This review is built from NVIDIA’s published specifications, reseller listings and third-party coverage rather than hands-on use. Figures are attributed in the text. Independent benchmarks of this system are still scarce.
The NVIDIA DGX Station is the most powerful AI computer you can put under a desk. Built around the GB300 Grace Blackwell Ultra Desktop Superchip, it offers up to 748GB of coherent memory and up to 20 petaflops of FP4 compute. NVIDIA launched it at Computex 2026, OEM systems began shipping through the year, and in October it announced a Windows version.
The honest answer for almost everyone reading this: you do not need one, and the price makes that clear. Resellers list systems at roughly $90,000 to $125,000. For that money you can rent a lot of cloud GPU time, or build several multi-GPU workstations. The DGX Station makes sense for a narrow group: teams that must keep frontier-scale models and data on their own premises, run them every day, and want one quiet machine rather than a server rack. For them it has no real rival.
Table of Contents
- What the DGX Station Offers
- The Two Memory Pools Explained
- DGX Station Pricing in 2026
- What a DGX Station Can Run
- DGX Station Performance: What We Know
- DGX Station for Windows
- OEM Versions and Availability
- Networking and Linking Two Stations
- Sharing One Machine With MIG
- DGX Station vs DGX Spark and RTX Pro 6000
- DGX Station vs Renting GPUs
- Power, Noise and Practicalities
- Who Actually Buys One
- Who Should Buy What
- FAQ
- Planning a DGX Station Deployment
- Verdict
- Sources and Further Reading

What the DGX Station Offers
The specifications read like a server, because in many ways it is one. Pi3g, a German NVIDIA distributor, lists the following.
| Specification | NVIDIA DGX Station (GB300) |
|---|---|
| Superchip | GB300 Grace Blackwell Ultra Desktop |
| CPU | 72-core Grace, Arm Neoverse V2 |
| Memory | 748GB coherent: 496GB LPDDR5X plus 252GB HBM3e |
| Memory bandwidth | 396 GB/s on LPDDR5X, 7.1 TB/s on HBM3e |
| AI compute | Up to 20 petaflops FP4 |
| Networking | ConnectX-8 SuperNIC up to 800 Gbps, plus 10GbE |
| Storage | Four M.2 NVMe slots |
Guru3D described the shared memory pool as the system’s defining feature: models can stay in one address space instead of being split between system memory and graphics memory. That is what lets a single desk-side box hold models that would otherwise need several servers.
The Two Memory Pools Explained
The 748GB headline hides an important detail. It is two very different kinds of memory joined coherently. 252GB is HBM3e attached to the GPU, running at 7.1 TB/s. The other 496GB is LPDDR5X attached to the Grace CPU, running at 396 GB/s.
Pulse reported that shipping units have 252GB of HBM3e, about 12% less than NVIDIA originally planned. That still makes the fast pool enormous by desktop standards: more than three times the 96GB on an RTX Pro 6000.
For language model inference, the split matters a great deal. Weights that sit in HBM3e can be read at data-centre speed. Weights that spill into the LPDDR5X pool are read at roughly a seventeenth of that rate. A model that fits in 252GB will be extremely fast. A model that needs most of the 748GB will work, but generation speed will be limited by the slower pool for the layers stored there.
Plan accordingly. The DGX Station is best for models and contexts that live mostly in HBM, with the large LPDDR5X pool holding overflow, long contexts, data and additional models.
DGX Station Pricing in 2026
NVIDIA did not publish a list price. Tom’s Hardware asked repeatedly at Computex and got no answer, then found an Exxact listing in August at $94,930, confirming the $100,000 ballpark that had circulated.
Other figures fall in the same range. VRLA Tech quoted roughly $90,000 to $115,000. Pulse cited a range of $80,000 to $125,000, noting that DRAM and NAND supply constraints were part of the reason. TechRadar reported mid-range reseller configurations at $94,000 or more, with higher-end options approaching $200,000. A commenter on ServeTheHome questioned a $125,000 figure by comparing it with roughly 25 DGX Sparks.
Expect to pay around $100,000 for a typical configuration, more with storage and support contracts. This is enterprise procurement territory, not a credit card purchase.
What a DGX Station Can Run
With 252GB of fast memory and 748GB in total, the machine can hold models that no other desk-side system can. Guru3D noted NVIDIA’s positioning around running very large models locally, and HP described its version as able to handle trillion-parameter models.
- Frontier-scale mixture-of-experts models at 4-bit. Open models in the several-hundred-billion to trillion-parameter range fit within the coherent pool.
- Large dense models in HBM. 70B to 120B-class dense models at FP8 or FP4 fit entirely in the fast pool.
- Several models at once. A team can keep a large reasoning model, a coding model and embedding models resident together.
