NVIDIA GB300 NVL72 Supercluster: Inside the 8-Rack Containerised AI Factory Node

The NVIDIA GB300 NVL72 supercluster is a containerised AI factory node that packs eight GB300 NVL72 racks into a single pre-integrated module — 576 Blackwell Ultra B300 GPUs and 288 Grace CPUs delivered, powered, cooled, networked, and commissioned as one unit. Rather than assembling superchips, boards, power, and fabric on-site over many months, the node arrives as a high-cube shipping module that hooks into grid and network and runs. This reference explains what is inside the node, how it behaves as a single accelerator, and where it fits for sovereign-AI, hyperscale, and enterprise AI-factory buildouts.

What the GB300 NVL72 supercluster node is

The node is the AI-factory building block above a standalone GB300 superchip or a single NVL72 rack. Traditional piecemeal architecture — sourcing boards, power, cooling, and switching separately and integrating them on the data-center floor — does not scale to next-generation training and inference. The GB300 NVL72 supercluster answers that by shipping a completely integrated, production-ready supercomputer as one turnkey node. Eight liquid-cooled GB300 NVL72 racks are mounted inside a single high-cube container along with power distribution, cooling, storage, and fabric, so the customer receives compute capacity rather than a parts list.

Compute core: 576 Blackwell Ultra GPUs as one accelerator

Each node contains 576 NVIDIA Blackwell Ultra B300 GPUs and 288 Grace CPUs across its eight NVL72 racks. Inside the node the GPUs are joined by an NVLink domain delivering 1,040 TB/s of aggregate bandwidth, which lets all 576 GPUs operate as a single coherent accelerator with effectively zero-bottleneck communication. That coherence is what makes trillion-parameter models practical: model state and activations move across the domain at NVLink speed instead of being fragmented across slower inter-node links.

Node-level unified memory

The node exposes roughly 165.6 TB of HBM3e alongside about 320 TB of fast system memory. This large, tightly-coupled memory pool is what allows very large models and long context windows to stay resident rather than being paged, which is the difference between a supercluster that can hold a frontier model and one that constantly stalls on data movement.

Performance: engineered for agentic AI

Per node, the GB300 NVL72 supercluster delivers approximately 11.5 EFLOPS of peak FP4 (sparse) compute and around 8.8 EFLOPS dense. These are sustained figures targeted at real-time, trillion-parameter reasoning and next-generation agentic AI, where responsiveness under load matters as much as raw peak throughput. Against the prior Hopper generation, NVIDIA positions the platform at roughly 50× the overall AI-factory output, 10× faster user responsiveness, and 5× more throughput per watt.

Specification	Per containerised node
Racks	8× GB300 NVL72
GPUs	576× Blackwell Ultra B300
CPUs	288× Grace
GPU memory	~165.6 TB HBM3e
System memory	~320 TB fast system memory
NVLink domain bandwidth	1,040 TB/s aggregate
Peak compute (FP4 sparse)	~11.5 EFLOPS
Dense compute	~8.8 EFLOPS
Cooling	Warm-water direct liquid cooling, in-row CDU up to 1.8 MW
Power	64× 33 kW shelves, redundant busbars + BMS
Fabric	ConnectX-8 (800 Gb/s), BlueField-3 DPUs; Quantum-X800 InfiniBand or Spectrum-X Ethernet
Software	NVOS + NVIDIA AI Enterprise (576 GPU subscriptions), Mission Control, DOCA

Critical infrastructure: power, cooling, and the data spine

Thermal and power delivery are built into the node. Cooling is warm-water direct liquid cooling (DLC) with an in-row coolant distribution unit rated for up to 1.8 MW of thermal dissipation. Power uses a heavy-duty redundant architecture of 64× 33 kW shelves with integrated busbars and comprehensive safety/BMS features. The data spine is a high-speed fabric built on ConnectX-8 (800 Gb/s) NICs and BlueField-3 DPUs, configurable with either Quantum-X800 InfiniBand or Spectrum-X Ethernet, and high-speed NVMe storage arrays are integrated directly with the internal management switches.

The orchestration and software stack

The node ships as a full stack, not bare metal. At the foundation, NVOS (the NVIDIA operating system) manages the hardware layer directly. Above it, NVIDIA AI Enterprise is included with 576 individual GPU subscriptions, and fleet orchestration is handled by Mission Control and DOCA services. This is what lets a strategic partner treat the node as a managed AI factory rather than a rack of servers to hand-configure.

Deployment: turnkey site integration

Because the node is a pre-integrated high-cube module, it bypasses years of traditional data-center construction. Commissioning follows three stages: the module arrives pre-integrated and containerised; a plug-and-play architecture is designed for immediate grid and network hookup; and handover includes comprehensive training and detailed operations-and-maintenance documentation. The result is time-to-first-token measured against logistics and grid readiness rather than multi-year build cycles.

Where it fits

The GB300 NVL72 supercluster is optimised for large-scale training, post-training alignment, real-time inference, and agentic reasoning at trillion-parameter model scale. NVIDIA positions it for sovereign-AI initiatives, hyperscale deployments, top-tier research institutions, and enterprise AI factories. For India-based sovereign-AI programs in particular, the containerised, self-contained form factor supports on-premises and colocation deployment where data residency and in-country control are requirements. Availability is a practical constraint: this is premium, allocation-limited inventory prioritised for strategic partners, so lead time and eligibility are part of the planning conversation, not an afterthought.