AIBRIEF
The Super AI Station pairs a 72 core Nvidia Grace CPU with a Blackwell generation B300 AI GPU, sharing 748GB of memory so enterprises can run trillion parameter AI models in house, with no cloud required.
Supermicro has begun selling Nvidia's first deskside GB300 workstation at $92,887.47 before tax and shipping, putting a small data-center AI node on an enterprise desk. The Super AI Station pairs Nvidia's Grace CPU Superchip, with 72 Arm cores, and a Blackwell B300 GPU inside a single coherent-memory platform that totals 748GB of RAM: 496GB of LPDDR5x on the Grace side and 252GB of HBM3e on the GPU.
Nvidia says the box is built to run trillion-parameter models locally and rates it at up to 20 PFLOPS of FP4 AI performance, with roughly 3x H200 inference throughput and up to 1,200 tokens per second for continuous agent workloads. The Supermicro spec sheet lists closed-loop direct-to-chip liquid cooling, dual 400GbE via ConnectX-8, and a deskside tower or 5U rack chassis.
To put the sticker in perspective, a maxed Mac Studio M5 Ultra with 256GB lists for $12,299, about seven times cheaper, but the two machines target different jobs. The Mac Studio is a creator workstation. The GB300 box is built for post-training, long-context inference, and always-on on-prem agents where data residency rules out the cloud.
Nvidia announced DGX Station for Windows at GTC Taipei, with Q4 delivery.