NVIDIA has confirmed a version of its DGX Station that runs on Microsoft Windows instead of Linux, packing up to 748 GB of coherent memory and 20 petaFLOPs of FP4 AI compute into a desk-sized workstation. The move, detailed in NVIDIA’s own DGX Station for Windows product page and amplified this week by outlets including Wccftech and HotHardware, ends nearly a decade of DGX hardware being locked to Linux-only deployments. Six manufacturers plan to ship units by the fourth quarter of 2026.
For hardware buyers tracking NVIDIA’s desktop AI push, this is the second release in a two-tier strategy that started with the smaller DGX Spark. The Windows variant targets a different buyer: the enterprise developer who already lives inside a Windows-standardized company and doesn’t want to dual-boot into Linux just to fine-tune a model. Here’s what NVIDIA has confirmed, what remains unannounced, and how the machine stacks up against the competition closing in from AMD and Apple.
What NVIDIA Just Confirmed for Windows-Based AI Workstations
The headline spec is memory. NVIDIA DGX Station for Windows carries up to 748 GB of coherent memory, split between 252 GB of HBM3e tied directly to the GPU and 496 GB of LPDDR5X serving the CPU side. NVIDIA says that pool is large enough to hold and run AI models scaling to 1 trillion parameters locally, without offloading to a cluster or a cloud instance. Compute tops out at 20 petaFLOPs of FP4 precision, the low-bit format NVIDIA has pushed as the default for inference-heavy workloads since the Blackwell generation launched.
Power draw sits at 1,600 watts, which NVIDIA notes requires a dedicated 20-amp circuit, a detail that matters more for IT procurement than for marketing copy. A 20-amp circuit isn’t standard in every office cubicle, so deployment will likely cluster around dedicated workstation rooms or lab space rather than someone’s desk in the literal sense the product name implies.
Inside the GB300 Grace Blackwell Ultra Desktop Superchip
The system is built around a single NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip, pairing one Blackwell Ultra GPU with a 72-core Grace CPU built on Arm’s Neoverse V2 cores. That CPU-GPU pairing is the same basic architecture NVIDIA uses in its data-center GB300 NVL72 racks, just shrunk down to fit a single tower instead of a server rack. The link between the two chips runs over NVIDIA’s NVLink-C2C interconnect, which keeps CPU and GPU memory addressable as one pool rather than forcing software to manage two separate memory spaces.
That unified-memory design is the same philosophy behind NVIDIA’s smaller DGX Spark, which shattered.io covered in detail when NVIDIA split it into 64GB and 128GB configurations. The Station pushes the same architecture to a scale that competes directly with multi-GPU data-center nodes rather than single-GPU workstations.
Buyers who need more graphics horsepower for visualization or simulation work alongside the AI workload can pair the system with an NVIDIA RTX PRO 6000 Blackwell Workstation GPU, according to NVIDIA’s announcement. That optional add-on signals NVIDIA is positioning the Station as a hybrid machine, one that can double as a rendering or CAD workstation and not just an isolated AI training box.
748GB of Coherent Memory, Broken Down
Splitting 748 GB into two memory types rather than one pool of uniform RAM is a deliberate tradeoff. The 252 GB of HBM3e sits closest to the GPU and handles the active weights and activations during inference or training, the data that needs the fastest possible access. The 496 GB of LPDDR5X, slower but far cheaper per gigabyte than HBM3e, holds everything else, including larger portions of a model that aren’t being actively computed against at a given moment.
Why the split matters for model size
A model’s full weight set rarely needs to sit entirely in the fastest memory tier at once, especially during inference. By tiering memory this way, NVIDIA can advertise a trillion-parameter ceiling without needing to pack three-quarters of a terabyte of HBM3e into a desktop chassis, something that would be both prohibitively expensive and thermally brutal to cool in a workstation form factor rather than a server rack with industrial airflow.
20 PetaFLOPS and the 1-Trillion-Parameter Model Claim
NVIDIA’s claim that the Station can run models up to 1 trillion parameters locally is notable mainly because of what it implies about where AI development is heading: out of hyperscaler data centers and onto desks inside regulated industries that can’t send proprietary data to a third-party API. Banks, defense contractors, and healthcare systems have spent much of 2026 wrestling with exactly that constraint, and a trillion-parameter ceiling on a single workstation changes the calculus for what can be kept entirely in-house.
