NVIDIA DGX Station

Scroll sideways for more →

NVIDIA DGX Station is a line of deskside artificial intelligence workstations from NVIDIA, each marketed as a "personal AI supercomputer" that puts data-center-class compute next to a developer's desk rather than in a server room. The line began in 2017 with a water-cooled tower built around four Tesla V100 GPUs, was rebuilt in 2020 around four A100 GPUs as the DGX Station A100, and was reinvented again in 2025 around a single NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip. ^[1][2][4] The current GB300 model carries up to 748GB of coherent memory and up to 20 petaflops of FP4 compute, and NVIDIA says it can run AI models of up to 1 trillion parameters locally. ^[2][11]

The DGX Station sits between NVIDIA's smaller NVIDIA DGX Spark desktop unit and its rack-scale data-center systems, and it is part of the broader NVIDIA DGX family. This article covers the product line and its history. For the exhaustive GB300 specification table, the trillion-parameter claim in detail, NVIDIA OpenShell, and the separate Windows edition, see NVIDIA DGX Station for Windows. ^[2][17]

History of the DGX Station

Across three generations the DGX Station has kept one idea constant, a full data-center-class AI machine that plugs into a wall outlet and sits under a desk, while the silicon inside changed completely. The first two generations packed four discrete GPUs into a tower; the third replaced that with a single CPU-plus-GPU superchip.

Scroll sideways for more →

The original DGX Station (2017)

NVIDIA unveiled the first DGX Station at its GPU Technology Conference in May 2017, alongside a Volta-based refresh of the rack-mounted DGX-1. It was an entirely new product rather than an upgrade of an earlier machine, positioned as a quiet, standalone workstation for organizations that wanted to train deep neural networks without access to a climate-controlled data center. ^[4][5] NVIDIA's own white paper titled it "The First Personal AI Supercomputer." ^[6]

The machine held four Tesla V100 Tensor Core GPUs (the initial configuration used the 16GB V100, and NVIDIA later offered a 32GB version) fully connected in a four-way topology using second-generation NVLink. That gave it 20,480 CUDA cores and 2,560 Tensor Cores. The rest of the system was a 20-core Intel Xeon E5-2698 v4 CPU, 256GB of DDR4 RDIMM memory, four 1.92TB SSDs (three in a RAID 0 data array plus one for the operating system), three DisplayPort outputs, and dual 10 gigabit Ethernet. The whole thing was water-cooled, ran inside a 1,500W power envelope quiet enough for an office (NVIDIA quoted under 35 dB), and launched at 69,000 US dollars. ^[5][6][7] NVIDIA rated the 16GB launch configuration at up to 480 teraflops of mixed-precision performance and the later 32GB configuration at 500 teraflops. ^[5][6]

DGX Station A100 (2020)

The second generation, the DGX Station A100, was announced on November 16, 2020, at the SC20 supercomputing conference. NVIDIA kept the deskside tower form factor and the self-contained cooling system that let the machine plug into a normal wall socket, but swapped in four Ampere-architecture A100 Tensor Core GPUs and pitched it as an "AI data-center-in-a-box." ^[8][9] Charlie Boyle, NVIDIA's vice president and general manager of DGX systems, said the product "brings AI out of the data center with a server-class system that can plug in anywhere." ^[8]

It came in two memory configurations: four A100 40GB GPUs for 160GB of total GPU memory, or four A100 80GB GPUs for 320GB. The four GPUs were fully connected with third-generation NVLink, the CPU was a single 64-core AMD EPYC 7742, and the system supported up to 512GB of DDR4 memory. NVIDIA rated it at 2.5 petaflops of AI performance. A distinctive feature was support for Multi-Instance GPU (MIG), which could partition the GPUs into as many as 28 separate instances so a small team could share one machine without interfering with each other's jobs. ^[8][9][10] NVIDIA sold the DGX Station A100 through its reseller partner network rather than publishing a fixed consumer price. ^[8][10]

DGX Station GB300 (2025)

The current generation was announced on March 18, 2025, at GTC, launched together with the smaller DGX Spark under the banner of "personal AI computers." ^[1][3] It is a clean break from the previous two designs. Instead of four discrete GPUs, it is built around a single NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip, which fuses a 72-core NVIDIA Grace CPU and an NVIDIA Blackwell Ultra GPU into one coherent-memory package. NVIDIA CEO Jensen Huang framed the two new machines as a new class of computer "designed for AI-native developers." ^[1][3]

