Desktop supercomputer goes home
Karpathy‑level hardware is mainstreaming: the DGX station GB300 targets desktop supercomputing with ~748GB of memory and ~20 petaflops of compute, built to support 1‑trillion‑parameter models and priced around $200K–$300K. That changes the calculus for teams that want to run large models locally for sensitive data or low‑latency analytics. (x.com)
NVIDIA’s DGX Station GB300 is now available to order through OEM partners including ASUS, Dell, GIGABYTE, MSI and Supermicro, with vendors listing systems on their commercial channels this week. (aihola.com) The GB300 package pairs a 72‑core Arm “Grace” CPU with a Blackwell Ultra GPU and uses NVIDIA’s NVLink‑C2C fabric and Blackwell‑generation tensor cores that enable NVFP4 (4‑bit floating‑point) AI execution. (nvidia.com) Hardware partners are splitting memory between high‑bandwidth HBM3e on the GPU and LPDDR5X on the CPU — independent reports show GPU HBM and CPU LPDDR capacities of roughly 252GB and 496GB respectively and the use of HBM3e to feed large models. (guru3d.com) Market listings for the same DGX Station architecture vary widely: an Exxact reseller configurator shows a configured price of $87,411.50, while TechRadar and other outlets have reported MSI‑branded listings around $85,000 and some OEM writeups cited figures near $97,000. (configurator.exxactcorp.com) NVIDIA first demonstrated the DGX Station concept at its GTC event last year, and the recent wave of OEM product announcements and press releases began on March 16–17, 2026 as partners moved from prototypes to commercial availability. (tomshardware.com) MSI’s XpertStation WS300 and other branded DGX Station builds advertise enterprise features such as dual high‑speed QSFP ports (400GbE in some SKUs) and integration with NVIDIA’s DGX software and ConnectX‑8 networking for datacenter connectivity. (prnewswire.com)
