Specifications
GPU
1× NVIDIA Blackwell Ultra GB300
GPU Memory
252GB HBM3e, 7.1 TB/s
CPU
Grace 72-core Neoverse V2
CPU Memory
496GB LPDDR5X, 396 GB/s
Unified Memory
748GB total (GPU + CPU coherent)
NVLink-C2C
900 GB/s GPU-CPU interconnect
FP4 Performance
20 PFLOPS
Networking
ConnectX-8 SuperNIC, up to 800Gb/s Ethernet
Power
~1.6kW, standard outlet
Cooling
Air-cooled, deskside
OS Support
Ubuntu Linux / Windows
Availability
Shipping June 2026 via ASUS, Dell, HP, Gigabyte, MSI, Supermicro
Overview
The NVIDIA DGX Station GB300 brings data-center-class AI compute to the deskside. Powered by a single Grace Blackwell Ultra GB300 Superchip with 748GB of coherent unified memory (252GB HBM3e + 496GB LPDDR5X), it delivers 20 PFLOPS of FP4 AI performance — enough to run frontier LLMs, fine-tune models, and serve AI agents without a data center. No liquid cooling, no colocation — just plug into a standard outlet (1.6kW) and start building. Available through OEM partners including ASUS, Dell, HP, Gigabyte, MSI, and Supermicro at approximately $90K-$115K.
Key Benefits
748GB Unified Memory: Largest coherent memory footprint in a desktop form factor — load entire 70B+ parameter models without quantization or offloading. Deskside AI Factory: Train, fine-tune, and serve LLMs locally with full data privacy and zero cloud dependency. ConnectX-8 Networking: 800Gb/s Ethernet enables direct integration with enterprise storage clusters and multi-node scaling. Windows + Ubuntu: Dual-OS support — run NVIDIA AI Enterprise on Ubuntu or integrate with Windows-based enterprise toolchains. No Data Center Required: Air-cooled, 1.6kW, fits under a desk — deploy in research labs, secure facilities, and branch offices.
Request a Quote — NVIDIA DGX Station GB300
QS Compute — NVIDIA Goldendisk Partner. Available through authorized OEM channels (ASUS, Dell, HP, Gigabyte, MSI, Supermicro).
Get Your Quote →