Last reviewed
Jun 2, 2026
Sources
7 citations
Review status
Source-backed
Revision
v1 · 1,884 words
Add missing citations, update stale details, or suggest a clearer explanation.
NVIDIA DGX Station for Windows is a deskside artificial intelligence supercomputer announced by NVIDIA on June 1, 2026, at NVIDIA GTC Taipei during COMPUTEX 2026. It is built around the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip and is the first DGX Station configured to run Microsoft Windows, a project NVIDIA developed in collaboration with Microsoft. The machine carries up to 748GB of coherent memory and up to 20 petaflops of FP4 compute in a workstation chassis, and NVIDIA says it can run frontier AI models with up to 1 trillion parameters locally. [1][2]
The Windows version is a variant of the DGX Station that NVIDIA first showed at GTC in March 2025, which shipped with an Ubuntu Linux software stack. The 2026 announcement keeps the same underlying system design and the same GB300-class hardware, but adds native Windows support along with a new agent runtime called NVIDIA OpenShell. NVIDIA positions the product for enterprises that want to build and operate always-on AI agents directly inside the Windows applications and workflows their employees already use. [1][3]
| Attribute | Detail |
|---|---|
| Product type | Deskside AI workstation / supercomputer |
| Manufacturer | NVIDIA (system design); built by OEM partners |
| Announced | June 1, 2026, at NVIDIA GTC Taipei (COMPUTEX 2026) |
| Operating system | Microsoft Windows (with Windows Subsystem for Linux) |
| Superchip | NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip |
| GPU | 1x NVIDIA Blackwell Ultra |
| CPU | 1x 72-core NVIDIA Grace (Arm Neoverse V2) |
| CPU-GPU link | NVLink-C2C, 900 GB/s |
| Coherent memory | Up to 748GB (252GB HBM3e + 496GB LPDDR5X) |
| FP4 compute | Up to 20 petaflops (153 with sparsity) |
| Networking | NVIDIA ConnectX-8 SuperNIC, up to 800 Gb/s |
| Optional GPU | NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
| Local model size | Up to 1 trillion parameters (per NVIDIA) |
| Availability | Q4 2026 |
| OEM partners | ASUS, Dell Technologies, GIGABYTE, HP, MSI, Supermicro |
NVIDIA introduced the DGX Station for Windows on June 1, 2026, at NVIDIA GTC Taipei, the company's developer conference held alongside COMPUTEX in Taiwan. The reveal came during CEO Jensen Huang's keynote, which ran on the morning of June 1 (Taiwan time) at the Taipei Music Center and framed much of NVIDIA's roadmap around what Huang called the "age of agents." [4][5]
In its press release, NVIDIA described the system as "the world's most powerful deskside AI supercomputer designed to build, run and connect always-on AI agents to Windows applications and workflows, capable of running frontier AI models of up to 1 trillion parameters locally." The company tied the launch to a broader argument: as companies move from chatbots to autonomous agents that observe, reason, plan, and act, they need a tier of AI hardware that sits closer to the desktop than a data center but is far more capable than a consumer PC. [1][2]
NVIDIA explained the gap it is trying to fill. Cloud AI is powerful but introduces latency, recurring cost, and questions about where sensitive data goes, while a typical workstation cannot hold a trillion-parameter model in memory. The company said the DGX Station for Windows "bridges this gap as the first deskside AI supercomputer to bring NVIDIA GB300 Grace Blackwell-class AI infrastructure directly into the Windows ecosystem." [1]
At the center of the machine is the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip. The superchip pairs a single NVIDIA Blackwell Ultra GPU with a 72-core NVIDIA Grace CPU based on Arm Neoverse V2 cores. The two are joined by NVLink-C2C, NVIDIA's chip-to-chip interconnect, which provides 900 GB/s of bandwidth between the CPU and GPU and lets them share a single coherent memory space. [2][6]
This is the same family of silicon NVIDIA uses in its rack-scale data center systems such as the NVIDIA GB300 NVL72, reworked here for a single desktop unit rather than a rack of 72 GPUs. The Blackwell Ultra GPU carries the latest-generation Tensor Cores and native FP4 precision, the low-precision format NVIDIA leans on to push inference throughput for very large models. Because the Grace CPU and Blackwell Ultra GPU address one coherent pool of memory, a model and its working data do not have to be copied back and forth across a slower bus, which matters a great deal when a model is large enough to spill out of the GPU's own memory. [2][6]
The coherent memory figure of 748GB comes from combining the GPU's high-bandwidth memory with the CPU's larger but slower memory. NVIDIA's published specifications break the system down as follows. [6]
| Specification | Value |
|---|---|
| GPU | 1x NVIDIA Blackwell Ultra |
| CPU | 1x Grace 72-core (Neoverse V2) |
| GPU memory | 252GB HBM3e at 7.1 TB/s |
| CPU memory | 496GB LPDDR5X at 396 GB/s |
| Total coherent memory | Up to 748GB |
| FP4 Tensor Core | 20 / 153 PFLOPS (dense / with sparsity) |
| FP8 / FP6 Tensor Core | 10 PFLOPS |
| FP16 / BF16 Tensor Core | 5 PFLOPS |
