NVIDIA DGX Station is a desk-side computer built on the GB300 Grace Blackwell Ultra Superchip: one Blackwell Ultra GPU with 252 GB of HBM3e, a 72-core Grace CPU with 496 GB of LPDDR5X, and 748 GB of coherent memory in total. It draws up to 1,600 W, NVIDIA claims up to 20 petaFLOPS of AI compute and models up to 1 trillion parameters, and partners sell it at roughly $100,000, depending on the vendor.
This guide covers what is in the box, what it runs, what it costs to own, and when renting a datacenter GPU by the hour is the better call. It is the larger sibling of the DGX Spark. For the full chip map, see Datacenter GPUs in 2026.
TL;DR
- One GPU, not eight. DGX Station is a single GB300 superchip with 252 GB of HBM3e at 7.1 TB/s, plus 496 GB of CPU memory at 396 GB/s, joined by NVLink-C2C at 900 GB/s.
- NVIDIA does not list a price. Pricing from partners and distributors runs from about $99,000 to $188,500 depending on vendor and configuration (pi3g, updated October 8, 2026), with one distributor quoting about $100,000 as the indicative figure and warning that prices are rising on memory costs.
- Power is 1,600 W on a 20 A circuit, so it needs a proper outlet, not a desk power strip.
- It makes sense for a team that needs a private, always-on box with a very large memory pool. For a few hundred GPU-hours of work, rent instead.
- Verdict: a six-figure purchase pays back only at sustained high utilization. Divide the price by the live hourly rate below to see your own break-even.
DGX Station GB300 specs
Figures are from NVIDIA's DGX Station page. Performance numbers are NVIDIA's claims.
| Spec | DGX Station (GB300) |
|---|---|
| GPU | 1x NVIDIA Blackwell Ultra |
| CPU | 1x Grace, 72-core Neoverse V2 |
| GPU memory | 252 GB HBM3e, 7.1 TB/s |
| CPU memory | 496 GB LPDDR5X, 396 GB/s |
| Total coherent memory | 748 GB |
| CPU-GPU link | NVLink-C2C, 900 GB/s |
| Networking | ConnectX-8 SuperNIC, up to 800 Gb/s; 2x QSFP112 (400 Gb/s per port), 1x 10 GbE RJ45, 1x 1 GbE RJ45 (BMC) |
| Storage | 4x M.2 Gen 5 slots |
| PCIe | 1x Gen 5 x16 and 2x Gen 5 x16 (x8 electrical) |
| MIG instances | 7 |
| Power | 1,600 W, 20 A circuit required |
| AI compute | Up to 20 petaFLOPS (FP4 Tensor Core); 10 PFLOPS FP8/FP6; 5 PFLOPS FP16/BF16 |
| Model size | Up to 1 trillion parameters (NVIDIA's claim) |
NVIDIA's page notes that Tensor Core figures include sparsity unless stated otherwise and does not label which FP4 number is sparse versus dense, so read "20 petaFLOPS" as a ceiling. The platform also supports adding RTX PRO Blackwell cards, including the RTX PRO 6000 Workstation Edition.
An early report is worth knowing: ServeTheHome noted the GPU memory on the shipping part is 252 GB, down from the 288 GB in the original 2025 spec. Quote the 252 GB figure.
Architecture: a superchip with a coherent pool
The Grace CPU and the Blackwell Ultra GPU share one address space over NVLink-C2C. The GPU's own 252 GB of HBM3e is the fast tier, and the 496 GB of LPDDR5X is a slower tier the GPU can reach without a PCIe copy. That is how one GPU gets a 748 GB pool.
The same GB300 silicon sits at the heart of the rack-scale GB300 NVL72, where 72 GPUs are joined by NVLink. DGX Station is one of those superchips in a box, without the rack. For the contrast with a full eight-GPU server, see HGX vs DGX vs NVL72. The GPU generation is covered in the B300 Blackwell Ultra guide.
The key difference from a rented B300 server: the Station has one GPU. An eight-GPU DGX B300 is specified at 2.1 TB of total GPU memory and 144 PFLOPS sparse FP4 (108 dense). The Station holds a large model in its 748 GB pool, but a single GPU's compute and HBM bandwidth set the speed.
What DGX Station runs
NVIDIA says the system supports models up to 1 trillion parameters. Memory arithmetic backs the order of magnitude: at 4-bit weights, 1 trillion parameters is about 500 GB (1 trillion x 0.5 bytes), which fits in the 748 GB pool with room for context. That is a computed fit, not a speed claim.
Good fits:
- Local inference on very large models, especially mixture-of-experts models where only part of the weights is read per token. See the mixture-of-experts and KV cache entries.
- Fine-tuning and evaluation where data cannot leave the building.
- A shared team box. MIG splits the GPU into up to 7 instances.
- Development against the datacenter stack. Same architecture family as GB300 and B300 servers.
Poor fits:
- Multi-GPU training. There is one GPU and a 800 Gb/s NIC, so scaling out means buying more Stations.
- Bursty work. You pay for the whole box whether or not it computes.
