NVIDIA unveils local AI hardware at GTC Berlin

Summary

NVIDIA showcased its local AI hardware stack at GTC Berlin: RTX Spark for consumer PCs, DGX Spark for developer clusters, and DGX Station for the enterprise segment. Each product is accompanied by new software integrations, model support, and steps toward on-premise agentic workflows.

Why it matters

The strategic bet is this: Enterprise customers and developers will increasingly want to perform edge-class inference without sending data to the cloud. NVIDIA is attempting to own this entire local stack, from silicon to orchestration. If clustering, model launchers, and enterprise management layers work as claimed, they could change how mid-sized teams and manufacturers deploy AI, reducing reliance on hyperscaler APIs.

Highlights

• Eight DGX Sparks can be clustered into a node with 1 TB of unified memory. This is sufficient to run edge models like GLM 5.3 locally.
• DGX Station for Windows (December) will integrate production applications from Siemens, Dassault, MathWorks, and others with on-device LLM agents.
• NVIDIA Cloud Functions will allow a DGX Station to be used as the team's shared, OpenAI-compatible local inference endpoint. It can be swapped in without code changes.

Critical take

This is a vendor blog, not news: 'Reviews' consist of selected single-sentence quotes. About half of the announcements are 'coming this month' or in 'early access' stages. The clustering presentation (requiring eight units to match one DGX Station) silently reveals how far a single Spark is from real edge workloads. The local AI narrative is compelling; however, the article conflates released products with the December roadmap. It is difficult to distinguish what can actually be purchased today from a presentation slide.