NVIDIA GB300 NVL72 Reference Architecture & Rack Specs

The GB300 NVL72 is NVIDIA’s rack-scale flagship: a single, factory-integrated, liquid-cooled rack that connects 72 GB300 (Grace Blackwell Ultra) GPUs and 36 Grace CPUs — arranged as 18 compute nodes of 4 GPUs each — into one coherent accelerator over a fifth-generation NVLink fabric.

Rather than a server you rack yourself, the NVL72 is delivered as a complete rack: compute trays, NVLink switch trays, networking, and liquid cooling integrated at the factory. The result is roughly 20.7 TB of HBM3e addressable across a single NVLink domain, for the largest AI training and inference workloads. It is the Blackwell Ultra successor to the GB200 NVL72.

The GB300 NVL72 reference configuration, aggregated from the current integrator builds. Per-GPU and per-rack figures below are the fixed characteristics of the platform; site power draw, weight, and final configuration are confirmed at quote.

Each GB300 GPU carries 288 GB of HBM3e — the Blackwell Ultra memory step-up over the 180 GB B200 — so a full rack holds roughly 20.7 TB of HBM3e. What makes the NVL72 more than 72 discrete GPUs is the fabric: all 72 GPUs share a single fifth-generation NVLink domain at 1.8 TB/s per GPU, with all-to-all connectivity in-rack (about 130 TB/s aggregate) via the NVLink switch trays.

This all-to-all NVLink topology — every GPU one hop from every other GPU through the switch trays — is what separates the NVL72 from a cluster of networked nodes. Because the whole rack is one NVLink domain, a model can be sharded across all 72 GPUs with far lower communication overhead than networking separate 8-GPU nodes — the rack behaves like one very large accelerator with ~20.7 TB of pooled high-bandwidth memory. For the full rack topology and datasheet, see the GB300 NVL72 specs.

On compute, each GB300 GPU delivers ~5 PFLOPS of dense FP8 and ~15 PFLOPS of dense NVFP4, so a full rack reaches roughly ~360 PFLOPS of dense FP8 and ~1.1 EFLOPS of dense FP4 (72 × per-GPU) — the low-precision path NVIDIA positions for large-scale reasoning inference. Figures are dense; sparsity doubles them.

Beyond the in-rack NVLink domain, the GB300 NVL72 scales out to multi-rack clusters over NVIDIA ConnectX-8 SuperNICs at 800 Gb/s per GPU, on Quantum-X800 InfiniBand or Spectrum-X Ethernet fabrics. This is the plane that stitches many NVL72 racks into a larger SuperPOD-class training cluster while NVLink handles the dense all-to-all traffic inside each rack.

The GB300 NVL72 is a high-density, liquid-cooled rack cooled by an in-rack coolant distribution unit (CDU) — direct-to-chip liquid cooling across CPUs, GPUs, and NVLink switches. Liquid cooling is what makes 72 Blackwell Ultra GPUs in one rack thermally viable.

Exact site power draw and rack weight depend on the integrator build and the facility, and are confirmed at quote — plan for a high-density liquid-cooling deployment with facility water and substantial per-rack power. (Some integrator builds specify an in-rack CDU rated around 250 kW; the precise figure for your configuration is confirmed at quote.)

The GB300 NVL72 ships as a factory-integrated rack from several OEMs, each with its own chassis, cooling, and support model:

• Supermicro — GB300 NVL72 on Supermicro MGX, liquid-cooled with an in-rack CDU.
• Lenovo — GB300 NVL72 on the Lenovo 48U MGX reference rack with direct water cooling.
• HPE — GB300 NVL72 on HPE’s 48U MGX-compliant build with HPE Direct Liquid Cooling and Cray deployment services.
• Pegatron — GB300 NVL72 on the Pegatron NVL72 rack, liquid-cooled with an in-rack CDU.

All are new, factory-integrated, and NVLink-domain-complete. The GPU, memory, and NVLink specifications are common across builds; chassis, cooling detail, warranty, and lead time differ by integrator.

Every GB300 NVL72 build is quoted per configuration on request — the figure depends on integrator, cooling, scale-out fabric, and support scope. Lead times run from a few weeks to a few months by integrator, confirmed at quote. Tell us the cluster you are standing up and we will return pricing, availability, and a facility-fit review. Browse the GPU catalog to compare integrations.