Nvidia's DGX Station is a desktop supercomputer that runs trillion-parameter AI models without the cloud

Nvidia on Monday unveiled a deskside supercomputer powerful enough to run AI models with up to one trillion parameters — roughly the scale of GPT-4 — without touching the cloud. The machine, called the DGX Station, packs 748 gigabytes of coherent memory and 20 petaflops of compute into a box that sits next to a monitor, and it may be the most significant personal computing product since the original Mac Pro convinced creative professionals to abandon workstations.

The announcement, made at the company's annual GTC conference in San Jose, lands at a moment when the AI industry is grappling with a fundamental tension: the most powerful models in the world require enormous data center infrastructure, but the developers and enterprises building on those models increasingly want to keep their data, their agents, and their intellectual property local. The DGX Station is Nvidia's answer — a six-figure machine that collapses the distance between AI's frontier and a single engineer's desk.

What 20 petaflops on your desktop actually means

The DGX Station is built around the new GB300 Grace Blackwell Ultra Desktop Superchip, which fuses a 72-core Grace CPU and a Blackwell Ultra GPU through Nvidia's NVLink-C2C interconnect. That link provides 1.8 terabytes per second of coherent bandwidth between the two processors — seven times the speed of PCIe Gen 6 — which means the CPU and GPU share a single, seamless pool of memory without the bottlenecks that typically cripple desktop AI work.

Twenty petaflops — 20 quadrillion operations per second — would have ranked this machine among the world's top supercomputers less than a decade ago. The Summit system at Oak Ridge National Laboratory, which held the global No. 1 spot in 2018, delivered roughly ten times that performance but occupied a room the size of two basketball courts. Nvidia is packaging a meaningful fraction of that capability into something that plugs into a wall outlet.

The 748 GB of unified memory is arguably the more important number. Trillion-parameter models are enormous neural networks that must be loaded entirely into memory to run. Without sufficient memory, no amount of processing speed matters — the model simply won't fit. The DGX Station clears that bar, and it does so with a coherent architecture that eliminates the latency penalties of shuttling data between CPU and GPU memory pools.

Always-on agents need always-on hardware

Nvidia designed the DGX Station explicitly for what it sees as the next phase of AI: autonomous agents that reason, plan, write code, and execute tasks continuously — not just systems that respond to prompts. Every major announcement at GTC 2026 reinforced this "agentic AI" thesis, and the DGX Station is where those agents are meant to be built and run.

The key pairing is NemoClaw, a new open-source stack that Nvidia also announced Monday. NemoClaw bundles Nvidia's Nemotron open models with OpenShell, a secure runtime that enforces policy-based security, network, and privacy guardrails for autonomous agents. A single command installs the entire stack. Jensen Huang, Nvidia's founder and CEO, framed the combination in unmistakable terms, calling OpenClaw — the broader agent platform NemoClaw supports — "the operating system for personal AI" and comparing it directly to Mac and Windows.

The argument is straightforward: cloud instances spin up and down on demand, but always-on agents need persistent compute, persistent memory, and persistent state. A machine under your desk, running 24/7 with local data and local models inside a security sandbox, is architecturally better suited to that workload than a rented GPU in someone else's data center. The DGX Station can operate as a personal supercomputer for a solo developer or as a shared compute node for teams, and it supports air-gapped configurations for classified or regulated environments where data can never leave the building.

From desk prototype to data center production in zero rewrites

One of the cleverest aspects of the DGX Station's design is what Nvidia calls architectural continuity. Applications built on the machine migrate seamlessly to the company's GB300 NVL72 data center systems — 72-GPU racks designed for hyperscale AI factories — without rearchitecting a single line of code. Nvidia is selling a vertically integrated pipeline: prototype at your desk, then scale to the cloud when you're ready.

This matters because the biggest hidden cost in AI development today isn't compute — it's the engineering time lost to rewriting code for different hardware configurations. A model fine-tuned on a local GPU cluster often requires substantial rework to deploy on cloud infrastructure with different memory architectures, networking stacks, and software dependencies. The DGX Station eliminates that friction by running the same NVIDIA AI software stack that powers every tier of Nvidia's infrastructure, from the DGX Spark to the Vera Rubin NVL72.

Nvidia also expanded the DGX Spark, the Station's smaller sibling, with new clustering support. Up to four Spark units can now operate as a unified system with near-linear performance scaling — a "desktop data center" that fits on a conference table without rack infrastructure or an IT ticket. For teams that need to fine-tune mid-size models or develop smaller-scale agents, clustered Sparks offer a credible departmental AI platform at a fraction of the Station's cost.

The early buyers reveal where the market is heading

The initial customer roster for DGX Station maps the industries where AI is transitioning fastest from experiment to daily operating tool. Snowflake is using the system to locally test its open-source Arctic training framework. EPRI, the Electric Power Research Institute, is advancing AI-powered weather forecasting to strengthen electrical grid reliability. Medivis is integrating vision language models into surgical workflows. Microsoft Research and Cornell have deployed the systems for hands-on AI training at scale.

Systems are available to order now and will ship in the coming months from ASUS, Dell Technologies, GIGABYTE, MSI, and Supermicro, with HP joining later in the year. Nvidia hasn't disclosed pricing, but the GB300 components and the company's historical DGX pricing suggest a six-figure investment — expensive by workstation standards, but remarkably cheap compared to the cloud GPU costs of running trillion-parameter inference at scale.

The list of supported models underscores how open the AI ecosystem has become: developers can run and fine-tune OpenAI's gpt-oss-120b, Google Gemma 3, Qwen3, Mistral Large 3, DeepSeek V3.2, and Nvidia's own Nemotron models, among others. The DGX Station is model-agnostic by design — a hardware Switzerland in an industry where model allegiances shift quarterly.

Nvidia's real strategy: own every layer of the AI stack, from orbit to office

The DGX Station didn't arrive in a vacuum. It was one piece of a sweeping set of GTC 2026 announcements that collectively map Nvidia's ambition to supply AI compute at literally every physical scale.

At the top, Nvidia unveiled the Vera Rubin platform — seven new chips in full production — anchored by the Vera Rubin NVL72 rack, which integrates 72 next-generation Rubin GPUs and claims up to 10x higher inference throughput per watt compared to the current Blackwell generation. The Vera CPU, with 88 custom Olympus cores, targets the orchestration layer that agentic workloads increasingly demand. At the far frontier, Nvidia announced the Vera Rubin Space Module for orbital data centers, delivering 25x more AI compute for space-based inference than the H100.

Between orbit and office, Nvidia revealed partnerships spanning Adobe for creative AI, automakers like BYD and Nissan for Level 4 autonomous vehicles, a coalition with Mistral AI and seven other labs to build open frontier models, and Dynamo 1.0, an open-source inference operating system already adopted by A