Nvidia on Monday unveiled a desktop supercomputer powerful enough to run artificial intelligence models with up to a trillion parameters — roughly the scale of GPT-4 — without ever touching the cloud. It’s called a car DGX stationIt packs 748 gigabytes of coherent memory and 20 petaflops of computing into a box that sits next to the monitor, making it the most significant personal computing product since the original Mac Pro convinced creative professionals to ditch their workstations.
This was reported at the company’s annual meeting GTC conference It comes at a time when the AI industry in San Jose is grappling with a fundamental tension: the world’s most powerful models require massive data center infrastructure, but the developers and businesses that build on those models increasingly want to keep their data, agents and intellectual property local. The DGX Station is Nvidia’s answer – a six-figure machine that bridges the gap between the frontier of artificial intelligence and a single engineer’s desk.
What 20 petaflops really means on your desktop
The DGX station is built around the new GB300 Grace Blackwell Ultra Desktop SuperchipIt combines a 72-core Grace CPU and a Blackwell Ultra GPU via Nvidia’s NVLink-C2C interconnect. This connection provides 1.8 terabytes per second of coherent bandwidth between the two processors—seven times higher than PCIe Gen 6—which means the CPU and GPU share a single, seamless memory pool without the bottlenecks that typically make desktop AI work difficult.
Twenty petaflops—20 quadrillion operations per second—would have made this machine one of the world’s top supercomputers less than a decade ago. Peak system Oak Ridge National LaboratoryThe global No. 1 in 2018 performed nearly ten times as well in 2018, but occupied a room the size of two basketball courts. Nvidia packs a significant amount of that capability into something that plugs into a wall outlet.
748GB of single storage is arguably the more important number. Models with trillions of parameters are huge neural networks that must be loaded entirely into memory to run. Without enough memory, no processing speed matters – the model simply won’t fit. DGX Station clears that bar and does so with a consistent architecture that eliminates the latency penalties of data transfers between CPU and GPU memory pools.
Always-on agents need always-on hardware
Nvidia designed it DGX station for what he clearly sees as the next phase of artificial intelligence: not just systems that respond to commands, but autonomous agents that continuously think, plan, code, and execute tasks. In every major announcement GTC 2026 reinforced it "agent AI" thesis and DGX Station is the place where those agents are planned to be built and operated.
The key is pairing NemoClawA new open source stack that Nvidia also announced on Monday. NemoClaw integrates Nvidia’s Nemotron outdoor models OpenShella secure runtime that implements policy-based security, network, and privacy safeguards for autonomous agents. A single command installs the entire stack. Nvidia founder and CEO Jensen Huang made the combination clear, OpenClaw — a wider agent platform supports NemoClaw — "an operating system for personal AI" and compare directly to Mac and Windows.
The argument is simple: 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 local data and local models within a security sandbox, is architecturally more suitable for this workload than a rented GPU in someone else’s data center. DGX Station can function as a personal supercomputer for the solo developer or as a shared computing hub for teams, and supports air-gapped configurations for classified or regulated environments where data can never leave the building.
From desktop prototyping to data center production in zero rewrites
One of the cleverest aspects of 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 rewriting a single line of code. Nvidia sells a vertically integrated pipeline: prototype on your desk, then scale in the cloud when you’re ready.
This is important because the biggest hidden cost in AI development today isn’t computation—it’s engineering time lost rewriting code for different hardware configurations. A fine-tuned model on an on-premises GPU cluster often requires significant rework to deploy on a cloud infrastructure with different memory architectures, network stacks, and software dependencies. DGX Station eliminates this friction by running the same NVIDIA AI software that powers every layer of the Nvidia infrastructure. DGX Spark Vera Rubin for NVL72.
Nvidia has also extended Station’s little sister, the DGX Spark, with new clustering support. Up to four Spark devices can now operate as a single system with near-linear performance scaling — a "desktop data center" fits a conference table with no rack infrastructure or IT ticket. For teams that need to tune medium-sized models or develop smaller-scale agents, clustered Sparks offers a reliable departmental AI platform at a fraction of Station’s cost.
