Runs Windows natively
No new group policies, no OS migration. The station deploys inside existing Windows infrastructure, with the NVIDIA AI software stack and WSL support for Linux toolchains.
NVIDIA DGX Station · GB300 Grace Blackwell Ultra
Up to 20 petaFLOPS of FP4 compute and 748 GB of coherent memory in a deskside tower — enough to run trillion-parameter models and always-on AI agents inside your own building, on Windows.
Why DGX Station
Most AI programmes stall because the compute lives somewhere the team cannot touch. DGX Station puts a GB300-class system on the floor beside the team and keeps it inside the infrastructure IT already manages.
No new group policies, no OS migration. The station deploys inside existing Windows infrastructure, with the NVIDIA AI software stack and WSL support for Linux toolchains.
252 GB of HBM3e and 496 GB of LPDDR5X are unified over NVLink-C2C at 900 GB/s, so a trillion-parameter model fits in a single pool instead of being sharded across rented nodes.
A ConnectX-8 SuperNIC at up to 800 Gb/s links multiple stations together — or connects them to a GB300 cluster in the cloud or data centre, same software stack.
Three OEM builds, one reference design
The GB300 Grace Blackwell Ultra platform is available as NVIDIA's own DGX Station and through authorised OEM partners. The silicon is identical; the chassis, service terms and fleet compatibility are not.
The reference system from NVIDIA — GB300 Grace Blackwell Ultra, 748 GB coherent memory, NVIDIA AI software stack and NVIDIA support. Available alongside the OEM builds below.





Technical specifications
Platform specifications as published by NVIDIA for DGX Station. Final configuration depends on the OEM system you order.
| NVIDIA GPU | 1x NVIDIA Blackwell Ultra |
|---|---|
| NVIDIA CPU | 1x Grace 72-core Neoverse V2 |
| GPU memory | 252 GB HBM3e ECC · 7.1 TB/s |
| CPU memory | 496 GB LPDDR5X ECC · 396 GB/s |
| NVLink-C2C | 900 GB/s |
| Networking | NVIDIA ConnectX-8 SuperNIC · up to 800 Gb/s Ethernet |
| Supported OS | Microsoft Windows (DGX Station for Windows) · DGX OS |
| FP4 Tensor Core | 20 PFLOPS (15 PFLOPS without sparsity) |
| FP8 / FP6 Tensor Core | 10 PFLOPS |
| INT8 Tensor Core | 330 TOPS |
| FP16 / BF16 Tensor Core | 5 PFLOPS |
| TF32 Tensor Core | 2.5 PFLOPS |
| FP32 | 80 TFLOPS |
| FP64 / FP64 Tensor Core | 1.3 TFLOPS |
| Total system power | 1,600 W · 20 A circuit required |
| Storage | 4x M.2 Gen 5 slots |
| Ethernet ports | 2x QSFP112 (400 Gb/s each) · 1x RJ45 10GbE · 1x RJ45 1GbE (BMC) |
| PCIe slots | 1x PCIe Gen 5 x16 · 2x PCIe Gen 5 x16 (x8 electrical) |
| Optional RTX PRO GPU | RTX PRO 6000 Blackwell Workstation / Max-Q · RTX PRO 4000 Blackwell SFF · RTX PRO 2000 Blackwell |
Peak rates are based on GPU boost clock. Tensor Core figures are with sparsity unless noted. RTX PRO support varies by OEM system — confirm the exact SKU before ordering.
What it runs
Run always-on agents connected to enterprise apps, with token generation happening locally instead of on a metered API.
Serve models up to a trillion parameters inside your own network — prompts and data never leave the building.
748 GB of coherent memory means full fine-tunes that would otherwise need a multi-GPU cluster.
Ingest and process large data lakes in memory, end to end, without paging to disk.
Pair GB300 with an RTX PRO 6000 for ray-traced simulation alongside AI training and inference.
Run Linux AI frameworks on Windows through WSL, without a second machine or a second image.
The DGX Spark is the GB10 deskside system — 1 PFLOPS FP4, 128 GB unified memory and models up to 200B parameters, at roughly 240 W. We stock the NVIDIA DGX Spark in 1TB and 4TB, plus GIGABYTE, ASUS and MSI builds of the same platform. It is the right first box for pilots and edge deployments; DGX Station is the step up when a single model needs the full 748 GB pool.
Homepage placement (preview)
A flagship strip directly under the hero, above the category cards — one platform, four numbers, one action.
Tell us the model size, the concurrency and the room it will sit in — including whether a 20 A circuit is available. We will confirm the OEM system, memory configuration and power requirement before you commit.