From DGX Station to GB300 NVL72: How NVIDIA Scales AI From Your Desk to the Data Centre

By sharing architecture rather than size. NVIDIA describes the DGX Station GB300 as a place to "rapidly develop, fine-tune, and inference with large AI models and long-running AI agents locally" and then "locally and seamlessly deploy to the data center or cloud". Its DGX Station Development Guide puts it more plainly: "Applications can be developed and debugged locally with DGX Station and then brought to massive scale on the cloud." The link to GB300 NVL72 is the hardware underneath. Both use the Blackwell Ultra GPU, the Arm-based Grace CPU and 900 GB/s NVLink-C2C between them. Code, containers and quantised models built on the desk run on the same instruction sets and precision formats in the rack. One thing does not carry over: a Station has one GPU and no NVLink to other GPUs, so multi-GPU performance has to be proven on the target system.

How does a DGX Station compare with a GB300 NVL72?

Source: NVIDIA DGX Station product page (updated 4 Nov 2026) and DGX Station Development Guide (updated 30 Oct 2026); NVIDIA GB300 NVL72 page (accessed 6 Nov 2026); NVIDIA Inside Blackwell Ultra blog (22 Aug 2025); OEM rack power figures (Supermicro, Vertiv, Schneider Electric). Note that NVIDIA's Station pages speak of deploying "to the data center or cloud" and say the Station "features the same architecture found in cloud-scale NVIDIA solutions"; they do not name NVL72. The pairing in this guide is ours, based on the shared GB300 platform.

What carries over from the desk to the rack?

• The CPU architecture. Both are Arm64 (aarch64) Grace systems. If your Python wheels, compiled libraries and containers build and run on the Station, the CPU side of your stack is already ported. This is the step that catches most x86 teams; see the 72-core Grace Arm software guide.
• The GPU architecture. The Station's GPU is a Blackwell Ultra, as in the rack. CUDA kernels, NVFP4 and FP8 quantisation, TensorRT-LLM engines and vLLM or SGLang settings are validated on the same GPU generation.
• The CPU-GPU memory model. NVLink-C2C coherence works the same way, so code that relies on the GPU reading Grace memory behaves alike, at similar LPDDR5X speeds.
• The software stack. The Station ships Ubuntu 24.04 with NVIDIA's CUDA, NCCL, cuDNN, TensorRT, PyTorch, Nsight and container toolkit. NGC containers are multi-architecture. An NVIDIA AI Enterprise licence, sold separately, covers supported deployment.

The DGX Station lets you get the code right on the same chip. NVL72 is where you find out whether 72 of those chips can work as one, and only the rack can answer that.

What does not carry over?

• Multi-GPU performance over NVLink. With one GPU, you cannot measure tensor-parallel or expert-parallel scaling across NVLink. Two Stations cabled back-to-back connect over ConnectX-8 RoCEv2. NVIDIA's own guide says "there is no inter-node NVLink". That link is enough to test that NCCL, Ray or vLLM multi-node code runs correctly. Its bandwidth is 100 GB/s per direction in total, against 900 GB/s per direction per GPU on NVLink, so it says little about performance in the rack.
• Memory per GPU. A Station GPU has 252 GB of HBM3e, and an NVL72 GPU has 288 GB. Budgets set on the Station are conservative for the rack. But the Station's 496 GB of LPDDR5X lets a model spill out of HBM in ways the rack may not need, so do not tune for spillover you will not have.
• Scheduling and operations. A Station is one machine you can log into. A rack is shared through a scheduler, with job queues, quotas, monitoring and failure recovery. Package jobs so that they restart from checkpoints.

How do you move a workload from DGX Station to NVL72, step by step?

