NVIDIA DGX Station runs trillion-parameter models locally
Key insights
- NVIDIA's GB300-powered DGX Station delivers 20 PFLOPS FP4 and 748GB memory, enabling 1-trillion-parameter inference on a single deskside unit.
- NVIDIA OpenShell, co-built with Microsoft, is an open-source agentic runtime providing sandboxed Windows app access for AI agents.
- Six OEM partners including Dell, HP, and ASUS are shipping DGX Station systems with deliveries beginning June 2026.
Why this matters
Local inference at trillion-parameter scale removes the latency, cost, and data-governance friction of cloud-dependent AI workflows, directly reshaping how enterprises budget and architect AI pipelines. NVIDIA's open-source agentic runtime with sandboxed Windows integration signals a bid to own the software layer of local AI compute alongside the hardware, positioning NVIDIA against both cloud providers and agent framework companies. OEM distribution through six major manufacturers means this capability enters standard enterprise procurement channels immediately rather than remaining a specialist research configuration.
Summary
NVIDIA's DGX Station for Windows moves what was until recently a data center workload onto a single desk. The GB300 Grace Blackwell Ultra Desktop Superchip delivers 20 petaflops of FP4 compute and up to 748GB of unified memory, enough to run models at the 1-trillion-parameter scale without a cloud dependency.
The system runs NVIDIA OpenShell on Windows, an open-source agentic runtime co-developed with Microsoft that gives AI agents sandboxed access to Windows applications. Dell, HP, ASUS, GIGABYTE, MSI, and Supermicro begin partner shipments this month.
Essentially: (NVIDIA, Microsoft) are making enterprise-grade local inference a standard workstation configuration.
- 20 PFLOPS FP4 and 748GB memory matches or exceeds many cloud inference nodes available today.
- OpenShell's isolated sandbox design lets agents act within Windows apps without unrestricted system access.
- Six major OEM partners shipping simultaneously signals a new hardware category, not a niche product.
The real shift is that air-gapped, fully local inference at frontier scale is now a procurement decision, not an infrastructure project.
Potential risks and opportunities
Risks
- Enterprise IT departments unprepared for 20-PFLOPS deskside hardware could create security and compliance gaps if DGX Stations are deployed without updated governance policies in the next 90 days.
- OpenShell's open-source agentic runtime, if vulnerabilities are discovered post-launch, could expose Windows environments to agent-level privilege escalation across all six OEM partner deployments simultaneously.
- Cloud inference providers (AWS, Azure, Google Cloud) face accelerating margin pressure as enterprises migrate frontier-model workloads to owned hardware over the next 12 to 18 months.
Opportunities
- Enterprise AI software vendors (Scale AI, Weights & Biases, Hugging Face) can target DGX Station buyers with local fine-tuning and model management tooling before cloud-native competitors adapt their positioning.
- Data-sensitive verticals including healthcare, legal, and defense contractors gain a clear procurement path for frontier AI that satisfies air-gap and data residency requirements, unlocking budget previously blocked by compliance constraints.
- NVIDIA's six OEM partners gain a premium workstation category commanding higher ASPs than traditional GPU workstations, with Dell and HP positioned to bundle lucrative multi-year enterprise support contracts.
What we don't know yet
- Pricing has not been disclosed; whether DGX Station for Windows falls within typical enterprise workstation budgets or requires capital procurement processes is unknown.
- Whether OpenShell's sandbox isolation has undergone independent security audit before partner shipments begin this month.
- Power draw and cooling requirements for the GB300 desktop superchip relative to standard office electrical infrastructure have not been specified in available reporting.
Originally reported by nvidianews.nvidia.com
Read the original article →Original headline: NVIDIA DGX Station for Windows Launches at GTC Taipei — GB300 Superchip, 20 PFLOPS FP4, Runs 1-Trillion-Parameter Models Locally, Partner Shipments Begin This Month