| TF32 Tensor Core | 2.5 PFLOPS |
| FP32 | 80 TFLOPS |
| FP64 / FP64 Tensor Core | 1.3 TFLOPS |
| INT8 Tensor Core | 330 TOPS |
| CPU-GPU interconnect | NVLink-C2C, 900 GB/s |
| Networking | NVIDIA ConnectX-8 SuperNIC, up to 800 Gb/s Ethernet |
| Network ports | 2x QSFP112 (400 Gb/s each), 1x RJ45 10GbE, 1x RJ45 1GbE (BMC) |
| Storage | 4x M.2 Gen 5 slots |
| Total system power | 1,600 W |
| Operating system | Microsoft Windows |
The networking is worth calling out. The built-in NVIDIA ConnectX-8 SuperNIC supports speeds up to 800 Gb/s, which means several DGX Station units can be linked together into a small cluster for jobs that outgrow a single box. [1][6]
NVIDIA's headline claim is that a single DGX Station for Windows can run frontier AI models of up to 1 trillion parameters locally, or alternatively serve hundreds of AI agents in parallel. This claim should be read as NVIDIA's own figure. It leans heavily on two things: the 748GB coherent memory pool, which is large enough to hold a trillion-parameter model when that model is quantized to a low-precision format, and the FP4 path on the Blackwell Ultra Tensor Cores, which is what makes inference at that scale practical on one machine rather than a rack. [1][2][3]
For context, running a model "locally" at this size has until now generally meant renting time on cloud GPUs or owning a multi-GPU server. Putting it under a desk changes the calculus for organizations that care about keeping proprietary data and model weights on premises, and that is the audience NVIDIA is aiming at. [3]
The defining feature of this variant is that it runs Windows rather than the Linux stack used by the original DGX Station. NVIDIA built the product in collaboration with Microsoft so that the full NVIDIA AI software stack runs on Windows, with Windows Subsystem for Linux available for tooling that expects a Linux environment. The intent is to plug data-center-class AI into the operating system that most enterprise IT departments already manage, secure, and patch. [1][2]
Alongside the hardware, NVIDIA announced NVIDIA OpenShell, described as an open-source, secure-by-design runtime for autonomous agents. OpenShell uses new Windows security and containment primitives so that agents can take actions inside applications and across a system while staying sandboxed. This is the software layer meant to make the "always-on agents" pitch safe enough for corporate deployment. [1]
Two executives spoke to the partnership. Chris Marriott, NVIDIA's vice president of enterprise platforms, said the system "delivers supercomputing-class AI directly into Windows, where millions already design, engineer, research and create every day." Pavan Davuluri, Microsoft's executive vice president of Windows and Devices, said the collaboration "unlocks a new class of AI performance on Windows, the platform enterprises trust for security, manageability and compatibility." [1]
For workloads that mix AI with graphics, the DGX Station for Windows can be configured with an additional NVIDIA RTX PRO 6000 Blackwell Workstation GPU, pairing the frontier AI compute of the GB300 with ray-traced visualization and simulation. NVIDIA pointed to use across productivity, creative, design, and physical AI applications. [1][6]
The DGX Station for Windows sits in a family of NVIDIA "personal AI computer" products, and it helps to place it against its siblings.
The closest relative is the original DGX Station, which NVIDIA announced at GTC on March 18, 2025. That machine introduced the GB300 Grace Blackwell Ultra Desktop Superchip, the 748GB coherent memory pool, the 20 petaflops of FP4, and the ConnectX-8 SuperNIC. It shipped with an Ubuntu-based NVIDIA software stack rather than Windows. The 2026 announcement is, in effect, the same system design opened up to the Windows ecosystem, which is why the hardware specifications line up so closely. [3][7]
At the same March 2025 event NVIDIA also unveiled the NVIDIA DGX Spark, a much smaller and cheaper desktop unit built on the GB10 Grace Blackwell Superchip. DGX Spark, marketed as "the world's smallest AI supercomputer," targets individual developers and researchers who want to prototype and fine-tune models on a compact box, whereas the DGX Station class is built for far larger local models. [7]
Further up the stack sits the data center hardware. The NVIDIA DGX B300 and rack-scale systems like the NVIDIA GB300 NVL72 and the earlier NVIDIA GB200 NVL72 deliver many times the compute and memory of a single DGX Station, but they live in racks inside a server room, not on a desk. The DGX Station for Windows is best understood as the deskside tier of NVIDIA's Blackwell and Blackwell Ultra lineup, sitting above DGX Spark and below the rack-scale machines. [3][7]
NVIDIA said the DGX Station for Windows is on track to be available in Q4 2026 from a set of OEM partners: ASUS, Dell Technologies, GIGABYTE, HP, MSI, and Supermicro. NVIDIA designs the reference system, and these manufacturers build and sell their own branded versions. NVIDIA did not announce pricing at GTC Taipei, which is typical for a system that ships through OEMs at varying configurations. [1][2]