- Fine-tuning. The memory and compute make full fine-tuning of mid-sized models and parameter-efficient tuning of large ones practical on one box.
The practical question is not whether a model fits but how much of it sits in the slower pool. Check that before committing to a specific model plan.
DGX Station Performance: What We Know
Independent benchmarks are scarce. Atcyrus titled its profile of the system as the local AI box everyone wants benchmarked, and noted that real runs matter more than spec sheets. At this price, few reviewers have units, and buyers rarely publish results.
What can be said with confidence comes from the architecture. The GB300 GPU with HBM3e at 7.1 TB/s is data-centre class, so models that fit in HBM will generate tokens and process prompts at speeds comparable to a GB300 node slot. The 20 petaflops of FP4 compute give it enormous prefill and batched-serving capacity.
Prompt processing deserves a separate mention. Long documents, retrieval pipelines and coding agents send large contexts with every request, and the time to process that context often matters more to users than generation speed. This is where the compute of the GB300 shines. A model in HBM with plenty of FP4 throughput should handle long prompts far faster than any consumer or workstation card, which is a large part of what makes it practical for teams serving many requests a day.
The weaker point is any work that depends on the LPDDR5X pool. At 396 GB/s, it is faster than a DGX Spark but far slower than HBM. Ask your reseller for benchmarks on your actual models before you buy, and treat vendor slides with caution.
DGX Station for Windows
On October 7, 2026, NVIDIA announced DGX Station for Windows. Tech Insider reported that it moves NVIDIA’s most powerful desktop AI system out of Linux-only territory, with partner systems planned for December. The hardware is the same; the change is software.
That matters for organisations whose staff and tools live on Windows. Data scientists can use familiar applications alongside local models, and IT departments can manage the machine like other Windows workstations. Tech Insider noted that NVIDIA had not published a definitive retail price for the Windows version either.
For teams already comfortable with Linux and NVIDIA’s DGX OS, the Windows version changes little. For enterprise desktops, it removes a real barrier.
Two practical details are easy to miss. First, the Grace CPU is Arm-based, so any x86-only tools in your workflow need Arm builds or replacements; most mainstream AI software already supports Arm, but older internal scripts and some commercial applications may not. Second, NVIDIA positions the machine for AI agents connected to enterprise applications, so expect it to be managed like a server even though it sits beside a desk. Budget time for user accounts, monitoring, backups and security updates, the same as any shared infrastructure.
OEM Versions and Availability
NVIDIA supplies the motherboard and superchip, and partners design systems around it while meeting NVIDIA’s specifications, as ServeTheHome explained. Tech Stories India listed ASUS, Dell, Gigabyte, HP, MSI and Supermicro with availability in the fourth quarter of 2026 for the Windows-ready systems.
Availability through the year was uneven. ServeTheHome reported that HP was targeting August for its system, and some vendors took longer. Exxact and specialist integrators such as VRLA Tech offered systems earlier. If timing matters, ask several vendors for delivery dates.
OEM choice affects support, warranty and service more than performance, since the core hardware is the same.
Networking and Linking Two Stations
The ConnectX-8 SuperNIC provides up to 800 Gbps of networking, and Pi3g notes it supports linking two DGX Stations together. That roughly doubles the memory and compute available to a single job, at the cost of a second machine.
Fast networking also matters for moving data. Loading hundreds of gigabytes of model weights or training data over a slow link wastes hours. Plan for a high-speed connection to your storage before the machine arrives. Our 10GbE guide covers the basics, though this machine is in a different league.
Sharing One Machine With MIG
NVIDIA’s Multi-Instance GPU technology lets the DGX Station split its GPU into up to seven isolated instances, according to Pi3g. Each instance gets its own share of compute and memory, so several people or services can use the machine at once without interfering with each other.
For a team, that changes the economics. One machine can serve several developers’ experiments during the day and run a large batch job overnight. Used this way, the cost per person looks more reasonable than the sticker price suggests.
DGX Station vs DGX Spark and RTX Pro 6000
| System | Memory | Approximate price | Best for |
|---|---|---|---|
| DGX Station (GB300) | 748GB coherent, 252GB HBM3e | About $90,000 to $125,000 | Frontier-scale models on premises |
| DGX Spark | 128GB at 273 GB/s | A few thousand dollars | Individual developers |
| RTX Pro 6000 workstation | 96GB GDDR7 per card | Depends on card count | Teams needing fast 70B-class models |
The comparisons are not close in scale. DGX Spark is a developer box. A workstation with one to four RTX Pro 6000 cards gives up to 384GB of fast GPU memory, split across cards, for well under the DGX Station’s price. The Station’s advantages are a single coherent pool, HBM bandwidth and NVIDIA’s full DGX software and support.