The GB300 Grace Blackwell Ultra model

At the heart of the current DGX Station is the GB300 Grace Blackwell Ultra Desktop Superchip. It pairs one 72-core Grace CPU (built on Arm Neoverse V2 cores) with one Blackwell Ultra GPU, joined by NVLink-C2C, a chip-to-chip link that carries 900 GB/s between the two and lets them share a single coherent pool of memory. This is the same class of silicon NVIDIA uses in its rack-scale NVIDIA GB300 NVL72, packaged here for one desktop unit. ^[2][11][13]

Scroll sideways for more →

The 748GB figure combines the GPU's high-bandwidth memory with the CPU's larger but slower memory, so a very large model and its working data can live in one address space rather than being copied across a slow bus. The built-in ConnectX-8 SuperNIC runs at up to 800 Gb/s, which is what lets two or more DGX Stations be lashed together for a job that outgrows one box. ^[2][11][12] This page summarizes the hardware; the full specification table, including every precision tier, storage, and total system power, lives on the NVIDIA DGX Station for Windows page, which documents the identical GB300 platform. ^[2][17]

Software: DGX OS versus the Windows variant

The DGX Station ships with NVIDIA DGX OS, an Ubuntu-based Linux distribution (ServeTheHome reported the shipping systems run a DGX OS build derived from Ubuntu 24.04) preloaded with NVIDIA's CUDA-X AI stack, container runtime, and developer tools. Linux has been the native environment for the whole DGX line since 2017, and it remains the default for the GB300 model. ^[2][12][6]

In June 2026 NVIDIA introduced a second software flavor of the same hardware, the DGX Station for Windows, developed with Microsoft so that the NVIDIA AI stack runs on Microsoft Windows (with Windows Subsystem for Linux available for Linux-only tooling) and adding an agent runtime called NVIDIA OpenShell. That variant has its own detailed page, NVIDIA DGX Station for Windows (linked in the intro and See also). The two share the same GB300 silicon, up to 748GB of coherent memory, and the trillion-parameter claim; they differ in operating system and target buyer. ^[17]

Deskside AI for agents (SIGGRAPH 2026)

At SIGGRAPH 2026, described in an NVIDIA blog post dated July 20, 2026, NVIDIA positioned the DGX Station as the deskside home for agentic AI. The pitch is that a single machine can run a frontier agent stack entirely on-premises. According to NVIDIA, "On DGX Station, NVIDIA Agent Toolkit brings together NVIDIA NemoClaw, the NVIDIA Nemotron 3 Ultra open model ... and a secure runtime in a single local system, no internet required." ^[14]

The NVIDIA NemoClaw agent blueprint runs on the NVIDIA Agent Toolkit using Nemotron 3 Ultra, which NVIDIA describes as a "frontier 550-billion-parameter open model" optimized to run on DGX Station GB300 systems. NVIDIA says the stack "can be running in roughly 30 minutes" and published step-by-step playbooks on its DGX Station project page, including "Connect Two DGX Stations for Distributed Workloads" and a guide to running NemoClaw with a local LLM on a dual DGX Station. The framing NVIDIA used was blunt: "Super agents have arrived on the desktop." ^[14][15] Running a 550-billion-parameter model this way trades the recurring per-token cost of a cloud API for a one-time hardware purchase, which is the economic argument NVIDIA makes for keeping agents local. ^[14]

Availability and partners

When it was announced in March 2025, the GB300 DGX Station was slated to arrive "later in 2025." ^[1][3] In practice, orders opened about a year later, around GTC 2026. On March 17, 2026, Tom's Hardware reported that the DGX Station was "now available to order and will begin shipping in the coming months," and ServeTheHome titled its March 20, 2026 coverage "Available At Last." ^[11][12] As of the SIGGRAPH 2026 timeframe in mid-2026, the systems were orderable and shipping was rolling out through the year. ^[11][14]

NVIDIA designs the reference system and its OEM and system-integrator partners build and sell their own branded configurations. The launch partners were ASUS, Dell Technologies, GIGABYTE, MSI, and Supermicro, with HP joining later in the year; NVIDIA's product page also lists Exxact, and BIZON sells a GB300 "AI Station" built to the DGX Station architecture. ^{[2][11][18][19]} Examples of branded units include Exxact's Valence DGX Station and MSI's XpertStation WS300. NVIDIA did not publish a list price; ServeTheHome estimated retail configurations in roughly the 100,000 to 125,000 US dollar range. ^[12][18][19]

Where it sits in NVIDIA's lineup

The DGX Station is the middle tier of NVIDIA's "personal AI computer" range and the top of its deskside range.

