72× B300 GPUs + 36× Grace CPUs · ~1.4 EFLOPS FP4 · Liquid-Cooled · NVLink 72
72× B300 (Blackwell Ultra)
36× NVIDIA Grace
Up to 20.7 TB
~1.4 EFLOPS FP4 (inference)
NVLink 72, all-to-all
130 TB/s per rack
Direct liquid cooling
~1.36 t (3,000 lb)
42U rack-scale system
Frontier AI training & inference
The NVIDIA GB300 NVL72 is the Blackwell Ultra rack-scale system, interconnecting 72 B300 GPUs and 36 Grace CPUs over a fifth-generation NVLink 72 all-to-all fabric. With up to 20.7 TB of HBM3e and ~1.4 EFLOPS of FP4 inference compute, a single cabinet behaves like one giant GPU for trillion-parameter model training.
The 42U cabinet is direct-liquid-cooled, weighs approximately 1.36 tonnes, and collapses what previously required multiple racks of interconnected nodes into a single, coherent NVLink domain. It is the reference platform for Blackwell Ultra-era exascale AI clusters.
QS Compute can procure and deploy GB300 NVL72 cabinets and sub-rack GB300 HGX configurations for customers building large-scale AI infrastructure across APAC.
Need NVIDIA GB300 NVL72?
Enterprise hardware — configured, tested, deployed. Contact us for pricing and availability.
Request Quote