Early buyers discover where the market is headed
The initial customer list for DGX Station maps the industries where AI is making the fastest transition from experience to everyday operational tool. Snowflake uses the system to locally test the open-source Arctic training framework. EPRIThe Electric Power Research Institute is developing AI-powered weather forecasting to strengthen power grid reliability. Medivis integrates visual language models into surgical workflows. Microsoft Research and Cornell have widely implemented systems for hands-on artificial intelligence training.
The systems are available to order now and will ship in the coming months ASUS, Dell Technologies, GIGABYTE, MSIand Supermicrowith HP joins at the end of the year. Nvidia hasn’t disclosed pricing, but the GB300 components and the company’s historical DGX pricing suggests a six-figure investment—expensive by workstation standards, but pretty cheap compared to cloud GPU costs for trillion-parameter inference at scale.
The list of supported models highlights how open the AI ecosystem is: developers can run and fine-tune OpenAI. gpt-oss-120bGoogle Gemma 3, Gwen3, Mistral Big 3, DeepSeek V3.2and Nvidia’s own Nemotron models, among others. The DGX Station is model-agnostic by design – Swiss hardware in an industry where model allegiances change quarterly.
Nvidia’s real strategy: own every layer of the AI stack, from orbit to the office
The DGX station it did not come in a vacuum. It was part of a broom set GTC 2026 The announcements collectively reflect Nvidia’s ambition to enable AI computing at every physical scale.
At the top, Nvidia presented Vera Rubin platform — seven new chips in full production — are anchored by the Vera Rubin NVL72 rack, which combines 72 next-generation Rubin GPUs and delivers up to 10 times more output per watt than the current Blackwell generation. The Real CPUWith 88 dedicated Olympus cores, the agent targets the orchestration layer increasingly demanding workloads. On the far frontier, Nvidia announced the Vera Rubin Space Module for orbital data centers, providing 25 times more AI computation for space-based inference than the H100.
Nvidia revealed partnerships that include Adobe for creative artificial intelligence between orbit and office, automakers like BYD and Nissan for Level 4 autonomous vehicles, a coalition with Mistral AI and seven other labs to create open boundary models, and Dynamo 1.0, an already adopted open source inference operating system, including Google’s Az-Cloud companies. Cursor and Confusion.
The pattern is unmistakable: Nvidia wants to be the go-to computing platform for every AI workload—hardware, software, and models—everywhere. The DGX station it is the fabric that fills the gap between the cloud and the individual.
The cloud isn’t dead, but its monopoly on serious AI work is coming to an end
For the past few years, the default assumption in AI has been that serious business requires cloud GPU instances — renting Nvidia hardware. AWS, Azureor Google Cloud. This model works, but it comes with real costs: data access fees, latency, the security risk of sending proprietary data to a third-party infrastructure, and the fundamental loss of control inherent in renting someone else’s computer.
The DGX station isn’t killing the cloud — Nvidia’s data center business dwarfs desktop revenue and is accelerating. But it provides a reliable on-premise alternative for an important and growing category of workloads. Building a boundary model from scratch still requires thousands of GPUs in storage. Fine-tuning an open model with a trillion parameters over proprietary data? Are you inferring for an internal agent processing sensitive documents? Prototyping before incurring cloud costs? The machine under your desk is starting to look like a rational choice.
This is the strategic elegance of the product: while strengthening its cloud business, it also expands Nvidia’s addressable market into personal AI infrastructure, as everything built on-premises is designed to extend Nvidia’s data center platforms. It’s not cloud versus table. It’s a cloud and table and Nvidia provides both.
A supercomputer on every desk and an agent who never sleeps on it
It was the defining slogan of the PC revolution "a computer on every desk and in every home." After four decades, Nvidia is updating the building with a disturbing escalation. The DGX station It puts the original supercomputing power at the national labs next to the keyboard, and NemoClaw puts an autonomous artificial intelligence agent on top of it that works around the clock, writing code, calling tools, and performing tasks while the owner sleeps.
Whether the future is exciting or worrisome depends on your point of view. But one thing is no longer in dispute: the infrastructure required to build, manage, and own frontier AI has simply been moved from the server room to a desk drawer. The company that sells just about every serious AI chip on the planet has made sure to sell the desk drawer, too.
