NVIDIA launches DGX Station desktop supercomputer with Nemotron 3 Ultra, a 550-billion-parameter open model

At SIGGRAPH 2026 (running through July 23 in Los Angeles), NVIDIA unveiled DGX Station, a desktop supercomputer built around the GB300 Grace Blackwell Ultra Desktop Superchip, paired with Nemotron 3 Ultra — a new open-weight model — and NemoClaw, a set of open blueprints for building autonomous agents.

What's new

• DGX Station hardware: the GB300 Grace Blackwell Ultra Desktop Superchip delivers 20 petaflops of FP4 AI compute with 748GB of coherent memory in a desktop form factor. NVIDIA says the system is "built and available to order" through manufacturing partners including ASUS, Dell, HP, and MSI.
• Nemotron 3 Ultra: NVIDIA describes it as "a frontier 550-billion-parameter open model" that "is optimized to run on DGX Station," serving as the reasoning backbone for agents built on the system and available under an open-source license for developers to fine-tune.
• NemoClaw: open blueprints that package the model, an agent harness, and a runtime together as a starting point for custom, domain-specific agents. NVIDIA says setup "takes just three steps, and the system can be running in roughly 30 minutes."
• The stack also bundles NVIDIA OpenShell, an open-source sandboxed runtime for keeping agents governed, and Omniverse libraries exposed as agent-accessible tools.
• Partner example: Nous Research has already customized Nemotron 3 Ultra for its Hermes Agent harness.
• Resources and ordering are available now at build.nvidia.com/station.

Context

DGX Station is NVIDIA's latest entry in its DGX line aimed at bringing datacenter-class AI compute to a desktop footprint, and Nemotron 3 Ultra extends NVIDIA's open Nemotron model family to frontier scale. The launch was one of several SIGGRAPH 2026 announcements from NVIDIA, alongside creative-tool integrations via the Model Context Protocol and a new Synthetic Video Detector NIM microservice for identifying AI-generated video.

Why it matters

Bundling a 550B-parameter open-weight model with turnkey desktop hardware and ready-made agent blueprints lowers the barrier for developers and enterprises to run and customize frontier-scale reasoning models on-premises rather than exclusively through cloud APIs — a meaningful option for teams with data-sovereignty, latency, or cost constraints on agentic workloads. The emphasis on a three-step, 30-minute setup also signals NVIDIA's push to productize "agent infrastructure" as a single packaged stack — silicon, model, runtime, and blueprint together — rather than components enterprises must assemble themselves.