RACK-SCALE SYSTEM
NVIDIA GB300 NVL72
Blackwell Ultra GPUs
- 36 Grace CPUs
- 20 TB GPU memory per rack
- NVLink 5 · CUDA
- Whole rack or defined provider allocation
Home / Computing / Rack-scale AI
Rack-scale AI
Integrated GPU, CPU and networking systems for demanding AI workloads. Explore whole-rack systems or defined allocations where a provider offers them.

Rack-scale options
Your proposal defines the allocation, resource isolation, operating location and delivery window for the selected system.
72 GPUs
Logical GPU count.RACK-SCALE SYSTEM
Blackwell Ultra GPUs
RACK-SCALE SYSTEM
Rubin GPUs
RACK-SCALE SYSTEM
Instinct GPUs
Swipe to see every specification on a small screen.
| Platform | GPUs | CPU / reference design | GPU memory | Interconnect / software |
|---|---|---|---|---|
| NVIDIA GB300 NVL72 | 72 Blackwell Ultra GPUs | 36 Grace CPUs | 20 TB GPU memory per rack | NVLink 5 · CUDA |
| NVIDIA Vera Rubin NVL72 | 72 Rubin GPUs | 36 Vera CPUs | 20.7 TB HBM4 per rack | NVLink 6 · CUDA |
| AMD Helios | 72 Instinct GPUs | MI455X / MI450X reference designs | Up to 31 TB HBM4, variant-dependent | UALink architecture · ROCm |
Manufacturer reference architectures. The selected OEM build, allocation, software support and delivery are established in the proposal. Maximum memory depends on the GPU variant.
WHAT YOU BUY
Your proposal states the GPU count and which CPU, memory, networking and storage resources are dedicated or shared. Access follows the selected platform’s operating model.
Understand the service scopePower delivery, liquid cooling and heat rejection designed for the actual system.
Connections beyond the rack, data movement and storage for the workload.
System tests, any agreed customer benchmark, and a supplier-backed delivery window.
Rack or allocation pricing
Quoted for the configuration, operating location, service scope and term you choose. Your offer identifies the delivery window and any upfront payment.
Bring your scale and target start. We can explore a whole rack, a defined allocation or a phased plan where available.