NVIDIA GB200 NVL72 Specs & Datasheet (72-GPU Rack)

The GB200 NVL72 is NVIDIA’s Grace Blackwell rack-scale system: a single, factory-integrated, liquid-cooled rack that connects 72 B200 GPUs and 36 Grace CPUs into one coherent accelerator over a fifth-generation NVLink fabric.

The building block is the GB200 Grace Blackwell superchip — one Grace CPU paired with two Blackwell GPUs on a single module. Thirty-six of those superchips, two per compute tray across 18 trays, give the rack its 72 GPUs and 36 CPUs. Unlike an HGX B200 node that you rack yourself in a standard chassis, the NVL72 arrives as a complete rack: compute trays, NVLink switch trays, power shelves, networking, and liquid cooling integrated at the factory.

The result is roughly 13.8 TB of HBM3e addressable across a single NVLink domain. It is the Blackwell predecessor to the Blackwell Ultra GB300 NVL72.

The GB200 NVL72 rack specification, aggregated from the current integrator builds. The per-GPU and per-rack GPU, memory, and NVLink figures below are fixed characteristics of the platform; site power draw, rack weight, and final configuration vary by integrator and facility, and are confirmed at quote.

Each Blackwell GPU in the GB200 NVL72 carries 192 GB of HBM3e — the on-package memory available to the model. Across 72 GPUs that is roughly 13.8 TB of HBM3e per rack (72 × 192 GB = 13,824 GB), pooled and addressable over NVLink so a frontier-scale model and its long-context KV cache can span the entire rack.

Per-GPU memory bandwidth is about ~8 TB/s. What makes that number matter is the fabric: because all 72 GPUs sit in one NVLink domain, the rack behaves like a single very large accelerator rather than 72 networked cards, and the practical ceiling on model size is the pooled 13.8 TB rather than any one GPU’s 192 GB.

On compute, each GPU delivers ~5 PFLOPS of dense FP8 and ~10 PFLOPS of dense FP4, so a full rack reaches roughly ~360 PFLOPS of dense FP8 and ~0.72 EFLOPS of dense FP4. All figures here are dense; NVIDIA’s published rack headline of 1,440 PFLOPS FP4 is the "with sparsity" number, i.e. 2× the dense figure.

The NVL72 rack is built from three repeating elements plus power and cooling:

• 18 compute trays. Each holds two GB200 superchips — 2 Grace CPUs and 4 Blackwell GPUs per tray — for 72 GPUs and 36 Grace CPUs across the rack.
• 9 NVLink switch trays. These carry the fifth-generation NVLink switch silicon that gives every GPU all-to-all connectivity to every other GPU in the rack, for roughly ~130 TB/s of aggregate in-rack NVLink bandwidth.
• Power shelves. Rack-level shelves convert facility input and feed the compute and switch trays over a common busbar, rather than per-server PSUs.

The cabling is what distinguishes NVL72 from a cluster of nodes: an NVLink spine connects the compute trays to the switch trays inside the rack, so GPU-to-GPU traffic never leaves the chassis for the network fabric. That is the topology that makes 72 GPUs act as one NVLink domain — and it is why the rack ships factory-integrated rather than assembled on site.

The GB200 NVL72 is a high-density, liquid-cooled rack served by an in-rack coolant distribution unit (CDU) with direct-to-chip liquid cooling across the Grace CPUs, Blackwell GPUs, and NVLink switches. At roughly 1,200 W per GPU for 72 GPUs — before CPUs, switches, and conversion losses — liquid cooling is not optional; it is what makes the density thermally viable at all.

NVIDIA and its integrators commonly describe the NVL72 as an approximately 120 kW rack, and that is the right order of magnitude for facility planning. Treat it as a planning figure rather than a spec: the exact site power draw and rack weight depend on the integrator build, the configuration, and the facility, and are confirmed at quote. Plan for facility water and high-density power distribution.

Beyond the in-rack NVLink domain, the GB200 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. NVLink handles the dense all-to-all traffic inside each rack; ConnectX stitches many racks into a larger training cluster.

The GB300 NVL72 is the Blackwell Ultra successor on the same rack-scale platform. The defining difference is memory: 288 GB of HBM3e per GPU versus 192 GB, so a GB300 rack holds roughly 20.7 TB against the GB200’s 13.8 TB. Blackwell Ultra also lifts dense FP4 from ~10 to ~15 PFLOPS per GPU (~1.1 EFLOPS vs ~0.72 EFLOPS per rack), while dense FP8 is unchanged at ~5 PFLOPS.

Everything else is shared: 72 GPUs, 36 Grace CPUs, one fifth-generation NVLink domain at 1.8 TB/s per GPU, and ConnectX-8 scale-out. For the full side-by-side see GB300 NVL72 vs GB200 NVL72.

The GB200 NVL72 ships as a factory-integrated rack from several OEMs, each with its own chassis, cooling detail, and support model. The GPU, memory, and NVLink specifications are common across builds; chassis, cooling, warranty, and lead time differ by integrator. Each build below links to its full specification.

Every GB200 NVL72 build is quoted per configuration on request — the figure depends on integrator, cooling, scale-out fabric, and support scope. Lead times vary by integrator and are confirmed at quote. Tell us the cluster you are standing up and we will return pricing, availability, and a facility-fit review.

For a new rack-scale buildout the current generation is the GB300 NVL72, and that is where we point most buyers; we carry the GB200 NVL72 for requirements that specifically call for it. Browse the GPU catalog to compare integrations.