NVIDIA GB200 NVL72 GPU Infrastructure & Deployment_
72 Blackwell GPUs per rack at 120 kW with mandatory liquid cooling. The GB200 NVL72 represents the first generation of NVIDIA's Blackwell architecture at rack scale.
What Is the NVIDIA GB200?_
The NVIDIA GB200 NVL72 is the first Blackwell-architecture GPU system designed for rack-scale AI training. Each rack contains 72 GPUs with 192 GB HBM3e per GPU, connected via NVLink 5.0 for massive inter-GPU bandwidth within the rack.
At approximately 120 kW per rack with mandatory 100% direct liquid cooling, the GB200 NVL72 requires purpose-built data center infrastructure. Traditional air-cooled facilities cannot support this platform without significant upgrades to power distribution and cooling systems.
Technical Specifications_
| Specification | GB200 NVL72 |
|---|---|
| Architecture | Blackwell |
| GPU Memory | 192 GB HBM3e per GPU |
| GPUs per Rack | 72 (NVL72 configuration) |
| TDP per Rack | ~120 kW |
| Rack Weight | ~1,360 kg (3,000 lbs) |
| Interconnect | NVLink 5.0, InfiniBand NDR |
| Networking | 400GbE / 800GbE |
| Cooling | 100% Direct Liquid Cooling (mandatory) |
Deployment Considerations_
120 kW per rack requires dedicated high-voltage feeds. Standard 208V circuits are insufficient. Plan for 480V 3-phase distribution with bus plugs or overhead busway.
400GbE/800GbE networking with InfiniBand NDR fabric. High-count MPO/MTP fiber trunks required for the dense interconnect topology.
100% direct liquid cooling is mandatory. No air-cooled option exists. Facility must provide CDUs, chilled water loops, and rack-level manifolds.
Related Resources
Ready to Deploy Your GPU Infrastructure?_
Tell us about your project. We’ll respond within 48 hours with a scope assessment and timeline estimate.