- Throughput serving. One GPU's HBM bandwidth caps tokens per second regardless of pool size. Compare with B200's eight-GPU HGX systems, which carry 1,440 GB across 8 GPUs at 64 TB/s.
Station vs Spark vs an eight-GPU server
Where the desk-side boxes sit against a rack server, using only NVIDIA's published figures:
| DGX Spark | DGX Station GB300 | DGX B300 (8 GPUs) | |
|---|---|---|---|
| GPU | Blackwell (GB10) | 1x Blackwell Ultra | 8x Blackwell Ultra SXM |
| Accelerator memory | 128 GB unified | 252 GB HBM3e (748 GB coherent with CPU) | 2.1 TB total |
| Memory bandwidth | 273 GB/s | 7.1 TB/s (GPU), 396 GB/s (CPU) | not compared here |
| Power | 240 W supply | 1,600 W | see NVIDIA datasheet |
| Where it lives | Desk | Under a desk, 20 A circuit | Datacenter |
Sources: Spark, Station, DGX B300. The jump from Spark to Station is about 26 times the memory bandwidth on the GPU side (7,100 / 273), which is what turns a prototyping box into something that can serve a large model at usable speed.
Software and operations
DGX Station runs NVIDIA's CUDA stack, so containers built for Blackwell Ultra servers run unchanged. In practice that means you can develop on the Station and move the identical image to a rented B300 or GB300 for a larger run. Plan for three operational chores that a rental removes: driver and firmware updates, a failed SSD or fan swap on a machine you own, and remote access to a box that lives on your office network. The 1 GbE BMC port exists for exactly that management job.
What it costs to own
NVIDIA's page gives no price and says to contact a partner. Partners named for the Station include ASUS, Dell, Exxact, Gigabyte, HP, MSI and Supermicro. The figures we can cite:
| Source | Figure | As of |
|---|---|---|
| pi3g, European distributor | Indicative about $100,000 excluding VAT and shipping; vendor range about $99,000 (Supermicro) to $188,500 (Dell); expects rises of 10 to 20% or more | October 8, 2026 |
| ServeTheHome | Unofficial estimates heard: around $100K, up to roughly $120K to $125K, unlikely below $80K to $85K; most vendors do not publish list prices | March 20, 2026 |
Treat all of these as estimates, since vendors quote individually. Lead times per pi3g run from 2 to 16 weeks depending on the vendor.
Ownership costs beyond the sticker:
- Power. At the rated 1,600 W, continuous use is 38.4 kWh a day (1.6 kW x 24 h).
- Electrical work. The 20 A circuit requirement may mean a new outlet.
- Cooling and noise. 1.6 kW of heat in an office is a real load.
- Price drift. Distributors report memory costs pushing prices up, and a Rubin-generation successor will follow.
Rent vs buy: the break-even
Break-even in GPU-hours is the purchase price divided by the hourly price of a comparable rented GPU. Using the roughly $100,000 indicative price:
break-even hours = 100,000 / (live hourly price in dollars)
For every $1 of hourly price, break-even is 100,000 GPU-hours, which is about 4,167 days or 11.4 years of continuous use (100,000 / 24 / 365). Read the live price from the box below and divide. Then adjust for:
- Utilization. At 8 hours a day, the payback period triples compared with running around the clock.
- What you rent instead. An eight-GPU server gives about eight GPUs of compute for eight GPUs of hourly price. If your job scales across GPUs, the finished-job cost can favor renting by a wide margin.
- Electricity and operations. The Station adds power and IT time; a rental adds neither.
In plain terms: the Station is a purchase for privacy, latency to your desk, or a steady daily workload on one big model. For everything else, rent a B300 or B200 for the hours you need.
Rent today
Aquanode manages and optimizes GPUs for training and inference workloads, and you can rent the GPUs in the box below on demand. Prices update live. Use them in the break-even formula above.
What's next
NVIDIA's Rubin generation is the next datacenter step; its status and specs are covered in the Rubin guide. A Windows-compatible Station was reported for Q4 2026 by secondary sources, not by NVIDIA's product page, so treat it as unconfirmed.
FAQ
How much does DGX Station cost?
NVIDIA publishes no price. A European distributor quotes about $100,000 excluding VAT and shipping, with a vendor range of roughly $99,000 to $188,500 (pi3g, October 8, 2026). Contact partners for a quote.
What GPU is in DGX Station?
One NVIDIA Blackwell Ultra GPU with 252 GB of HBM3e, joined to a 72-core Grace CPU with 496 GB of LPDDR5X over NVLink-C2C at 900 GB/s.
How much memory does DGX Station have?
748 GB of coherent memory: 252 GB HBM3e at 7.1 TB/s plus 496 GB LPDDR5X at 396 GB/s.
How much power does DGX Station need?
1,600 W, on a 20 A circuit.
Is DGX Station the same as a DGX B300?
No. DGX B300 is an eight-GPU rack server with 2.1 TB of GPU memory. DGX Station is a single GB300 superchip in a desk-side chassis.
Is it better to rent or buy?
Rent unless you will run it at high utilization for years. Divide the price by the live hourly price to see how many GPU-hours break-even takes.
