, and then slide it in. Power rails are connected to the rack’s busbar, NVLink connectors mate, and even the liquid cooling tubing from the nodes to the rack manifolds.
We noted that the manifolds in the ASUS designs we saw were Cooler Master. At GTC 2025, we saw eight different vendors making the rack manifolds. Here they are from Pinda, LiteOn, Lead Wealth, Foxconn, Delta, Cooler Master, AVC, and Auras.
While the size, nozzles, positioning, hose lengths, and more are very standardized due to the NVIDIA MGX design, companies are able to differentiate on some of the materials used and their manufacturing capabilities.
Final Words
The NVIDIA MGX realm is very standardized. This is one of the more varied parts of liquid cooling for the NVIDIA MGX systems. Looking at the liquid cooling blocks, for example, it is much more difficult to see differences. Still, it is neat to see these components not just out of the racks but also on display side-by-side.
Also, if you want to see this in a factory, here is a tour with these being installed in racks:
