NVIDIA NVL72 AI Factory
Type: Reference Architecture Tags: NVIDIA, NVL72 AI Factory, Enterprise Reference Architecture, GB300 NVL72, Blackwell Ultra, Spectrum-X, NVLink, Mission Control Related: NVIDIA-Enterprise-Reference-Architectures, NVIDIA-Enterprise-AI-Factory, NVIDIA-GB300-NVL72, NVIDIA-DGX-B300, NVIDIA-DGX-SuperPOD, NVIDIA-Blackwell-Architecture, NVIDIA-AI-Enterprise, NVIDIA-AI-Enterprise-Software-Reference-Architecture, NVIDIA-Enterprise-RA-Observability-Guide, NVIDIA-Mission-Control, NVIDIA-Dynamo, NVIDIA-Run-ai, NGC, NVIDIA-NetQ, NVIDIA-Spectrum-X, NVIDIA-BlueField-DPU, NVLink, GPUDirect-RDMA, GPU-Direct-Storage Sources: https://docs.nvidia.com/enterprise-reference-architectures/index.html, https://docs.nvidia.com/enterprise-reference-architectures/nvl72-ai-factory-with-gb300-nvl72-dual-plane-networking-architecture.pdf Last Updated: 2026-04-29
Summary
NVIDIA NVL72 AI Factory is the Enterprise Reference Architecture for GB300 NVL72 rack-scale AI factories. The current NVIDIA PDF describes a GB300 NVL72 and Spectrum-X architecture for massive enterprise AI training and inference, with each scalable unit containing 18 compute trays/nodes and 72 Blackwell Ultra GPUs.
Detail
Purpose
The NVL72 AI Factory pattern is for the most intensive enterprise AI workloads: large-scale foundation model training, fine-tuning, real-time reasoning, and complex agentic AI pipelines. It turns GB300 NVL72 racks into modular building blocks for AI factory data centers.
Architecture notes
- The RA uses NVIDIA-GB300-NVL72, BlueField-3 DPUs, ConnectX-8 SuperNICs, Spectrum-X Ethernet, AI Enterprise software, Mission Control, and NVLink/NVLink Switch.
- One scalable unit has 18 trays/nodes and 72 GPUs.
- The document includes 2-rack, 4-rack, and 8-rack designs with 144, 288, and 576 Blackwell Ultra GPUs.
- The compute east-west fabric is separated and rail-optimized for high-bandwidth, low-latency GPU traffic.
- Dual-plane Spectrum fabrics are used for high availability, resiliency, and load balancing.
- Storage guidance uses the same 12.5 Gb/s per GPU planning signal that appears in other Enterprise RAs.
Software stack
The NVL72 RA includes NVIDIA-AI-Enterprise, NVIDIA-Mission-Control, NVIDIA-Dynamo, NVIDIA-Run-ai, NGC, and NVIDIA-NetQ. Treat NVIDIA-Mission-Control as the operations-plane page for NVL72 deployment and lifecycle management.
NVIDIA context
Use this page for the Enterprise RA document and design pattern. Use NVIDIA-GB300-NVL72 for the system identity, NVIDIA-DGX-B300 for the DGX branded product path, and NVIDIA-Blackwell-Architecture for GPU architecture context.
Connections
- NVIDIA-Enterprise-Reference-Architectures - program-level RA hub.
- NVIDIA-Enterprise-AI-Factory - strategic AI factory planning context.
- NVIDIA-GB300-NVL72 - rack-scale Blackwell Ultra system used in this RA.
- NVIDIA-DGX-B300 - DGX system path adjacent to GB300 NVL72 deployments.
- NVIDIA-DGX-SuperPOD - large-scale AI factory and SuperPOD context.
- NVIDIA-Blackwell-Architecture - Blackwell Ultra architecture foundation.
- NVIDIA-AI-Enterprise - enterprise software suite included in the software stack.
- NVIDIA-AI-Enterprise-Software-Reference-Architecture - common AI Enterprise software pattern.
- NVIDIA-Enterprise-RA-Observability-Guide - observability guidance for Enterprise RAs.
- NVIDIA-Mission-Control - operations platform for NVL72 deployments.
- NVIDIA-Dynamo - open-source inference orchestration component included in the RA software table.
- NVIDIA-Run-ai - workload orchestration layer.
- NGC - GPU-optimized containers and AI/HPC artifacts.
- NVIDIA-NetQ - network operations and validation.
- NVIDIA-Spectrum-X - Ethernet networking platform for compute and converged connectivity.
- NVIDIA-BlueField-DPU - DPU role for infrastructure, networking, and storage paths.
- NVLink - defining rack-scale GPU interconnect for NVL72.
- GPUDirect-RDMA - direct GPU-network data path.
- GPU-Direct-Storage - direct storage-to-GPU data path for data-heavy AI workloads.
Source Excerpts
- NVIDIA’s Enterprise RA hub describes NVL72 AI Factory as the GB300 NVL72 rack-scale building block for the most intensive enterprise AI workloads.
- The March 2026 PDF lists GB300 NVL72, BlueField-3, ConnectX-8, Spectrum-X, AI Enterprise, Mission Control, and NVLink/NVLink Switch as core technologies.