NVIDIA GB300 NVL72
A liquid-cooled, rack-scale system that joins 72 Blackwell Ultra GPUs and 36 Grace CPUs in a single NVLink domain. NVIDIA lists it as available now.
NVIDIA · Current generationShipping
GB300
Blackwell Ultra · rack-scale NVL72
- Memory
- 279 GB HBM3E
- Bandwidth
- 8 TB/s
- Rack
- 72 GPUs · 36 CPUs
Specifications at a glance
As published by NVIDIA, checked 24 September 2026. Configurations vary by system builder.
- System
- 72 Blackwell Ultra GPUs and 36 Grace CPUs in one rack, with one 72-GPU NVLink domain
- Cooling
- Liquid-cooled
- GPU memory
- 279 GB HBM3E per GPU; 20 TB HBM3E per rack
- Memory bandwidth
- 8 TB/s per GPU; up to 576 TB/s per rack
- FP4 compute
- 1,440 PFLOPS with sparsity, 1,080 PFLOPS dense, per rack
- Interconnect
- Fifth-generation NVLink: 1.8 TB/s per GPU, 130 TB/s per rack; ConnectX-8 at 800 Gb/s per GPU
- Power
- Configurable up to 1,400 W per GPU; rack power not published
- Status
- Shipping; NVIDIA's product page lists it as available now
Configuration and site considerations
A GB300 NVL72 arrives as a complete rack, so the review looks at the site as closely as the system.
- Plan for a fully liquid-cooled rack and the facility cooling it needs.
- Confirm rack power and distribution with the system builder; NVIDIA publishes per-GPU power only.
- Document the number of racks, the system builder and the delivery schedule.
Prepare this system for a financing review
Document the configuration, the supplier and the deployment so the request can be assessed against the actual equipment.**
System configuration and quantity
Supplier and purchase order
Or fleet details and ownership information for deployed servers.
Deployment site and colocation agreement
Power and cooling readiness
Customer contract or documented on-demand revenue
