GB300 NVL72 Rack vs HGX 8-GPU Nodes

This is a form-factor and procurement decision, not a chip decision — both paths are NVIDIA Blackwell.

A GB300 NVL72 is a rack-scale unit: 72 Grace Blackwell Ultra GPUs wired into a single fifth-generation NVLink domain, liquid-cooled by an in-rack CDU, delivered as one factory-integrated 19-inch rack. You buy the rack as a whole. An HGX 8-GPU node — B200 or B300 — is a server you deploy incrementally: rack it yourself (or through an integrator) into standard facilities, in air- or liquid-cooled builds, and add nodes as you grow.

The rack wins when you need one very large accelerator — the largest single-domain training runs, maximum density, a greenfield build-out. Nodes win when you want to add capacity gradually, fit standard facilities, or run a mix of workloads that don’t need all 72 GPUs in one NVLink domain.

The two form factors compared. The rack trades granularity for a single 72-GPU NVLink domain and maximum density; nodes trade the rack-scale domain for incremental, facilities-flexible deployment.

The defining advantage of the GB300 NVL72 is the single NVLink domain: all 72 GPUs connect over fifth-generation NVLink into one coherent accelerator (roughly 130 TB/s all-to-all in-rack), rather than eight-GPU islands stitched together over the scale-out network. For the largest training runs, that lets a model shard across 72 GPUs with far less cross-node traffic than the same GPU count spread over nine separate nodes.

It is also the density and integration play. The rack arrives factory-integrated and NVLink-domain-complete — 18 compute nodes, 36 Grace CPUs, in-rack CDU, and scale-out fabric assembled and burned in — so a greenfield build-out stands up as whole racks instead of node-by-node assembly. The trade-off is liquid cooling: the NVL72 requires an in-rack CDU and a facility that can supply it.

HGX 8-GPU nodes win on flexibility. You can add capacity node-by-node — start with a few and grow as demand and budget allow, rather than committing to a full rack up front. They also fit standard facilities: air-cooled builds (e.g. 8U) drop into rooms with no liquid loop, and liquid-cooled builds (e.g. 4U) are available where the facility supports them.

Nodes also suit mixed workloads — inference, fine-tuning, and smaller training jobs that don’t need all 72 GPUs in one NVLink domain run happily on independent 8-GPU nodes. And they give you integrator choice: the B200 and B300 ship from Lenovo, Supermicro, GIGABYTE, HPE, and Dell, so you match the node to your rack, cooling, and support preferences.

Match the form factor to your scale and facilities:

• Largest single-domain training, max density, or a new build-out → the GB300 NVL72 rack. One 72-GPU NVLink domain, factory-integrated, liquid-cooled.
• Incremental capacity, standard facilities, or mixed workloads → HGX 8-GPU nodes such as the HGX B200 or the Blackwell Ultra HGX B300. Deploy node-by-node, air or liquid.

Many buildouts do both — nodes for flexible capacity, a rack for the largest jobs. For the full rack composition (18 compute nodes, 36 Grace CPUs, CDU, fabric) see the GB300 NVL72 reference architecture. Pricing across every configuration is quoted per configuration on request — browse the GPU catalog and tell us the buildout you need.