DGX Station GB300 or GB300 NVL72? Deskside vs. Rack Scale

At the Naval Postgraduate School, a DGX GB300 now serves more than 1,500 resident students and 600 faculty working on problems from weather prediction to cybersecurity. NVIDIA's account of the installation names the data platforms, racks, cooling, power work and commissioning behind it. That is a rack purchase. A small research group that needs one large model close to its data may need a DGX Station instead.

The shared GB300 name makes this an easy comparison to get wrong. Station is a deskside system with one Blackwell Ultra GPU. The DGX GB300 NVL72 is a liquid-cooled rack with 72 GPUs. The decision begins with who will run the system, how often, and what facility can support it.

DGX Station places one NVIDIA Blackwell Ultra GPU beside a 72-core Grace CPU in a deskside chassis. NVIDIA lists 252 GB of HBM3e GPU memory and 496 GB of LPDDR5X CPU memory, connected through NVLink-C2C. Its published peak is up to 20 PFLOPS of FP4 Tensor Core performance, with the usual precision and sparsity qualifications. NVIDIA positions it for large-model development, local inference, data science, and agent workflows. It can also serve several developers as a local compute node; NVIDIA lists support for up to seven Multi-Instance GPU partitions.

NVIDIA DGX GB300 is a different scale of infrastructure. NVIDIA lists 72 Blackwell Ultra GPUs, 36 Grace CPUs, 20 TB of aggregate GPU memory, and 37 TB of total fast memory for the liquid-cooled rack. Its NVLink switch system links the GPUs inside the rack. The system targets training, post-training, and inference at AI-factory scale. The rack also needs an appropriate fabric and operations plan if it will connect to other racks, storage, or users.

The Station's 748 GB is a coherent CPU-and-GPU system memory figure. It does not mean the Station has 748 GB of HBM on its GPU. Its GPU HBM3e capacity is 252 GB; the other 496 GB is Grace CPU memory. Coherence lets software work across the two memory types, while their bandwidth and behavior still differ.

The NVL72 rack's 20 TB is GPU HBM aggregated across 72 GPUs. NVIDIA separately lists 37 TB of total fast memory when CPU memory is included. That aggregate is useful for distributed workloads, but it should not be described as the HBM capacity of a single GPU or compared directly with one Station's coherent memory as though both figures represent the same resource.

For either system, model parameter count alone is a weak sizing rule. Precision, quantization, context length, batch size, fine-tuning method, and the number of simultaneous users all affect memory and throughput. NVIDIA says DGX Station supports models up to 1 trillion parameters; validate the proposed model and workflow before treating that ceiling as a performance guarantee.

Choose a DGX Station path when the workload fits one deskside system and a small team needs local prototyping, fine-tuning or inference without waiting for shared cluster time. If several people will use it, define access and scheduling; MIG partitions share one GPU rather than multiplying it. An additional RTX PRO GPU is an OEM-dependent visualization and simulation option, not another GB300 accelerator.

Choose the DGX GB300 rack path when the requirement is sustained, large-scale training or high-throughput inference and the organization is prepared to run rack-scale AI infrastructure. The NVL72 architecture is an integrated, liquid-cooled system, so the purchase includes decisions about facility power, cooling distribution, network fabric, storage, operations software, commissioning, and service ownership. A rack quote that covers only compute hardware is incomplete.

If one Station misses the memory or throughput target but a rack is not justified, compare an eight-GPU DGX server or a smaller cluster. That middle path still needs networking, software and operations planning, but it does not presume the full NVL72 facility project.

NVIDIA says the Naval Postgraduate School commissioned its DGX GB300 in July 2026 for campus research and instruction. DDN and VAST provided data infrastructure. Vertiv supplied racks, cooling and power infrastructure and supported installation sequencing, testing, commissioning and fluid management. That list is more useful to a prospective rack buyer than comparing theoretical PFLOPS with a workstation.

You do not need the same partners or campus scale. You do need owners for the same categories: data path, liquid cooling, electrical supply, network fabric, deployment sequence, acceptance testing and long-term operations. If those owners are not yet identified, a rack quote is still a design exercise.

NVIDIA lists 1,600 W total system power and a 20 A circuit requirement for DGX Station. Confirm the exact OEM configuration, circuit, plug, room ventilation, floor placement, network connection, and delivery route with the site team. These checks are smaller than a data-center build, but they still affect whether the machine can be installed and used on arrival.

For the DGX GB300 rack, involve facilities and data-center operations before a commercial quote is finalized. Ask for the rack's electrical design envelope, liquid-cooling interfaces and capacity, residual air-cooling load, floor and receiving requirements, network topology, storage path, and commissioning plan. NVIDIA describes the system as a liquid-cooled rack-scale platform; a standard air-cooled server row should not be assumed sufficient.

The following pages identify concrete paths for discussing each system. Use the exact proposed configuration and current manufacturer documentation for the final technical review.

Uniqcli lists deskside Station-architecture systems from ASUS and MSI. The MSI entry is a barebones, request-pricing configuration; neither product card represents the 72-GPU rack. Use the HPE rack path above for an engineered GB300 NVL72 discussion.

For a Station, share the intended models and frameworks, whether the system is dedicated or shared, data location, storage requirement, software and support expectations, site power, quantity, and desired installation date. If RTX PRO visualization is required, state that explicitly so the supported OEM configuration can be checked.

For an NVL72 rack, add expected training and inference workloads, target throughput and concurrency, existing network and storage designs, facility power and cooling information, receiving constraints, security requirements, and the team that will operate the system. A quote can then define the full deployment boundary rather than only the accelerator count.