DGX Station vs an RTX PRO 6000 server
DGX Station puts one GB300 with 252 GB of HBM3e under your desk. An 8x RTX PRO 6000 server has 768 GB on eight cards. Specs, limits and prices compared.

NVIDIA DGX Station is a deskside tower built around one GB300 Grace Blackwell Ultra superchip: one very fast GPU with 252 GB of HBM3e, plus 496 GB of CPU memory it can share. A server with eight RTX PRO 6000 Blackwell cards spreads 768 GB of GDDR7 across eight GPUs in a data centre. Both run the same NVIDIA software. The difference is one big GPU against eight medium ones, and a desk against a rack.
Which one for whom
- One team that trains or fine-tunes large models and wants the hardware in its own office → DGX Station. 252 GB of HBM3e at 7.1 TB/s on one GPU, 748 GB of coherent memory in total, 1,600 W from a 20 A circuit. Public prices on 11 November 2026: from $104,830 at Exxact, $160,195 for the Dell Pro Max with GB300 in the US.
- Serving models to many users, running several models side by side, or a fixed monthly cost instead of a purchase → a server with eight RTX PRO 6000. 8 × 96 GB of GDDR7, 1.6 TB/s per card, power, cooling and a fixed address included. Our server in Sofia rents from $1.80 per GPU-hour, or about $10,510 a month for the whole machine. It is rented into early 2027 and we take reservations for the period after. See the server.
The specifications side by side
| DGX Station (GB300) | 1 × RTX PRO 6000 Server Edition | 8 × RTX PRO 6000 server | |
|---|---|---|---|
| GPUs | 1 Blackwell Ultra GPU | 1 Blackwell GPU | 8 Blackwell GPUs |
| GPU memory | 252 GB HBM3e | 96 GB GDDR7 | 768 GB GDDR7 |
| GPU memory bandwidth | 7.1 TB/s | 1.6 TB/s | 1.6 TB/s per card |
| CPU memory | 496 GB LPDDR5X at 396 GB/s, shared with the GPU over NVLink-C2C (900 GB/s) | the server's own RAM | |
| Coherent memory in total | 748 GB | ||
| FP4 | up to 20 PFLOPS (1) | 4 PFLOPS (1) | 8 × 4 PFLOPS |
| CPU | 72-core Grace (Arm Neoverse V2) | the server's x86 CPUs | |
| Network | ConnectX-8 SuperNIC, up to 800 Gb/s | data centre uplink | |
| Split into isolated GPUs (MIG) | up to 7 | up to 4 | up to 32 |
| Power | 1,600 W system, 20 A circuit | up to 600 W | about 6.5 kW at full GPU load (2) |
| Form | deskside tower, up to 38.7 kg (Dell) | dual-slot PCIe card | 4U rack server |
(1) NVIDIA does not say on these pages whether the FP4 figures assume sparsity, so treat the comparison as rough. (2) Measured from the wall on our own server. The measurements are in What an 8x RTX PRO 6000 server really draws.
One fast GPU against eight medium ones
When a model writes an answer, the GPU reads its weights again for every new token, so memory bandwidth sets the speed for a single user. Here DGX Station is strong: 7.1 TB/s is about four and a half times the 1.6 TB/s of one RTX PRO 6000. A model that fits in 252 GB runs on one GPU, with no need to split it.
Eight RTX PRO 6000 cards have more bandwidth added up, about 12.8 TB/s, but only if the model is split across them. Splitting a model across cards (tensor or pipeline parallelism) means the cards exchange data over PCIe on every step, so you never get the full sum. In return, a server can do something one GPU cannot: run eight different jobs at once. A large model on four cards, two smaller models on two cards each, and nobody waits for anybody.
We do not have our own measurements on a DGX Station, so we give no token rates here. For serving software and how it uses several cards, see vLLM vs Ollama vs SGLang.
748 GB is two kinds of memory
NVIDIA says DGX Station handles models up to 1 trillion parameters. That works because the GPU can reach the 496 GB of LPDDR5X on the Grace CPU as well as its own 252 GB. The two are not the same speed. The HBM3e on the GPU runs at 7.1 TB/s; the CPU memory at 396 GB/s, and the link between them at 900 GB/s. Whatever part of a model or its cache lives in CPU memory is read much more slowly than the part in HBM3e.
In a server with eight RTX PRO 6000 cards, all 768 GB is GPU memory at 1.6 TB/s per card. As a rough rule, weights take one byte per parameter in FP8 and half a byte in FP4, plus room for the cache of each conversation:
- Up to about 200B in FP8, or 400B in FP4: fits in the 252 GB of HBM3e on DGX Station, or on three to five RTX PRO 6000 cards.
