Available Now

NVIDIA GB300 NVL72 SYSTEMS, Ready to Deploy

Platform Overview

RACK-SCALE AI INFRASTRUCTURE

The NVIDIA GB300 NVL72 is a different class ofAI system. Rather than individual servers, it is a fully integrated, liquid-cooled rack that combines 72 Blackwell Ultra GPUs and 36 Grace CPUs intoa single compute domain. The entire rack operates as one unified system, connected by fifth-generation NVLink, and is purpose-built for large-scale AI training, reasoning, and high-throughput inference.

NVIDIA HGX B300 8-GPU baseboard render
NVIDIA Blackwell Architecture
Available Systems

NVIDIA GB300 NVL72 SYSTEMS FROM LEADING OEM PARTNERS

Choose from  validated GB300 NVL72 platforms designed for hyperscale AI deployments and  high-density data center environments.

Platform Specs

NVIDIA GB300 NVL72 RACK-SCALE SYSTEM

The GB300  NVL72 is a fully integrated, liquid-cooled rack that operates as a single unified AI system. It combines 72 Blackwell Ultra GPUs with 36 Grace CPUs,  connected by NVLink 5, to create one large compute domain capable of handling  the most demanding AI workloads in production today.

NVIDIA HGX B300 8-GPU baseboard render
Compute

72 Blackwell ULTRA GPUS

72 NVIDIA Blackwell Ultra GPUs in a single rack, operating as one unified compute domain for maximum training and inference throughput.

Interconnect

NVLINK 5 INTERCONNECT

Fifth-generation NVLink provides high-bandwidth, low-latency connectivity across all 72 GPUs, enabling the entire rack to function as a single compute node.

Processors

36 Grace CPUs

36 NVIDIA Grace CPUs tightly integrated with the GPU fabric, purpose-built to minimize CPU-GPU bottlenecks in AI workloads.

Reliability

LIQUID-COOLED DESIGN

Full rack liquid cooling designed for sustained, high-density operation. Required for the thermal envelope of 72 GPUs in a single rack.

Use Cases

Target workloads

Training

Large-Scale Model Training

Train frontier-scale foundation models that require massive, tightly coupled compute. The GB300 NVL72's unified 72-GPU domain eliminates the multi-node communication overhead that limits training throughput on traditional server-based architectures.

Inference

AI Reasoning & Inference at Scale

Deploy inference systems built for high-throughput AI reasoning workloads. The GB300 NVL72 delivers the memory capacity and compute density to serve large models with complex reasoning chains at production scale.

AI Factory

Hyperscale AI Factory Infrastructure

Build dedicated AI compute environments designed to run continuously at high utilization. The rack-scale architecture and liquid cooling are built for the sustained, dense operation that AI factory workloads demand.