NVIDIA has pulled its most powerful desktop AI system out of Linux-only territory. On October 7, 2026, the company announced DGX Station for Windows, a deskside workstation built around the GB300 Grace Blackwell Ultra Desktop Superchip that packs up to 748GB of coherent memory and 20 petaFLOPs of FP4 AI compute into a machine running Microsoft’s operating system instead of Ubuntu. Wccftech was first to frame the shift bluntly, reporting that NVIDIA’s “AI-powerhouse” DGX Station is “breaking out of Linux” and landing on Windows desks before the end of the year.
For a company that has spent a decade building its AI empire almost entirely on Linux containers, CUDA toolchains, and Ubuntu-based DGX OS images, putting a trillion-parameter-capable supercomputer inside a Windows box is a genuinely new direction. It also lands at a moment when NVIDIA’s RTX Spark Windows ARM PC and Microsoft’s own Surface Dev Box are already pushing local AI hardware into the Windows enterprise desktop. DGX Station for Windows is the top of that stack: a machine NVIDIA says is meant to put “data-center-level” AI performance on an individual developer’s desk.
What NVIDIA Actually Announced
According to NVIDIA’s own product page and an investor press release, DGX Station for Windows runs on the GB300 Grace Blackwell Ultra Desktop Superchip, combining a Blackwell Ultra GPU with a Grace CPU on a single coherent-memory platform. NVIDIA states the system is “the ultimate deskside AI supercomputer to build and run powerful AI agents connected to enterprise applications and workflows,” according to the company’s official DGX Station for Windows product page.
The headline specs are the same ones that made the Linux version of DGX Station notable when NVIDIA first detailed it earlier in 2026: up to 748GB of coherent memory shared between CPU and GPU, and up to 20 petaFLOPs of FP4 AI compute. What changes is the operating system layer sitting on top of that silicon. Instead of shipping preloaded with Ubuntu and NVIDIA’s Linux-based AI Enterprise stack, the Windows edition lets enterprise developers, data scientists, and researchers work inside the Windows environment most corporate IT departments already standardize on, while still tapping into CUDA, TensorRT, and NVIDIA’s broader AI Enterprise software suite underneath.
NVIDIA says the Windows edition of DGX Station “runs on the GB300 Grace Blackwell Ultra Desktop Superchip, delivering 748GB of coherent memory and up to 20 petaFLOPS of FP4 AI compute – enough to run models up to a trillion-parameter scale locally,” per the company’s own announcement on the NVIDIA blog covering the Windows-focused local AI event. That “trillion-parameter scale” framing is the number NVIDIA keeps repeating across its marketing for this launch, and it is the clearest signal of who this machine is actually built for: teams running frontier-scale open models locally rather than renting cloud GPU time by the hour.
Inside the GB300 Grace Blackwell Ultra Desktop Superchip
The GB300 Grace Blackwell Ultra Desktop Superchip is the same silicon foundation NVIDIA uses in both the Linux and Windows configurations of DGX Station, which means the performance ceiling of the two systems should be identical. The real differentiator is software, not hardware. Reporting on the chip’s architecture describes a 72-core Grace CPU built on Arm’s Neoverse V2 design, paired with a Blackwell Ultra GPU over NVIDIA’s NVLink-C2C chip-to-chip interconnect, which NVIDIA has previously rated at up to 900GB/s of bandwidth between the two processors.
Memory Architecture: Why 748GB Is Coherent, Not Just Large
The 748GB figure matters less for its raw size and more for how it’s organized. Rather than splitting memory into a separate GPU pool and a separate CPU pool that software has to shuttle data between, NVIDIA’s coherent memory architecture lets both the Grace CPU and the Blackwell Ultra GPU address the same unified memory space. Industry reporting on the GB300 Desktop Superchip has described the pool as a mix of roughly 252GB of HBM3e sitting close to the GPU and roughly 496GB of LPDDR5X attached to the Grace CPU side, combined into a single addressable space. That matters practically because a model that would otherwise need to be sharded across multiple GPUs just to fit in memory can instead run on one desktop box.
