NVIDIA ships the first DGX Station GB300 system to Andrej Karpathy’s lab

NVIDIAAIDev said on X on March 18, 2026 that Andrej Karpathy’s lab has received the first DGX Station GB300 system. In NVIDIA’s GTC 2026 coverage, the company says the first machine was a Dell Pro Max with GB300 delivered to Karpathy in Palo Alto on March 6, marking an early real-world placement for its new deskside AI system.

NVIDIA is positioning DGX Station GB300 as data-center-class AI hardware condensed into a workstation form factor. According to the GTC update, the system includes 748GB of coherent memory, delivers up to 20 petaflops of FP4 performance, and can support models up to 1 trillion parameters for frontier AI development on the desktop. The same GB300 architecture also means work developed locally can move to the cloud or the data center without a major rewrite.

That matters because the center of AI development is shifting from short prompt-response interactions toward long-running agents that need local files, local tools, and continuous iteration. NVIDIA explicitly ties DGX Station to that transition. Its GTC update says the early placements reflect demand for deskside systems that can support autonomous agents that reason, plan, and execute over time, rather than only serving as thin clients for remote infrastructure.

The company is also pairing the hardware story with software aimed at agent workflows. NVIDIA says DGX Station is being positioned alongside NemoClaw and OpenClaw, which it describes as frameworks for building and running more secure long-running agents locally. In practice, that means NVIDIA is not just selling a powerful desk machine, but trying to define a full local-to-cloud path for agent development, validation, and deployment.

The practical significance is straightforward. If developers can run serious agent experiments on a deskside GB300 box before scaling to larger clusters, the gap between prototyping and production narrows. That could make DGX Station GB300 attractive not only to frontier researchers, but also to enterprise teams that want tighter control over data, lower iteration latency, and a simpler path from local development to large-scale AI infrastructure.