HGX is NVIDIA’s 8-GPU board, built into servers by Dell, HPE, Lenovo, Supermicro, Gigabyte and others. An HGX B300 server has 8 GPUs with up to 288 GB each; the reference DGX B300 draws around 14 kW. Servers can be air- or liquid-cooled and fit many existing data halls.
NVL72 is a full rack: 72 GPUs and 36 Grace CPUs in one NVLink domain, liquid-cooled, drawing up to 142 kW for GB300. It behaves more like one very large accelerator and suits frontier-scale training and large-model inference.
How to choose: if the site can’t provide 130 kW+ per rack with liquid cooling, start with HGX. If it can, and workloads need very large models across many GPUs, NVL72 is the step up.
Also specify: CPUs, system RAM, storage, network cards and fabric (InfiniBand or Ethernet), and whether optics and cables are included. The same GPU can be priced very differently depending on these.