20 PFLOPs of FP4: What the Number Actually Buys You
FP4 is a four-bit floating-point format that trades numerical precision for raw throughput, and it’s become the standard yardstick NVIDIA uses to market Blackwell-generation AI performance. Twenty petaFLOPs of FP4 compute is a theoretical peak figure, not a sustained real-world benchmark, but it gives DGX Station for Windows enough headroom to run inference, and in some cases fine-tuning, on models that would otherwise require a multi-GPU server rack. NVIDIA pairs that compute claim with the same “up to 1 trillion parameters” framing it uses for the Linux version, repeated again in the company’s investor relations announcement, which notes the system “features up to 748GB of coherent memory and up to 20 petaflops of FP4 performance, and can be paired with an NVIDIA RTX PRO 6000 Blackwell Workstation GPU for frontier AI compute with ray-traced visualization and simulation,” a claim also detailed in Wccftech’s report on the Windows launch.
That optional RTX PRO 6000 Blackwell pairing is notable on its own. It means a fully loaded DGX Station for Windows isn’t purely an AI compute box, it’s also positioned as a visualization and simulation workstation, aimed at the same engineering and creative teams who might otherwise be shopping for workstations like the ASRock DeskSlim’s compact RTX Blackwell configuration.
DGX Station for Windows vs. DGX Station for Linux
NVIDIA isn’t retiring the Linux version. The original DGX Station, which ships preconfigured with Ubuntu and NVIDIA’s AI Developer Tools stack including CUDA, cuDNN, TensorRT, and the NVIDIA Container Toolkit according to NVIDIA’s own development documentation, remains available for teams already standardized on Linux-native AI pipelines. The Windows edition is a parallel product line aimed at the much larger population of enterprise developers whose laptops, desktops, and IT policies are built around Windows rather than Ubuntu.
| Specification | DGX Station for Windows | DGX Station for Linux |
|---|---|---|
| Operating system | Microsoft Windows | Ubuntu with NVIDIA AI Developer Tools |
| Core silicon | GB300 Grace Blackwell Ultra Desktop Superchip | GB300 Grace Blackwell Ultra Desktop Superchip |
| Coherent memory | Up to 748GB | Up to 748GB |
| Peak FP4 AI compute | Up to 20 petaFLOPs | Up to 20 petaFLOPs |
| Model scale supported | Up to ~1 trillion parameters | Up to ~1 trillion parameters |
| Primary audience | Enterprise developers on Windows IT stacks | Linux-native AI researchers and MLOps teams |
| Optional GPU pairing | NVIDIA RTX PRO 6000 Blackwell Workstation GPU | Not emphasized in NVIDIA’s Linux materials |
| Announced availability | Q4 2026, via OEM partners | Already shipping through NVIDIA partners |
The practical takeaway from that table is that NVIDIA isn’t choosing between platforms, it’s betting that the addressable market for a desk-side AI supercomputer is bigger if it doesn’t require every buyer to retrain their team on Linux first.
DGX Station vs. DGX Spark: The Memory and Compute Gap
DGX Station for Windows shouldn’t be confused with NVIDIA’s smaller, cheaper desktop AI box, DGX Spark, which tech-insider.org covered at launch earlier this year. DGX Spark is a Linux-oriented system built around a smaller Grace Blackwell configuration, with far less memory and compute headroom than the full-size Station. Reporting on the two product lines has described DGX Spark as topping out around 128GB of memory and roughly 1 petaFLOP of FP4 AI performance, aimed at models in the low hundreds of billions of parameters rather than trillion-parameter workloads.