Scroll sideways for more →

Below the DGX Station is the DGX Spark, a much smaller and cheaper unit built on the GB10 Grace Blackwell Superchip with 128GB of unified memory and about 1 petaflop of FP4, marketed as "the world's smallest AI supercomputer" for individual developers and priced near 3,000 to 4,000 US dollars. ^[1][11] Above the DGX Station are the rack-scale machines, the DGX B300 and systems such as the GB300 NVL72 and GB200 NVL72, which deliver many times the compute and memory but live in a server room rather than on a desk. The DGX Station is best understood as the deskside step between those two worlds. ^[2][11][12] For buyers who prioritize double-precision HPC throughput over AI density, ServeTheHome reported that Supermicro also offers a DGX Station variant based on the earlier Blackwell B200 GPU. ^[12]

Significance

The DGX Station line tracks NVIDIA's answer to a recurring question: where should serious AI work run when the cloud is too slow, too expensive, or too exposed for sensitive data, but a normal workstation is too weak? In 2017 that answer was four V100s under a desk; by 2025 it was a single Grace Blackwell Ultra superchip capable, on NVIDIA's numbers, of holding a trillion-parameter model in local memory. ^[4][6][2] The 2025 to 2026 generation lands at the moment the industry's attention shifted from chatbots to autonomous agents, and NVIDIA has leaned into that, using SIGGRAPH 2026 to recast the machine as an on-premises home for agent stacks like NemoClaw and Nemotron 3 Ultra. ^[14] The through-line is on-premises control: keeping model weights, proprietary data, and inference cost inside an organization instead of renting them from a cloud. ^[16][17]

Limitations and caveats

Several of the headline numbers are NVIDIA's own vendor claims and deserve context.

• The trillion-parameter and 20-petaflop figures depend on low precision. Running a model of up to 1 trillion parameters locally, and the up-to-20-petaflop compute figure, both rely on FP4, a 4-bit format, which means the model must be quantized. These are NVIDIA's figures under favorable precision assumptions, not a guarantee for any model at full precision. ^[2][14][16]
• The shipping memory is smaller than first announced. NVIDIA's March 2025 announcement described 784GB of coherent memory; the shipping 2026 product carries 748GB. ServeTheHome attributes the difference to a binned Blackwell Ultra (B300) GPU with seven of eight HBM3e stacks enabled, giving 252GB of GPU memory instead of 288GB, roughly 12 percent less memory and bandwidth (about 7.1 TB/s), and reports the shipping FP4 rate closer to 15 petaflops dense than the product page's higher figures. ^[1][2][12]
• No official pricing. NVIDIA has never published a list price for the GB300 DGX Station; the numbers in circulation are reseller estimates, and configurations vary by partner. ^[11][12]
• Single-superchip design. Unlike the four-GPU 2017 and 2020 machines, the current model has one GPU, so it cannot be expanded with more GPUs internally; scaling beyond one unit means networking two or more DGX Stations together over ConnectX-8. ^[2][12]

ELI5

Think of the DGX Station as a supercomputer shrunk down to fit under a desk and plug into a normal wall socket, so a researcher can train and run big AI models in their own office instead of renting time in a giant data center far away. The first one, in 2017, stuffed four powerful graphics chips into a quiet, water-cooled tower. The newest one, from 2025, uses a single super-chip that glues a big processor and a big graphics chip together and shares one huge pile of memory, big enough, NVIDIA says, to hold an AI model with a trillion adjustable numbers in it. In 2026 NVIDIA started pitching it as the machine that runs your own private AI "agents" (programs that can plan and do tasks for you) right on your desk, without sending anything to the internet.

See also

• NVIDIA DGX Station for Windows
• NVIDIA DGX Spark
• NVIDIA DGX
• NVIDIA GB300 NVL72
• NVIDIA Blackwell
• NVIDIA Grace
• Nemotron 3
• NVIDIA RTX PRO 6000

