Supermicro begins selling Nvidia GB300 workstation for $92,887

TechRadar reports that the workstation carries 748GB of memory, split between 496GB of system memory attached to the 72-core Grace CPU and 252GB of HBM3e memory on the Blackwell Ultra GPU. This combined pool gives buyers nearly three times the addressable memory of Apple's Mac Studio with an M5 Ultra chip, which retails at $12,299 and offers 256GB of RAM. The price difference makes the Supermicro machine roughly seven times costlier than the Mac Studio.

Nvidia claims the platform delivers up to 20 PFLOPS of FP4 performance, tripling inference speeds compared with its H200 chip. The system also supports continuous agent workloads, generating as much as 1,200 tokens per second at no additional token cost. Storage includes two 1.92TB and two 960GB M.2 NVMe drives, with additional M.2 slots supporting PCIe 5.0 and PCIe 6.0 standards for further expansion. Networking runs through dual 400GbE ports paired with Nvidia's ConnectX-8 SuperNIC, alongside a dedicated management LAN port and IPMI support for remote administration.

Cooling relies on a direct-to-chip liquid system paired with fans mounted inside a tower or rack-mount chassis. The system draws power from a single 1600W supply rated at 94% Titanium-level efficiency, and the chassis weighs about 88 pounds unloaded. Supermicro lists a 3% discount for new customers on the Gold Series model, bringing the price down to roughly $90,101.05. Volume buyers can request quotes for larger discounts, a common practice in research and laboratory sales.

For most buyers, a device like this only makes sense if trillion-parameter models must run entirely on-site, without cloud reliance. Otherwise, the price gap compared with prosumer alternatives seems difficult to justify for typical research budgets outside well-funded organizations.