DGX Station vs Renting GPUs
At about $100,000, a DGX Station costs the same as many thousands of hours of rented H100 or B200 time. For workloads that run occasionally, renting is far cheaper. Our RunPod review and Lambda vs Hetzner comparison cover the rental side.
The calculation flips when three conditions hold. The machine runs most of the working day, every day. The data or models cannot leave your premises for legal or commercial reasons. And the team values predictable costs over flexibility. Regulated industries, defence work and research groups handling sensitive data are the obvious buyers.
Depreciation matters too. AI hardware loses value quickly as new generations arrive. Plan to use the machine heavily for its first three years.
Power, Noise and Practicalities
A desk-side machine with this much compute draws substantial power and needs a suitable circuit, cooling and space. Check your office’s electrical capacity and ventilation before ordering, and ask the OEM for power and noise figures under sustained load.
Software is the other practicality. The machine runs NVIDIA’s full stack, so CUDA, TensorRT, vLLM and NVIDIA’s enterprise tools are all available. That makes it far easier to use productively than any non-NVIDIA alternative at this scale, which is part of what the price buys.
Who Actually Buys One
It helps to picture the real customers. A hospital research group that cannot send patient records to a cloud provider, but wants to run a large open model over them. A law firm or bank whose compliance team will not approve external inference for confidential documents. A defence contractor working on systems that cannot touch the public internet. A university lab that wants a shared machine for several PhD students, without the queueing of a central cluster.
What these buyers have in common is that the alternative is not a cheaper GPU. It is either a server rack with its own room, power and cooling, or no local capability at all. Against those options, a single quiet desk-side box with full NVIDIA support looks reasonable, even at six figures.
Individual developers, startups and hobbyists are not the market. For them, the money buys far more through rented GPUs or a smaller workstation.
Who Should Buy What
| If you are | Start with | Why | Link |
|---|---|---|---|
| A team that must run frontier models on premises daily | NVIDIA DGX Station | 748GB coherent, HBM3e, full NVIDIA stack | Request quote |
| A team needing fast 70B to 120B models | RTX Pro 6000 workstation | Far cheaper, very fast | Check price |
| An individual developer | DGX Spark | Same software family, tiny cost | Check price |
| Running large models occasionally | Rented GPUs | Pay only when you use it | Check price |
Frequently Asked Questions About the DGX Station
How much does a DGX Station cost?
NVIDIA has no public list price. Resellers list systems at roughly $90,000 to $125,000, with Exxact at $94,930 in August 2026.
How much memory does it have?
Up to 748GB of coherent memory: 252GB of HBM3e at 7.1 TB/s and 496GB of LPDDR5X at 396 GB/s.
Does it run Windows?
Yes. NVIDIA announced DGX Station for Windows on October 7, 2026, with partner systems from December.
Can it run trillion-parameter models?
Large mixture-of-experts models in that range fit in the coherent pool at low precision. Speed depends on how much sits in the slower LPDDR5X memory.
Is it worth it over renting?
Only if it runs most of the day and your data must stay on premises. For occasional use, renting is far cheaper.
Can several people share it?
Yes. MIG splits the GPU into up to seven isolated instances.
Planning a DGX Station Deployment
Start with your models, not the machine. List the models and context lengths you need, and check how much of each fits in the 252GB of HBM3e. Ask your reseller for benchmarks on those models.
Next, plan the site: power, cooling, noise and a fast network link to your data. Then decide on Linux or Windows, and how you will share the machine across the team with MIG.
Finally, compare the total cost over three years against renting the equivalent GPU hours. If the machine will not be busy most of the time, rent instead.
Verdict
Buy the NVIDIA DGX Station if your team must run frontier-scale models on premises every day, and you value one coherent machine with NVIDIA’s full software stack over a rack of servers. For that use, nothing else on a desk comes close.
The honest caveat is cost and evidence. Around $100,000 buys a great deal of rented compute, independent benchmarks are still scarce, and most of the 748GB runs at a fraction of HBM speed.
Next step: map your models against the 252GB of fast memory, get benchmarks from two resellers on your own workloads, and run the three-year rent-versus-buy numbers before you sign anything.
Sources and Further Reading
- Pi3g: DGX Station specifications, for the memory split, networking and MIG details.
- Tom’s Hardware: listed near $100,000, for real reseller pricing.
- Tech Insider: DGX Station for Windows, for the October announcement.
- Guru3D: 748GB on the desktop, for the coherent memory explanation.
- VRLA Tech: DGX Station vs custom GPU server, for an integrator’s comparison.
On this site: compare the smaller sibling in our DGX Spark review, see the workstation card option in the RTX Pro 6000 Blackwell review, weigh renting with our RunPod review, and check memory needs in how much VRAM for AI.