| Metric | DGX Station (Windows/Linux) | DGX Spark | RTX Spark (Windows ARM PC) |
|---|---|---|---|
| Coherent/unified memory | Up to 748GB | ~128GB | 128GB (per NVIDIA’s RTX Spark configuration) |
| Peak FP4 AI compute | Up to 20 petaFLOPs | ~1 petaFLOP | Lower than DGX-class systems |
| Target model scale | Up to ~1 trillion parameters | ~200 billion parameters | Smaller local inference workloads |
| Form factor | Full deskside workstation | Compact desktop box | Mini PC / developer box |
| Operating system | Windows or Ubuntu | Linux-based DGX OS | Windows on Arm |
The gap between DGX Spark and the full DGX Station is roughly the gap between a high-end AI development box and an actual on-premises inference cluster condensed into one machine. That spread is also why NVIDIA’s partners like Microsoft, which built its own Surface-branded RTX Spark dev box, and tool vendors such as the makers of CPU-Z and HWMonitor, which added RTX Spark monitoring support ahead of launch, are treating the Spark and Station tiers as genuinely different product categories rather than a simple good/better/best ladder.
Pricing and Availability: What’s Confirmed and What Isn’t
NVIDIA has not published a definitive public retail price for DGX Station for Windows. The company’s investor release and product pages confirm a Q4 2026 availability window through OEM partners, but pricing is being left to those partners to set, which is standard practice for NVIDIA’s DGX Station line generally, since these systems are typically ordered through a partner sales channel rather than bought off a retail shelf. Industry estimates circulating around the GB300-based Station configuration have placed likely street pricing anywhere from roughly $50,000 to $100,000 depending on memory configuration and whether the optional RTX PRO 6000 Blackwell GPU is included, though none of those figures come from an official NVIDIA price sheet.
That price range puts DGX Station for Windows firmly in small-business-server territory rather than consumer workstation territory. It’s a figure that makes sense once you consider what it’s replacing: a rack of cloud GPU instances running up a monthly bill, versus a single capital purchase that a team owns outright. Whether that trade-off pencils out depends entirely on how continuously a given team actually uses the hardware, which is the same calculation enterprise buyers have had to make with every on-premises AI system NVIDIA has shipped since the original DGX-1.
The OEM Partner Lineup
NVIDIA named six system builders expected to ship DGX Station for Windows configurations: ASUS, Dell Technologies, GIGABYTE, HP, MSI, and Supermicro, according to the company’s own announcement. That’s a notably broader partner list than NVIDIA has used for some of its past DGX hardware generations, where Dell and Supermicro historically carried most of the enterprise volume. Opening the lineup to gaming-adjacent brands like ASUS and MSI alongside traditional enterprise vendors like HP and Supermicro suggests NVIDIA wants DGX Station for Windows sold through as many existing enterprise and prosumer sales channels as possible, rather than treating it as a niche data-center SKU.
It also mirrors the channel strategy NVIDIA has already used for its smaller RTX Spark hardware, where Microsoft’s Surface Laptop Ultra leaked with a 128GB RAM, 20-core RTX Spark configuration months before the Windows DGX Station news broke. NVIDIA appears to be building a consistent ladder of Windows-native AI hardware, from laptop-class RTX Spark chips, up through developer boxes, and now to full DGX-class desktop supercomputers, all running the same operating system enterprise IT already manages.
Why Windows, and Why Now
NVIDIA’s own messaging frames the Windows pivot as removing friction for a huge population of developers. The company describes DGX Station for Windows as “the most powerful deskside AI supercomputer for building and running agents on Windows enterprise applications and workflows,” according to NVIDIA’s blog post covering its joint announcement with Microsoft. That phrase, “Windows enterprise applications and workflows,” is the real thesis of the launch: most large enterprises run their line-of-business software, identity systems, and developer tooling on Windows, and asking every AI team inside that organization to learn Ubuntu just to use a DGX box has always been a real adoption barrier.
The timing also lines up with NVIDIA and Microsoft’s broader push to get local AI hardware classified alongside traditional PCs rather than treated as a separate server category. Earlier 2026 announcements around RTX Spark, Copilot+ PCs, and Windows-native AI agents all point the same direction: NVIDIA wants “AI PC” to eventually mean something closer to “capable of running serious local models,” and DGX Station for Windows is the top rung of that ladder.