Twenty petaFLOPs of FP4 compute is a meaningful jump from the DGX Spark line, which tops out around 1 petaFLOP of FP4 performance in its 128 GB configuration. That’s roughly a twentyfold compute gap between NVIDIA’s two current desktop AI products, underscoring that these aren’t really competing SKUs so much as two rungs of the same ladder aimed at different budgets and different workloads.
Why NVIDIA Finally Brought DGX Station to Windows
Every prior DGX Station and DGX workstation shipped running Linux exclusively, which forced Windows-standardized enterprises to stand up and maintain a second operating environment just to touch NVIDIA’s highest-end desktop hardware. That friction fell disproportionately on large enterprises, the same Fortune 500 companies NVIDIA is now explicitly courting with a Windows-native option. The new Station preserves access to Linux toolchains through the Windows Subsystem for Linux, so developers who’ve built workflows around Linux-based AI frameworks aren’t locked out, they just no longer have to boot into a separate OS to use them.
The timing lines up with NVIDIA’s broader Windows push this year. The company already worked with Microsoft to bring RTX Spark and local AI agent support to consumer Windows PCs, a rollout that shipped to six PC makers earlier in 2026. DGX Station for Windows reads as the enterprise-grade extension of that same strategy: meet developers and IT departments on the operating system they already run, rather than asking them to adopt a new one.
Six Manufacturers, One Q4 2026 Launch Window
NVIDIA named six hardware partners building commercial versions of the Station: ASUS, Dell, GIGABYTE, HP, MSI, and Supermicro. That’s a notably broad OEM list for a first-generation Windows product, spanning everyone from consumer-facing brands like ASUS and MSI to enterprise-focused suppliers like Dell and Supermicro. It suggests NVIDIA expects demand across a wide range of buyer sizes, from individual research labs ordering a single unit to enterprise IT departments provisioning fleets of them.
NVIDIA’s official newsroom announcement places availability in the fourth quarter of 2026, which started October 1. That gives the six manufacturers roughly three months to get units into customers’ hands, a tight but not unusual window for workstation-class hardware that’s already cleared engineering validation.
Price and Availability: What’s Confirmed, What Isn’t
NVIDIA has not published a price for DGX Station for Windows. That’s worth stating plainly because several outlets covering the launch have floated unofficial estimates, generally in the six-figure range given the system’s data-center-class components, but none of those figures trace back to an NVIDIA price sheet. Buyers evaluating a purchase for Q4 2026 should treat any specific dollar figure circulating online as a guess rather than a quote until one of the six OEM partners posts an actual SKU price.
What is confirmed is the channel: the Station will ship through the same six manufacturers rather than direct from NVIDIA, which mirrors how DGX Spark reached buyers through partners and retail listings rather than an NVIDIA storefront.
DGX Station vs DGX Spark: NVIDIA’s Desktop AI Ladder
NVIDIA now sells two distinct desktop AI systems, and the gap between them is wide enough that they barely compete for the same buyer. DGX Spark is the entry point, a smaller box aimed at individual developers prototyping locally. DGX Station is the step up, aimed at teams running production-scale models without a data center. The table below lines up the two product lines using only specifications each company has publicly confirmed.
| Spec | DGX Spark (128GB) | DGX Station for Windows |
|---|---|---|
| Superchip | GB10 Grace Blackwell | GB300 Grace Blackwell Ultra Desktop |
| CPU cores | 20-core Arm (10x Cortex-X925, 10x Cortex-A725) | 72-core Grace (Neoverse V2) |
| Total memory | 128GB LPDDR5X unified | Up to 748GB (252GB HBM3e + 496GB LPDDR5X) |
| AI compute (FP4) | Up to 1 petaFLOP | Up to 20 petaFLOPs |
| Operating system | NVIDIA DGX OS (Linux) | Microsoft Windows (with WSL) |
| Street price (USD) | ~$4,699 to $4,999 (Founders Edition) | Not yet announced |
| Availability | Shipping since October 2025 | Q4 2026 |
The twentyfold compute jump between the two systems roughly tracks the price gap buyers should expect once NVIDIA or its partners confirm Station pricing. Spark was built to sit on an individual desk. Station is sized, and priced, for a department budget.
The Competition: AMD’s Threadripper Halo Station and Apple’s Mac Studio
NVIDIA isn’t the only chipmaker betting that AI compute belongs on a desk instead of in a rack. AMD and Apple have both shipped or announced competing hardware this year, and the contrast shows just how differently each company is solving the same problem.
