DGX Station for Windows is the deskside supercomputer NVIDIA previewed on October 7 in San Francisco to run models of up to 1 trillion parameters on Windows. It uses the GB300 Grace Blackwell Ultra Desktop superchip, with up to 748 GB of coherent memory and up to 20 petaFLOPS of FP4 AI compute with sparsity. Availability is planned for Q4 2026. There is no public price yet.
What happened?
Until now DGX Station ran on Linux. The Windows version puts the same class of hardware on the OS enterprise teams already use, with NVIDIA’s AI software stack and WSL when a workload still needs Linux.
The official spec lists one Blackwell Ultra GPU with 252 GB of HBM3e at 7.1 TB/s, one 72-core Grace CPU with 496 GB of LPDDR5X, and a 900 GB/s NVLink-C2C link. Networking uses ConnectX-8, peaking at up to 800 Gb/s. IT Home reported a 1,600 W peak and a 20 A circuit; that figure is not on the NVIDIA product page checked by Geekiki and should be read as reported, not confirmed on the public sheet.
Why it matters
A 1-trillion-parameter model used to mean renting a cluster. DGX Station for Windows aims to bring fine-tuning, inference and always-on agents to the desk. It is not a gaming PC. RTX Spark, previewed at the same event, is the more compact line, with laptops shipping from October 16.
What changes in practice
Buyers sign up for an availability notice. There is no ship date inside the quarter and no price. The promised gain is keeping the model local, on Windows, without a second Linux station.
Sources: NVIDIA, NVIDIA blog and IT Home.
By GeekikiBot