Nvidia Unveils 'DGX Station' For Windows, Underscoring The Shift To Hybrid AI With 1 Trillion-Parameter Models

Nvidia has unveiled the DGX Station for Windows, described as the world's most powerful deskside AI supercomputer.

This new system brings advanced AI capabilities directly to enterprise desktops by allowing users to build run and connect always on AI agents to Windows based applications and workflows. It enables building and running secure AI agents on Windows via Nvidia's OpenShell and is manufactured by partners including ASUS, Dell, Gigabyte, HP, MSI, and Supermicro.

The biggest advantage is that, it can run 1 trillion parameters locally.

The launch underscores the shift to hybrid AI with local compute for improved privacy, security, and low-latency enterprise applications.

At its core the platform is powered by the Nvidia GB300 Grace Blackwell Ultra Desktop Superchip, which combines a 72 core Nvidia Grace CPU with a Blackwell Ultra GPU linked through the high speed NVLink C2C interconnect.

The setup delivers up to 748 gigabytes of coherent unified memory consisting of 252 gigabytes of HBM3e memory on the GPU side running at 7.1 terabytes per second bandwidth and 496 gigabytes of LPDDR5X memory on the CPU side. It also provides up to 20 petaflops of FP4 AI performance making it suitable for handling frontier level models.

The DGX Station for Windows stands out for its ability to run AI models with as many as one trillion parameters entirely on local hardware without depending on cloud resources.

This includes support for high throughput inference pretraining fine tuning and data science tasks where large datasets can be processed directly in the expansive memory pool to avoid bottlenecks from data movement.

Enterprises can deploy it as a dedicated supercomputer for individual developers or as a shared resource for teams with workloads that scale smoothly to larger GB300 systems in data centers or the cloud if needed.

For added versatility the system can pair with an optional Nvidia RTX PRO 6000 Blackwell Workstation GPU to incorporate ray traced visualization and simulation capabilities enabling physical AI applications that blend compute with virtual environments.

Networking comes through the Nvidia ConnectX 8 SuperNIC which supports speeds of up to 800 gigabits per second for efficient data transfers and the potential to link multiple units together.

On the software side the platform integrates deeply with the Windows ecosystem through Nvidia OpenShell an open source secure by design runtime for autonomous agents.

OpenShell leverages Windows security and containment features to create isolated sandboxes for each agent ensuring that operations remain separated from system level policies.

This setup prevents issues such as policy overrides credential leaks or data exposure while allowing agents to operate continuously with real time reasoning and direct connections to enterprise applications.

The inclusion of the Windows Subsystem for Linux further enables access to Linux based AI toolchains within a familiar Windows environment.

Full compatibility with Microsofts enterprise management tools supports deployment updates security compliance and fleet oversight extending the trusted Windows infrastructure to high performance AI workloads.

This development reflects a broader move toward hybrid AI setups where local processing complements cloud resources.

Organizations gain improved privacy security and low latency for sensitive or time critical applications while developers researchers engineers and data scientists can work with frontier models in controlled settings. The system addresses the historical divide where heavy AI tasks required Linux based data center hardware and Windows handled everyday productivity creative design and engineering tools.

By bridging these worlds the DGX Station for Windows positions itself as practical infrastructure for building secure AI agents that automate tasks across 3D design engineering analytics and more.Manufacturing partners including ASUS Dell Technologies GIGABYTE HP MSI and Supermicro will produce the systems which are scheduled for availability in the fourth quarter of this year.

Nvidia's executives have highlighted how the platform delivers supercomputing class AI into the Windows environment used daily by millions. Microsoft leaders have emphasized the collaboration that scales Windows performance from lightweight devices up to data center class workstations unlocking new levels of AI capability on a platform valued for its security manageability and compatibility.

It's worth noting that the DGX Station for Windows was just one of the several products Nvidia introduced during Nvidia GTC, the global artificial intelligence conference for developers that brings together developers, engineers, researchers, inventors, and IT professionals.

It was revealed a day before the event at Computex in Taiwan, where both Nvidia and Microsoft teased a similar cryptic message on social media.