AMD’s answer: more memory, different software stack
AMD’s Threadripper Halo Station, which shattered.io covered when AMD unveiled it, pairs a 96-core Threadripper Pro 9995WX CPU with up to four Instinct MI350P GPUs, totaling roughly 2.6TB of system memory, well ahead of the Station’s 748GB. AMD’s own comparisons, cited by outlets including ITPro, put that at roughly 3.4 times the memory of NVIDIA’s Station. The catch is software: AMD’s MI350P GPUs run on the ROCm stack, which still trails CUDA in library support and the volume of research code already written against it. Threadripper Halo Station is also targeting a 2027 launch, a year behind NVIDIA’s Q4 2026 window.
Apple’s answer: unified memory, no discrete GPU needed
Apple’s approach looks nothing like either rival’s. The M5 Ultra Mac Studio, reviewed on this site at launch, packs 36 CPU/GPU cores and 1.2TB/s of unified memory bandwidth into a single chip with no discrete GPU at all, starting well under six figures. It won’t touch 748GB of coherent memory or NVIDIA’s CUDA ecosystem, but for smaller models and local inference it’s a fraction of the cost and power draw of either the Station or Threadripper Halo.
| System | Maker | Total memory | Software stack | Target launch |
|---|---|---|---|---|
| DGX Station for Windows | NVIDIA | 748GB (252GB HBM3e + 496GB LPDDR5X) | CUDA, Windows + WSL | Q4 2026 |
| Threadripper Halo Station | AMD | ~2.6TB | ROCm, Linux/Windows | 2027 |
| Mac Studio (M5 Ultra) | Apple | Up to 512GB unified | Apple MLX/Metal | Shipping now |
| DGX Spark (128GB) | NVIDIA | 128GB unified | CUDA, DGX OS (Linux) | Shipping now |
No single system wins on every axis. AMD leads on raw memory, Apple leads on price-to-performance for smaller workloads, and NVIDIA leads on software maturity and the breadth of its OEM partner list. Which one matters most depends entirely on whether a buyer’s bottleneck is memory capacity, budget, or the availability of CUDA-optimized tooling for their specific model.
A Short History of the DGX Line, From DGX-1 to GB300
NVIDIA’s DGX brand dates back to the DGX-1 in 2016, a Pascal-based server NVIDIA pitched as a turnkey deep-learning box for research labs that didn’t want to assemble their own GPU cluster. Each generation since has scaled compute aggressively while shrinking the gap between data-center and desktop form factors: Ampere-based DGX A100 systems, Hopper-based DGX H100 racks, and now the Blackwell and Blackwell Ultra generation that underpins both the data-center GB300 NVL72 and the desk-sized Station covered here.
That data-center sibling is itself in the news. NVIDIA’s Vera Rubin platform, the architecture slated to follow Blackwell, beat GB300 NVL72 by 3.7x in its first MLPerf inference run, a reminder that even as GB300-based hardware reaches desktops for the first time, NVIDIA’s data-center roadmap has already moved a full generation past it. The Station’s GB300 silicon is current-generation hardware for a workstation, but it’s not NVIDIA’s newest architecture overall.
Checking Hardware Against DGX-Class Specs
For developers trying to gauge whether their current workstation comes anywhere close to DGX Station territory, a quick memory and compute check is the fastest way to see the gap. On a Windows machine with an NVIDIA GPU installed, the following command surfaces total VRAM, driver version, and current memory utilization, the same baseline figures that matter when sizing a model against Station-class coherent memory.
nvidia-smi --query-gpu=name,memory.total,memory.used,driver_version --format=csvRunning that against a typical high-end consumer GPU will return a memory total measured in tens of gigabytes, not hundreds, which is the quickest illustration of why NVIDIA built a separate product line rather than just selling a faster version of its existing GeForce or RTX PRO cards for this use case.
Market Impact: What This Means for Enterprise AI Budgets
The business case for DGX Station for Windows hinges on a comparison most enterprise buyers are already running internally: the cost of renting cloud GPU time versus the cost of owning hardware outright. NVIDIA’s own B200 cloud instances have gotten more expensive this year, not cheaper. B200 cloud pricing climbed to $8.01 an hour, up 79%, according to reporting earlier in 2026, a trend that makes a fixed, if unconfirmed, upfront cost for owned hardware look more attractive the longer a model needs to stay in training or heavy inference.