References

1. NVIDIA Newsroom. "NVIDIA Announces DGX Spark and DGX Station Personal AI Computers." March 18, 2025. https://nvidianews.nvidia.com/news/nvidia-announces-dgx-spark-and-dgx-station-personal-ai-computers ↩
2. NVIDIA. "NVIDIA DGX Station: The Ultimate Desktop AI Supercomputer." Product page. https://www.nvidia.com/en-us/products/workstations/dgx-station/ ↩
3. Wiggers, Kyle. "Nvidia announces two 'personal AI supercomputers.'" TechCrunch, March 18, 2025. https://techcrunch.com/2025/03/18/nvidia-announces-two-personal-ai-supercomputers/ ↩
4. NVIDIA Newsroom. "NVIDIA Advances AI Computing Revolution With New Volta-Based DGX Systems." May 2017. https://nvidianews.nvidia.com/news/nvidia-advances-ai-computing-revolution-with-new-volta-based-dgx-systems ↩
5. Ragan, Brandon. "NVIDIA Volta-Powered DGX-1 And DGX Station AI Supercomputers Debut At GTC 2017." HotHardware, May 2017. https://hothardware.com/news/nvidia-volta-powered-dgx-1-and-dgx-station-ai-supercomputers-debut-at-gtc-2017 ↩
6. NVIDIA. "NVIDIA DGX Station: The First Personal AI Supercomputer." White paper (WP-V01). https://images.nvidia.com/content/newsletters/email/pdf/DGX-Station-WP.pdf ↩
7. Kennedy, Patrick. "NVIDIA DGX Station Upgraded to Tesla V100." ServeTheHome, October 28, 2017. https://www.servethehome.com/nvidia-dgx-station-upgraded-tesla-v100/ ↩
8. NVIDIA Newsroom. "NVIDIA DGX Station A100 Offers Researchers AI Data-Center-in-a-Box." November 16, 2020. https://nvidianews.nvidia.com/news/nvidia-dgx-station-a100-offers-researchers-ai-data-center-in-a-box ↩
9. Kennedy, Patrick. "NVIDIA DGX Station A100 Brings 320GB HBM2e to Desktops." ServeTheHome, November 16, 2020. https://www.servethehome.com/nvidia-dgx-station-a100-brings-320gb-hbm2e-to-desktops/ ↩
10. Alcorn, Paul (via wccftech). "NVIDIA Announces DGX Station A100 With Upgraded 80 GB A100 Tensor Core GPUs, Up To 320 GB Memory & 2.5 Petaflops of AI Horsepower." wccftech, November 16, 2020. https://wccftech.com/nvidia-dgx-station-a100-with-upgraded-80-gb-a100-tensor-core-gpus/ ↩
11. Ridley, Jacob. "Nvidia launches DGX Station with its bleeding-edge GB300 Grace Blackwell Superchip, now available to order and will begin shipping in the coming months." Tom's Hardware, March 17, 2026. https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidia-launches-dgx-station-with-its-bleeding-edge-gb300-grace-blackwell-superchip-now-available-to-order-and-will-begin-shipping-in-the-coming-months ↩
12. Kennedy, Patrick. "NVIDIA DGX Station Systems Available At Last: GB300 and GB200 Workstations For Your Desktop." ServeTheHome, March 20, 2026. https://www.servethehome.com/nvidia-dgx-station-systems-available-at-last-gb300-gb200-workstations-for-your-desktop/ ↩
13. Mujtaba, Hassan. "NVIDIA DGX Station Upgraded With GB300 Blackwell Ultra Desktop Superchip: 748 GB Memory, 20 PFLOPs AI Compute & AI Ready." wccftech, March 2026. https://wccftech.com/nvidia-dgx-station-upgraded-gb300-blackwell-ultra-desktop-superchip/ ↩
14. NVIDIA Blog. "At SIGGRAPH, NVIDIA Advances Graphics and Simulation With Agentic and Physical AI." July 20, 2026. https://blogs.nvidia.com/blog/siggraph-news-2026/ ↩
15. NVIDIA. "NVIDIA DGX Station" project and playbooks page. https://build.nvidia.com/station ↩
16. Deutscher, Maria. "Nvidia's DGX Station is a desktop supercomputer that runs trillion-parameter AI models without the cloud." VentureBeat, 2026. https://venturebeat.com/infrastructure/nvidias-dgx-station-is-a-desktop-supercomputer-that-runs-trillion-parameter ↩
17. NVIDIA Newsroom. "NVIDIA DGX Station for Windows Puts a Trillion-Parameter AI Supercomputer on Every Enterprise Desk." June 1, 2026. https://nvidianews.nvidia.com/news/nvidia-dgx-station-for-windows-puts-a-trillion-parameter-ai-supercomputer-on-every-enterprise-desk ↩
18. Exxact Corporation. "Exxact Valence NVIDIA DGX Station, 1x NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip." Product page. https://www.exxactcorp.com/Exxact-VWS-158270643-E158270643 ↩
19. BIZON. "AI Station with NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip (NVIDIA DGX Station GB300 architecture)." Product page. https://bizon-tech.com/gb300.html ↩