NVIDIA Announces DGX Station for Windows to Run Trillion-Parameter AI Models Locally
NVIDIA has announced the DGX Station for Windows, a desktop AI supercomputer equipped with the GB300 Grace Blackwell Ultra Desktop Superchip capable of running up to 1-trillion-parameter models locally. Scheduled for launch in Q4 2026, the machine delivers up to 20 PFLOPS of FP4 AI performance alongside 748GB of unified memory. By bringing data-center-class hardware and massive high-speed memory to a Windows desktop environment, developers can fine-tune and serve large models locally without relying on cloud infrastructure. This reduces latency, cloud expenses, and data privacy risks while accelerating AI agent development and local prototyping. The system combines a 72-core Grace CPU and a Blackwell Ultra GPU, providing 748GB of unified memory split between 252GB of ultra-fast HBM3e VRAM and 496GB of LPDDR5X system RAM. Operating as a heavy-duty workstation, it draws up to 1,600W of power and requires a 20A circuit.
## BACKGROUND
Running massive artificial intelligence models with hundreds of billions or trillions of parameters typically requires multi-node GPU clusters housed in specialized enterprise data centers due to strict VRAM capacity constraints. High Bandwidth Memory (HBM) addresses data bottlenecks by vertically stacking dynamic RAM modules directly onto the processor substrate, providing the multi-terabyte-per-second bandwidth essential for modern AI workloads.