Market Impact: On-Prem AI vs. the Cloud Compute Bill
The business case NVIDIA is making is straightforward: cloud GPU rental for frontier-scale inference and fine-tuning is expensive and unpredictable, and a single owned machine with 748GB of coherent memory removes a chunk of that monthly variability. For enterprises running sensitive or regulated data, there’s a second argument that matters just as much as cost, which is that a desk-side box keeps model weights and training data inside the building rather than inside a third-party cloud account.
That pitch is part of a broader pattern NVIDIA has leaned on throughout 2026, visible in how aggressively the company has pushed local inference hardware like its PAIR technology, which links idle PCs together for faster local AI. Whether DGX Station for Windows actually displaces meaningful cloud AI spend will depend on utilization rates inside the enterprises that buy it, and on how quickly NVIDIA’s own cloud partners respond with competitive on-demand pricing for similarly sized workloads.
Historical Context: From DGX-1 to a Trillion-Parameter Desktop
NVIDIA’s DGX line started in 2016 with DGX-1, a server-class box built to accelerate deep learning research, and it has stayed almost entirely Linux-first for a decade since. The jump to a Windows-native DGX Station in 2026 is arguably the biggest platform change the DGX brand has made since its debut, bigger in some ways than any single hardware generation bump, because it changes who can realistically deploy and manage the machine without hiring Linux-specialist IT staff.
It also tracks with how NVIDIA’s desktop AI ambitions have scaled up within just this calendar year. The company moved from positioning DGX Spark as a roughly 128GB, sub-petaFLOP developer box earlier in 2026, to shipping a 748GB, 20-petaFLOP DGX Station for Linux mid-year, to now extending that same top-tier hardware to Windows by Q4 2026. Each step has been about shrinking the gap between “data center” and “desk” rather than shrinking the chips themselves.
Competitive Landscape: Where Apple and AMD Fit
NVIDIA doesn’t have the desktop AI workstation category to itself. Apple’s unified-memory Mac Studio systems have made a similar pitch for local AI workloads using large pools of fast unified memory, though at nowhere near DGX Station’s compute ceiling. On the x86 side, AMD’s workstation and data-center silicon, including the large-core-count EPYC server chips the company has continued to push toward AI-adjacent workloads, competes more on raw CPU throughput and cost than on the coherent CPU-GPU memory architecture NVIDIA is selling with Grace Blackwell.
The honest comparison is that nobody else is currently shipping a desktop-form-factor system with 748GB of genuinely coherent CPU-GPU memory and 20 petaFLOPs of FP4 compute. NVIDIA’s advantage here isn’t just the silicon, it’s the vertical integration between CUDA, TensorRT, NVLink-C2C, and now Windows compatibility, all controlled by one company. Competing with that requires matching hardware, software, and OS-level enterprise reach simultaneously, which is a much harder target than matching a single spec sheet.
What NVIDIA Is Saying Publicly
NVIDIA has kept its public messaging around DGX Station for Windows consistent across every channel it has used to announce the system. On its own product page, the company states plainly: “NVIDIA DGX Station for Windows is the ultimate deskside AI supercomputer to build and run powerful AI agents connected to enterprise applications and workflows,” according to NVIDIA’s official DGX Station for Windows page.
In its investor-facing announcement, NVIDIA repeated the core spec claims almost word for word, noting the system “features up to 748GB of coherent memory and up to 20 petaflops of FP4 performance, and can be paired with an NVIDIA RTX PRO 6000 Blackwell Workstation GPU for frontier AI compute with ray-traced visualization and simulation,” per the official investor relations release. And on NVIDIA’s blog recapping the joint Windows-focused announcement with Microsoft, the company described the new system as “the most powerful deskside AI supercomputer for building and running agents on Windows enterprise applications and workflows.” Taken together, these statements, also referenced in NVIDIA’s own DGX Station development documentation, show NVIDIA deliberately targeting enterprise agent workloads, not general-purpose computing, as the core use case for this hardware.