Supply is the other half of the equation, and it’s less favorable. HBM3e, the memory type that makes up 252GB of the Station’s total pool, has been in tight supply across the industry. NVIDIA’s own next-generation Rubin Ultra platform reportedly lost a third of its planned memory allocation to the ongoing HBM shortage, which raises a real question about whether the six OEM partners building Station units will have consistent HBM3e supply to meet Q4 2026 demand without delays or allocation caps on early orders.
For enterprise IT budgets, that combination, rising cloud rates and tight memory supply, is likely to push some AI workloads toward owned hardware even without a confirmed Station price, simply because the alternative (cloud rental at current rates) keeps getting more expensive on its own.
5 Predictions for Desktop AI Supercomputers Through 2027
- NVIDIA will confirm Station pricing closer to the Q4 2026 shipping window, likely tiered by OEM rather than a single fixed price across ASUS, Dell, GIGABYTE, HP, MSI, and Supermicro.
- AMD’s Threadripper Halo Station, slated for 2027, will lean on its memory-capacity lead as its primary marketing angle against NVIDIA, since it can’t yet match CUDA’s software maturity.
- HBM3e supply constraints will delay at least one OEM’s Station shipments past the Q4 2026 window NVIDIA announced, mirroring the allocation pressure already reported around Rubin Ultra.
- Apple will continue positioning Mac Studio as the budget on-ramp to local AI rather than a direct Station competitor, given the gap in coherent memory and CUDA compatibility.
- A third DGX form factor, sized between Spark and Station, is plausible within the next product cycle as NVIDIA fills the gap between a 1-petaFLOP desktop box and a 20-petaFLOP one.
Related
- Nvidia Splits DGX Spark: $4,999 vs $6,950 128GB
- AMD Threadripper Halo Station: 96 Cores, 576GB
- M5 Ultra Mac Studio Review: 90% Faster, $12,299
- Vera Rubin Beats GB300 3.7x in First MLPerf Run
- RTX Spark Ships to 6 PC Makers, Erases 4.2% Drop
Frequently Asked Questions
What is NVIDIA DGX Station for Windows?
It’s a deskside AI workstation built around NVIDIA’s GB300 Grace Blackwell Ultra Desktop Superchip, running Microsoft Windows instead of the Linux-only setup of earlier DGX systems, with up to 748GB of coherent memory and 20 petaFLOPs of FP4 AI compute.
How much does DGX Station for Windows cost?
NVIDIA has not announced an official price. Treat any specific dollar figure circulating online as an unverified estimate until one of the six manufacturing partners publishes a real SKU price.
When will DGX Station for Windows be available?
NVIDIA’s announcement targets the fourth quarter of 2026, with ASUS, Dell, GIGABYTE, HP, MSI, and Supermicro building commercial units.
How is DGX Station different from DGX Spark?
DGX Spark is the smaller, entry-level system, topping out at 128GB of unified memory and around 1 petaFLOP of FP4 compute, and it runs NVIDIA’s Linux-based DGX OS. DGX Station for Windows scales up to 748GB of memory and 20 petaFLOPs of compute, and adds Windows as an operating system option.
Can DGX Station for Windows really run a 1-trillion-parameter model?
NVIDIA says the system’s 748GB coherent memory pool supports AI models of up to 1 trillion parameters. That figure comes directly from NVIDIA’s own announcement and hasn’t yet been independently benchmarked against a specific trillion-parameter model running locally.
Does DGX Station for Windows still support Linux tools?
Yes. NVIDIA says the system preserves access to Linux toolchains through the Windows Subsystem for Linux, so Linux-based AI frameworks remain usable without dual-booting into a separate OS.
What competes with DGX Station for Windows?
The closest rivals are AMD’s Threadripper Halo Station, which offers more total memory but runs on the less mature ROCm software stack and targets a 2027 launch, and Apple’s M5 Ultra Mac Studio, which costs far less but caps out at a smaller unified memory pool and lacks CUDA support.
Why does the Station need a 20-amp circuit?
NVIDIA lists the system’s power draw at 1,600 watts, which exceeds what a standard 15-amp household or office circuit can safely and continuously supply, making a dedicated 20-amp circuit a practical requirement for deployment.