- 671B in FP4 (DeepSeek-V3, about 336 GB of weights): on DGX Station it spills over into CPU memory. On eight RTX PRO 6000 cards it fits in GPU memory with room for the cache.
- Around 1 trillion parameters: NVIDIA's figure for DGX Station, using both kinds of memory. Eight RTX PRO 6000 cards hold about 768 GB, so the largest models need FP4 or a second machine.
How these cards compare with the data centre parts that use the same Blackwell Ultra GPU is in B300 vs H200 vs RTX PRO 6000.
Where DGX Station pulls ahead
DGX Station is built for work on one big model. Training and fine-tuning use a lot of memory on one GPU, and here 252 GB of HBM3e with 7.1 TB/s is hard to match with PCIe cards. It also has an 800 Gb/s network card, so two stations or a station and a cluster can share data quickly.
It is a workstation, not a data centre product. It runs Ubuntu with NVIDIA's AI tools, stands under a desk and needs a 20 A circuit, which in Europe usually means a dedicated line. Dell ships it with a second, smaller card for the display, an RTX PRO 2000. Choosing an RTX PRO 6000 Blackwell instead adds $13,560 on Dell's US site.
Where a server pulls ahead
A team of developers or a product with customers sends many short requests at the same time. That is a serving problem, and it rewards many GPUs over one: each card can hold its own copy of a model, or be split into up to four isolated MIG instances, so 32 isolated GPUs in one machine. A server also stays on 24 hours a day with redundant power, cooling and a fixed public address. A desktop needs someone to keep it running and to answer when it does not.
For running your own models without sending data to an API provider, see private LLM hosting.
Price: buying a tower against renting a machine
NVIDIA does not list a price for DGX Station; partners sell it. ASUS, Dell, Exxact, Gigabyte, HP, MSI and Supermicro build it, and orders opened at NVIDIA GTC in March 2026. On 11 November 2026 Exxact shows its DGX Station from $104,830. Dell's US site shows the Dell Pro Max with GB300 at $160,195. Neither price includes tax or the power and space to run it.
At about $10,510 a month for our whole machine, the Exxact price equals roughly ten months of rent, and the Dell price about fifteen. Buying makes sense when you need the same hardware for years and have somewhere to run it. Renting a dedicated GPU server gives one team the whole machine at a fixed monthly price, without a purchase whose value moves with the memory market. You can compare the options on our pricing page.
Questions
Is DGX Station the same as DGX Spark? No. DGX Spark uses the small GB10 chip with 128 GB of shared memory at 273 GB/s and fits in your hand. DGX Station uses the GB300 with 252 GB of HBM3e. We compare the Spark with our server in DGX Spark vs RTX PRO 6000 server.
Why 748 GB and not the 784 GB from the announcement? NVIDIA announced DGX Station in March 2025 with 784 GB of coherent memory. The shipping version lists 748 GB: 252 GB of HBM3e and 496 GB of CPU memory. ServeTheHome reported in March 2026 that 252 GB matches seven of the eight HBM3e stacks enabled, against 288 GB on a full chip.
Can I put a DGX Station in a data centre? It is a tower, not a rack server, and it has one 1,600 W power supply. Exxact lists it with no redundant supply. A rack server has several supplies and is built for a data centre.
Can a DGX Station run several users at once? Yes. It splits into up to seven MIG instances. A server with eight cards splits into up to 32.
Sources
- NVIDIA, DGX Station: GPU and CPU memory, bandwidth, FP4, ConnectX-8, 1,600 W and 20 A, MIG, models up to 1 trillion parameters, partners, checked 11 November 2026.
- Exxact, DGX Station VWS-158270643: "Starting at $104,830.00", one 1,600 W power supply without redundancy, checked 11 November 2026.
- Dell US, Dell Pro Max with GB300: $160,195.00 with RTX PRO 2000, RTX PRO 6000 option +$13,560, weight up to 38.67 kg, checked 11 November 2026.
- ServeTheHome, NVIDIA DGX Station Systems Available At Last: orders at GTC 2026, 252 GB of HBM3e against 288 GB on a full chip, 20 March 2026.
- NVIDIA Newsroom, DGX Spark and DGX Station: 784 GB of coherent memory at the announcement, 18 March 2025.
- NVIDIA, RTX PRO 6000 Blackwell Server Edition: 96 GB GDDR7, 1,597 GB/s, 4 PFLOPS FP4, up to 600 W, MIG.
- DeepSeek, DeepSeek-V3 model card: 671B total parameters.
- NEOXIS measurements of our 8 × RTX PRO 6000 server, published 20 October 2026.