What’s Still Unconfirmed
Several details that would normally accompany a hardware launch of this size remain open. NVIDIA has not released an official MSRP for any DGX Station for Windows configuration, leaving pricing entirely to its six named OEM partners. The company also hasn’t detailed exact SKU-level memory tiers below the 748GB maximum, meaning it’s unclear whether buyers will be able to purchase a lower-memory, lower-cost variant the way DGX Spark offers smaller configurations. NVIDIA likewise hasn’t confirmed exact benchmark results on real workloads, only the theoretical 20-petaFLOP FP4 peak figure, so independent performance testing once review units ship will be the first real check on how the Windows build performs against its Linux twin.
Predictions: Where This Goes Next
A handful of outcomes look likely as DGX Station for Windows moves from announcement to shipping hardware over the next two quarters.
- Pricing will land in enterprise-server territory, not prosumer territory. Expect OEM partners to announce configurations in the tens of thousands of dollars rather than anything resembling consumer-PC pricing, consistent with how NVIDIA’s partners have historically priced DGX-class hardware.
- A lower-memory Windows SKU is probable. Given how NVIDIA segmented DGX Spark into multiple memory tiers, a less expensive Windows DGX Station configuration below the full 748GB ceiling would broaden the buyer pool considerably.
- More OEM partners will likely add Windows AI workstation lines. With ASUS, Dell, GIGABYTE, HP, MSI, and Supermicro already confirmed, expect additional enterprise hardware vendors to follow once Q4 2026 sales data becomes available.
- Independent benchmarks will test the “1 trillion parameter” claim closely. Reviewers and enterprise IT buyers will want to see real throughput numbers on open models before trusting the theoretical FP4 ceiling NVIDIA has advertised.
- Microsoft’s Windows AI PC push will lean harder on DGX-class hardware as a halo product. Even if most buyers never touch a DGX Station, its existence strengthens Microsoft’s pitch that Windows, not just Linux, is a credible home for serious local AI work.
Frequently Asked Questions
What is NVIDIA DGX Station for Windows?
It’s a desktop AI workstation built on NVIDIA’s GB300 Grace Blackwell Ultra Desktop Superchip that runs Microsoft Windows instead of the Ubuntu-based Linux environment NVIDIA’s DGX systems have traditionally used. NVIDIA announced it on October 7, 2026, with up to 748GB of coherent memory and up to 20 petaFLOPs of FP4 AI compute.
How much memory does DGX Station for Windows have?
Up to 748GB of coherent memory shared between the Grace CPU and Blackwell Ultra GPU, according to NVIDIA’s official product page and investor press release.
How much does NVIDIA DGX Station for Windows cost?
NVIDIA has not published an official price. The system will be sold through OEM partners including ASUS, Dell Technologies, GIGABYTE, HP, MSI, and Supermicro starting in Q4 2026, and those partners will set final pricing.
What’s the difference between DGX Station for Windows and DGX Spark?
DGX Spark is a smaller, Linux-based desktop AI box with roughly 128GB of memory and around 1 petaFLOP of FP4 compute, aimed at models in the low hundreds of billions of parameters. DGX Station for Windows is a much larger system, with up to 748GB of memory and up to 20 petaFLOPs of compute, built for models up to roughly 1 trillion parameters.
When will DGX Station for Windows be available?
NVIDIA says the system is expected to become available through its OEM partners during the fourth quarter of 2026.
Is the Linux version of DGX Station going away?
No. NVIDIA continues to offer DGX Station preconfigured with Ubuntu and its AI Developer Tools stack. The Windows edition is an additional option alongside the existing Linux configuration, not a replacement for it.
What GPU can be paired with DGX Station for Windows?
NVIDIA says the system can be paired with an NVIDIA RTX PRO 6000 Blackwell Workstation GPU for added visualization and simulation performance alongside its core AI compute.
Who is DGX Station for Windows actually built for?
NVIDIA is marketing the system toward enterprise developers, researchers, and data-science teams who need to build and run large AI models and AI agents locally, inside organizations that are already standardized on Windows IT environments rather than Linux.
