Nvidia unveiled a desk-side machine Monday capable of running AI models at the scale of GPT-4 entirely offline, packaging 20 petaflops of compute and 748 gigabytes of unified memory into a box that sits beside a monitor.
The DGX Station, announced at the company’s GTC 2026 conference in San Jose, is built around the new GB300 Grace Blackwell Ultra Desktop Superchip — a single processor fusing a 72-core Grace CPU and a Blackwell Ultra GPU through Nvidia‘s NVLink-C2C interconnect. That connection delivers 1.8 terabytes per second of coherent bandwidth, seven times the throughput of PCIe Gen 6, according to the announcement.
The memory figure is the more consequential specification. Trillion-parameter models must be loaded entirely into memory to operate — processing speed is irrelevant if the model won’t fit. At 748 GB of coherent, unified memory, the DGX Station clears that threshold without the latency penalties that come from shuttling data between separate CPU and GPU memory pools.
A Fraction of a Room-Sized Machine
The 20 petaflop figure carries historical weight. The Summit system at Oak Ridge National Laboratory, which held the global top supercomputer ranking in 2018, delivered roughly ten times that performance — inside a facility the size of two basketball courts. Nvidia is delivering a meaningful portion of that capability through a machine that runs on a standard wall outlet.
The design targets what Nvidia describes as agentic AI: autonomous systems that reason, plan, write code, and execute tasks continuously rather than respond to isolated prompts. Every major product announcement at GTC 2026 reinforced that framing, and the DGX Station is positioned as the hardware where those agents are built and operated.
Paired with the hardware is NemoClaw, a new open-source software stack also announced Monday. It bundles Nvidia‘s Nemotron open models with OpenShell, a secure runtime that enforces policy-based security, network, and privacy controls for autonomous agents. The entire stack installs with a single command.
Jensen Huang’s Operating System Claim
Jensen Huang, Nvidia‘s founder and CEO, called OpenClaw — the broader agent platform that NemoClaw supports — “the operating system for personal AI,” drawing a direct comparison to Mac and Windows.
The commercial logic behind that framing is straightforward. Cloud instances spin up and down on demand, but always-on agents require persistent compute, persistent memory, and persistent state. A local machine running continuously, with data and models inside a security sandbox, is architecturally better suited to that workload than a rented GPU in a third-party data center. The DGX Station also supports air-gapped configurations for classified or regulated environments where data cannot leave the building.
The machine operates either as a personal system for a solo developer or as a shared compute node for teams. Nvidia is also selling continuity: applications built on the DGX Station migrate without code changes to the company’s GB300 NVL72 data center racks — 72-GPU systems built for hyperscale deployment. The pitch is a single development pipeline from a desk to a data center, removing the engineering cost of rewriting software for different hardware configurations.
Pricing was not disclosed in the announcement. The machine is described as a six-figure product aimed at developers and enterprises with strict data sovereignty requirements.
Photo by J. Kelly Brito on Pexels
This article is a curated summary based on third-party sources. Source: Read the original article
