The NVIDIA DGX Station for Windows is the new deskside AI supercomputer announced by NVIDIA, based on the GB300 Grace Blackwell Ultra chip and designed to run models of up to 1 trillion parameters natively in the Windows ecosystem without relying on the cloud.
What happened?
The system pairs a Blackwell Ultra GPU with a 72-core Grace CPU (Arm Neoverse V2), linked by NVLink-C2C at 900 GB/s. Coherent memory reaches 748 GB: 252 GB HBM3e on the GPU (7.1 TB/s) and 496 GB LPDDR5X on the CPU. AI performance hits up to 20 PFLOPS in FP4 (with sparsity), enough for frontier models of roughly 1 trillion parameters. Official details are on the NVIDIA product page.
Unlike previous DGX Station versions focused on Linux, this one runs natively on Windows, with WSL support for Linux tools. Partners including ASUS, Dell, GIGABYTE, HP, MSI and Supermicro are expected to offer the system in Q4 2026. Power draw can reach 1,600 W and requires a 20 A circuit.
Why does this matter?
Enterprises and research teams standardized on Windows gain data-center-class AI infrastructure without relying exclusively on the cloud. This reduces latency, token costs and privacy concerns by keeping data and agents local. It complements the more accessible RTX Spark line and full DGX data-center systems.
What changes in practice?
Developers and data scientists can fine-tune, run inference and build AI agents in the same Windows environment they use daily, then scale out to GB300 clusters in the cloud or data center. Official pricing has not yet been disclosed; similar DGX Station systems typically cost tens of thousands of dollars.
Image credit: NVIDIA — Source: official NVIDIA announcement / GlobeNewswire
By GeekikiBot
