NVIDIA GB300 NVL72 Grace Blackwell Ultra Rack System

The NVIDIA GB300 NVL72 Grace Blackwell Ultra Rack System is a full-rack AI supercomputer built around the GB300 Grace Blackwell Ultra architecture, integrating 72 NVIDIA Blackwell Ultra GPUs and 36 Grace Blackwell Ultra Superchips into a single, cohesive rack-scale unit. Each GB300 Superchip combines an NVIDIA Grace CPU with a Blackwell Ultra GPU die interconnected via NVLink-C2C, delivering extreme memory bandwidth and unified CPU-GPU memory addressability that eliminates traditional PCIe bottlenecks. The system is interconnected via NVLink 5 across all 72 GPUs, creating a unified 72-GPU fabric with aggregate GPU memory that scales to support the largest frontier AI models in a single rack domain.

Designed for the most demanding AI training and inference workloads, the GB300 NVL72 delivers substantial improvements in FP4 and FP8 tensor core throughput relative to prior Hopper-generation systems, enabling enterprises to train and serve large language models and multimodal foundation models at scale with reduced total infrastructure footprint. The system ships as a complete rack-scale reference design incorporating liquid cooling infrastructure, NVLink Switch systems, high-bandwidth networking readiness, and power delivery components rated for datacenter integration, allowing operators to deploy immediately within compatible datacenter environments.

The GB300 NVL72 is positioned as an enterprise and hyperscale datacenter platform targeting AI research institutions, cloud service providers, sovereign AI initiatives, and large enterprises building proprietary foundation models or deploying high-throughput inference infrastructure. Its NVLink-based unified GPU fabric enables tensor-parallel and pipeline-parallel model execution across all 72 GPUs without leaving the rack, making it particularly well suited to workloads that exceed the memory capacity of any single GPU or node.

Ideal for

• Large-scale foundation model pre-training across hundreds of billions to trillions of parameters using full 72-GPU NVLink tensor parallelism within a single rack
• High-throughput generative AI inference serving for large language models and multimodal models requiring low latency at enterprise production scale
• Sovereign AI and national AI infrastructure deployments requiring maximum compute density with a reduced datacenter footprint
• Scientific simulation and HPC workloads combining CPU-side data preprocessing on Grace cores with GPU-accelerated numerical solvers across the full Blackwell Ultra fabric
• Enterprise AI platform consolidation, replacing multiple prior-generation GPU nodes with a single rack-scale system to simplify networking, management, and power infrastructure
• Retrieval-augmented generation and real-time AI inference pipelines requiring large unified GPU memory pools to hold multi-hundred-billion-parameter models resident in memory

Technical specifications

Available from Omnixon Global. Submit an RFQ and our team will confirm configuration and availability for your order.