NVIDIA announced DGX Station for Windows on May 31, 2026, and expects ASUS, Dell Technologies, GIGABYTE, HP, MSI and Supermicro to offer the workstation during Q4 2026. Built around the GB300 Grace Blackwell Ultra Desktop Superchip, it pairs Microsoft Windows with access to Linux AI toolchains through WSL.
What Windows changes on the GB300 platform
The system combines a Blackwell Ultra GPU with a 72-core Grace CPU based on Neoverse V2. NVIDIA lists Microsoft Windows as the operating system for this configuration and says Windows Subsystem for Linux provides access to Linux AI toolchains. That gives teams using Windows a route to those tools within the workstation’s stated software setup.
NVIDIA frames the system for enterprise applications and workflows, including work with AI agents. The main distinction described for this configuration is its Windows environment and integration—not a separately announced processor generation.
What 748 GB and 20 PFLOPS mean
The advertised memory total is a coherent pool of up to 748 GB: 252 GB of HBM3e GPU memory plus 496 GB of LPDDR5X CPU memory. NVIDIA lists bandwidths of 7.1 TB/s for the GPU memory and 396 GB/s for the CPU memory, with NVLink-C2C rated at 900 GB/s.
NVIDIA lists peak FP4 Tensor Core performance of up to 20 PFLOPS with sparsity and 15 PFLOPS without it. A PFLOP per second represents a quadrillion floating-point operations each second. The sparsity-enabled figure takes advantage of zero-valued data where the workload supports that computation; it is a peak specification, not an application-level speed figure.
NVIDIA also advertises support for models with up to 1 trillion parameters. That is a model-scale claim, distinct from a stated generation speed or context length. The product page lists total system power at 1,600 W and specifies a 20 A circuit. It also describes configurations with up to one additional NVIDIA RTX PRO Blackwell-generation GPU, with support varying by system.
DGX Station for Windows vs. DGX Spark
NVIDIA’s product specifications distinguish the GB300 workstation from DGX Spark’s GB10 platform. The published memory, operating-system and model-scale details are:
| Product | Compute platform | Advertised memory | Operating system listed | Advertised model scale |
| DGX Station for Windows | GB300 Grace Blackwell Ultra Desktop Superchip | Up to 748 GB coherent memory | Microsoft Windows; Linux AI toolchains through WSL | Up to 1 trillion parameters |
| DGX Spark | GB10 Grace Blackwell Superchip | 128 GB unified memory | NVIDIA DGX OS | Workloads up to 200 billion parameters |
For readers following NVIDIA’s broader local-AI lineup, NeoTeo previously covered the push to bring local AI from the desk into the home. The comparison here highlights the different platforms and the operating environments NVIDIA lists for each system.
