Microsoft announces Windows-based DGX Spark and DGX Station for local AI

Microsoft announced the NVIDIA DGX Spark and DGX Station for Windows, targeting local execution of large artificial-intelligence models.

DGX Spark: 1 PFLOPS in a compact system

DGX Spark uses NVIDIA’s GB10 Grace Blackwell Superchip, combining a 20-core ARM processor with a Blackwell GPU. It provides 128 GB of unified memory at 273 GB/s and up to 1 PFLOPS of performance, enabling local execution of models with more than 200 billion parameters.

The device weighs just over 1 kg and consumes 240 W. It supports AI agents, generative models and computer-vision workloads; NVIDIA’s software stack supports FLUX.1, NVIDIA Cosmos Reason and a Spark-optimized Qwen3 chatbot. Its price is stated to be more than 32 times lower than that of DGX-1.

DGX Station: 20 PFLOPS and 784 GB

DGX Station is a workstation built around NVIDIA’s GB300 Grace Blackwell Ultra Desktop Superchip. It offers 784 GB of unified memory and 20 PFLOPS in FP4, with support for local trillion-parameter models.

The system weighs 40 kg and consumes up to 1,600 W. Its memory combines 496 GB of DDR5 attached to the 72-core Grace processor with 252 GB of HBM3e on the Blackwell Ultra GPU.

Storage includes two 1.92 TB SSDs and two 960 GB SSDs, with PCIe 5.0 and PCIe 6.0 expansion slots. Networking includes two 400GbE ports with a ConnectX-8 SuperNIC and a separate LAN port for remote IPMI administration.

Cooling combines direct liquid chip cooling with fans. The 1,600 W power supply is Titanium-certified with a stated 94% efficiency; the ConnectX-8 adapter supports speeds of up to 800 Gbps for linking multiple stations.

DGX Station is priced at $92,900, described in the source as seven times the price of a Mac Studio.