1. Containerise on Arm64. Build every service as an aarch64 container image on the Station, pinning CUDA, driver and framework versions. Check for libraries that assume 4 KB memory pages, because NVIDIA's porting guide notes the 64 KB page-size issue on Grace.
2. Get the model right on one GPU. Fine-tune, quantise (for example to NVFP4) and evaluate quality on the Station. Use the 748 GB of coherent memory to iterate on models that would not fit in a smaller workstation. Our post on 748 GB of coherent memory covers this.
3. Write the parallel version early. Express tensor, pipeline and expert parallelism through your framework's configuration, not hard-coded device counts. Run it at a size of one on the Station so that the code paths are exercised.
4. Test multi-node correctness with two Stations. Cable two units back-to-back with two 400G QSFP112 cables, set MTU 9000 and validate RoCE, as NVIDIA's connect-two-stations guide documents. Then run NCCL and multi-node inference or training to catch bugs before you pay for rack time. See DGX Station networking.
5. Profile what will not change. Use Nsight to find compute and memory bottlenecks inside one GPU. Those carry over. Communication time will not, because NVLink is about nine times faster per GPU than the Station's network link.
6. Book time on the target. Run a short scaling test on NVL72 capacity at a cloud provider or in a data centre. Keep tensor and expert parallelism inside the 72-GPU NVLink domain, and data parallelism across racks. Measure tokens per second and time per step at the sizes you will use.
7. Keep the Station in the loop. After deployment, the Station remains the place to reproduce bugs, test new model versions and run private evaluations without queueing for the rack.

For the rack itself, read how NVSwitch connects 72 GPUs and what 37 TB of memory enables. If your next step is an 8-GPU server rather than a rack, see DGX Station vs GPU servers.

Who should work this way?

Teams that will deploy on GB300-class infrastructure in the cloud or a data centre, and want developers to iterate without queueing for shared GPUs. It also suits teams that must keep data and models in-house during development, and teams that need to port x86 software to Arm before they commit to rack capacity. If your production target is x86 servers with PCIe GPUs, the Station's Grace CPU is a less direct match, and a workstation on the same architecture as production may suit you better.

Buying in Singapore

SourceIT lists six DGX Station GB300 systems, all price on request: the HP ZGX Fury G1n, MSI XpertStation WS300, Dell Pro Max with GB300, ASUS ExpertCenter Pro ET900N G3, GIGABYTE W775-V10-L01 and Supermicro ARS-511GD-NB-LCC. For HP and MSI side by side, see HP ZGX Fury vs MSI DGX Station. To link two Stations back-to-back, the HP ZGX 0.4 m QSFP112 DAC is S$365.00 (listed 6 Nov 2026); you need two. For a small lab switch, the MikroTik CRS804-4DDQ is S$2,255.00 (listed 6 Nov 2026). It has only four 400G ports, so it is not a cluster fabric. For dataset and checkpoint storage, the Micron 7600 PRO 3.84 TB Gen5 U.2 SSD is S$5,183.00 (listed 6 Nov 2026). We do not list GB300 NVL72 racks; for data-centre projects, ask us to source. For a quotation, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.

Frequently asked questions

Does NVIDIA say DGX Station is for developing before scaling up? Yes. The DGX Station Development Guide says applications "can be developed and debugged locally with DGX Station and then brought to massive scale on the cloud". The product page says you can "locally and seamlessly deploy to the data center or cloud". The pages do not name NVL72 specifically.

Will code written on a DGX Station run on GB300 NVL72? Code built for the same Blackwell Ultra GPU and Arm64 Grace CPU should run, given matching CUDA and driver versions. Multi-GPU performance must still be tested on the rack, because the Station has no NVLink to other GPUs.

Can I test tensor parallelism on a DGX Station? Only for correctness, by connecting two Stations over ConnectX-8 RoCEv2. That link is 100 GB/s per direction in total, against 900 GB/s per direction per GPU on NVLink, so scaling results will differ.

Is the DGX Station GPU identical to an NVL72 GPU? Both are Blackwell Ultra with NVLink-C2C to Grace. The Station's GPU has 252 GB of HBM3e at 7.1 TB/s; NVIDIA lists 288 GB at 8 TB/s for Blackwell Ultra in the rack.

Do I need NVIDIA AI Enterprise for this workflow? Not to develop. The Station ships with NVIDIA's AI developer tools on Ubuntu. NVIDIA AI Enterprise is a separate subscription for supported production deployment, and NVIDIA says it can fast-track deployment from the Station.