GB200 NVL72 Explained: Rack-Scale AI as One Giant GPU

The GB200 NVL72 is NVIDIA's answer to a problem that emerges at the frontier: when models grow past what any eight-GPU server can hold, the network between servers becomes the bottleneck. The NVL72 attacks that by making an entire rack behave like one accelerator. It is a different unit of compute — not a server you rack alongside others, but a fully integrated, liquid-cooled system designed and delivered as a whole. This is what is inside it and who it is for.

What a GB200 NVL72 actually contains

The NVL72 is a single rack housing 36 GB200 Superchips. Each GB200 Superchip pairs one Grace CPU with two Blackwell GPUs over a 900 GB/s NVLink-C2C link, so the rack totals 36 Grace CPUs and 72 Blackwell GPUs. Across those 72 GPUs it provides roughly 13.5TB of HBM3e. Compute is delivered in liquid-cooled compute trays, interconnected by dedicated NVLink switch trays.

The NVLink fabric is the whole point

Every one of the 72 GPUs connects to every other through fifth-generation NVLink at 1.8 TB/s per GPU, via the rack's NVLink switch system — an aggregate on the order of 130 TB/s of all-to-all bandwidth. The result is a single 72-GPU NVLink domain: a model can be split across all 72 GPUs and treat their combined HBM as one pool, with GPU-to-GPU communication an order of magnitude faster than InfiniBand between separate servers.

• 72 Blackwell GPUs and 36 Grace CPUs in one rack-scale system.
• Roughly 13.5TB of unified HBM3e across the GPU domain.
• Fifth-generation NVLink at 1.8 TB/s per GPU, fully connected through NVLink switch trays.
• Direct-to-chip liquid cooling as a baseline requirement, not an option.

Who needs rack-scale, and who does not

The NVL72 is built for trillion-parameter training and for real-time inference on the largest models, where keeping the model inside one ultra-fast NVLink domain delivers dramatic gains over multi-node setups. If your models fit comfortably in an eight-GPU HGX node, you do not need it — an HGX B200 or HGX H200 system will serve you better on cost and operational simplicity. The NVL72 is a frontier instrument, not a default.

Facility readiness is non-negotiable

A GB200 NVL72 draws on the order of 120 kilowatts per rack and is liquid-cooled by design. Deploying one assumes a facility with the coolant distribution, power delivery, and structural capacity to match — these are not requirements you retrofit casually. The integration, validation, and serviceability of a rack-scale system also differ fundamentally from racking individual servers.

Nexus Compute scopes, configures, and warranty-backs Grace Blackwell systems against your data-center's real power and cooling capacity, and validates them before they ship. If you are evaluating rack-scale AI, share your facility specs and model targets and we will return a concrete, tested plan and quote within 48 business hours.