Specifications
Configuration
NVIDIA GB300 NVL72 rack-scale AI
GPUs
72x NVIDIA B300 Blackwell Ultra, 288 GB HBM3e each
CPU
NVIDIA Grace, Arm Neoverse V2, 2,592 cores per rack
Pooled GPU Memory
Approximately 20.7 TB HBM3e
GPU-to-GPU Interconnect
NVIDIA NVLink 5 at 130 TB/s aggregate
NVLink Switch Trays
9 per rack, fully non-blocking
Cooling
Hybrid — liquid cooling on CPU/GPU, air for supporting components
Rack Platform
48U MGX rack
Power Consumption
120 kW to 140 kW per rack
Scale-Out Networking
NVIDIA ConnectX-8 SuperNIC at 800 Gb/s per port
South Networking
NVIDIA BlueField-3 DPU
Scale-Up / Scale-Out Fabric
Quantum-X800 InfiniBand or Spectrum-X Ethernet
Internal Drive Options
E1.S NVMe including Samsung PM9D3a 7.68 TB E1.S 9.5 mm SED
Operating Systems
Ubuntu 22.04 with NVIDIA optimized HWE kernel
Red Hat Certification
RHEL 9.6 (vendor certified), OpenShift 4.19, AI Inference Server 3.3
Software Stack
NVIDIA AI Enterprise
Overview
Lenovo's GB300 NVL72 is the Blackwell Ultra generation of the rack-scale NVL72 architecture: 72 B300 GPUs at 288 GB of HBM3e apiece and 36 Grace CPUs, delivering roughly 20.7 TB of pooled high-bandwidth memory and 130 TB/s of NVLink bandwidth inside a single 48U MGX cabinet.
Lenovo uses a hybrid cooling design in which the CPU and GPU cold plates carry the bulk of the thermal load while supporting components remain air-cooled - a smaller CDU and manifold envelope than a fully liquid rack, which matters for retrofit into existing halls.
The scale-out fabric is ConnectX-8 SuperNIC at 800 Gb/s per port with BlueField-3 DPUs for north-south offload, running over either Quantum-X800 InfiniBand or Spectrum-X Ethernet. Lenovo ships the stack with NVIDIA AI Enterprise and Red Hat certification for RHEL 9.6, OpenShift 4.19 and AI Inference Server 3.3.
Key Benefits
Hybrid cooling: liquid cold plates on the hot silicon only, reducing CDU capacity and plumbing complexity relative to a fully liquid rack. Validated software stack: AI Enterprise plus Red Hat certification across RHEL, OpenShift and the inference server shortens time-to-first-workload. Enterprise storage integration: E1.S NVMe options including 7.68 TB SED drives are factory-validated in the rack.
Applications
Enterprise and sovereign AI training, reasoning-model inference, retrieval-augmented generation at trillion-parameter scale, HPC simulation, and on-premises AI factories that require a vendor-validated enterprise software stack.
Request a Quote — Lenovo GB300 NVL72 Rack-Scale AI — 48U MGX Hybrid-Cooled Rack
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →