DGX Station GB300 for Universities: How Research Labs Can Build Their Own AI Infrastructure

A research lab can run a DGX Station GB300 as its own AI node, but the hardware is the easy part. What makes it work is the operating model: who looks after it, how students and projects share one GPU, how research data is handled, and how it fits alongside national compute at NSCC. One Station gives a lab a single Blackwell Ultra GPU with 748 GB of coherent memory, which can be split into up to seven MIG instances. That suits interactive development, data that should stay in the lab, fine-tuning and model serving between allocation windows. It does not replace a supercomputer for multi-node training. This guide covers the operating decisions; for hardware tiers and prices, see our AI workstation guide for universities and research labs.

How does a lab-owned GB300 compare with NSCC national compute?

Source: NVIDIA DGX Station datasheet (Aug 2026) and product page (fetched 4 Nov 2026); NSCC ASPIRE 2A, 2A+ and 2B pages and the NSCC SRA policy ("Updated as of 1 June 2026"), accessed 9 Nov 2026; NSCC press release, 8 Jun 2026. NSCC's per-hour rates are published on its policy page and are not reproduced here.

Who should administer the node?

Give it a named owner and a named deputy before it arrives. In many labs that is a research assistant or postdoc with Linux experience, backed by department or central IT. NUS, for example, runs central HPC clusters (Hopper, with 48 H100 GPUs, and Vanda, with 102 A40 GPUs) for staff and student researchers, and notes that some departments and groups run their own local resources. A lab-owned Station sits in that second group, so the lab carries the admin load.

The admin job is real but bounded. The Station ships with NVIDIA's "Ubuntu with NVIDIA AI Developer Tools" image (Ubuntu 24.04, kernel 6.17, the open NVIDIA driver and CUDA 13.1, per NVIDIA's documentation in November 2026). Firmware and BMC updates follow each OEM's own process. Every Station has a dedicated 1GbE management port (the BMC) with IPMI, Redfish and remote console, which should go on a management network, not the student Wi-Fi. Our security hardening checklist covers accounts, patching and logging, and what happens after your DGX Station arrives covers acceptance and handover.

How do students and projects share one GPU?

NVIDIA says the Station's GPU can be partitioned into as many as seven MIG instances. Profile names seen on the Station include 1g.34gb, 2g.67gb and 7g.269gb (the full profile table isn't published). Two practical points from NVIDIA's MIG playbook shape the operating model: changing the MIG layout needs administrator rights, and it affects every user on the machine. So the layout should be set by the owner on a published schedule, not changed by whoever logs in.

A simple pattern many labs could adopt (our suggestion, not an NVIDIA rule):

• Teaching or development layout: several small MIG instances, one per student group or project, during the day or term weeks.
• Research layout: the whole GPU for one job, booked in blocks (nights, weekends, before a paper deadline).
• Accounts and storage: one Linux account per person, one directory per project with a quota, and no shared logins.
• Booking: a shared calendar works for a small lab; a scheduler or Kubernetes suits a department. NVIDIA doesn't document Slurm or Base Command for the Station, so the choice is yours. See integrating the Station with Kubernetes and our earlier posts on sharing one GPU with MIG and shared remote access.

Most users will reach it over the network, so most labs won't need a display card. The GB300 has no display output of its own, and adding an RTX PRO card over 300 W cuts the GB300's share of the 1,600 W power budget to 950 W.

Treat a lab GB300 as shared infrastructure with a named owner, a booking rule and a data rule, not as one researcher's personal workstation.

How should research data be handled on it?

The ethics approval and data-management plan for each project still decide what data may go on the machine. A lab-owned node keeps data on hardware the lab controls, which can make it easier to meet a funder or partner condition that data stays on site. It also makes the lab responsible for protecting it. Where personal data is involved, the PDPA's protection obligation (section 24) calls for "reasonable security arrangements", and a lost or stolen storage device is covered by the same section. In practice:

• Keep each project's data in its own directory, readable only by that project's members.
• Note which datasets carry ethics or contract conditions, and check them before anyone copies data across projects.
• Plan where checkpoints and outputs go when a project ends; see storage planning for the Station.

This is general information; your institution's research office, data protection officer and ethics board set the actual rules.

How do you justify it against NSCC or cloud?

NSCC and a lab node do different jobs, and the strongest case uses both. The facts that favour NSCC are its scale and its low RIE rate. The facts that make a lab node worth having are the friction points: allocation calls twice a year, an expectation of jobs using at least four GPUs, higher charges for industry- or agency-funded work, and a higher Premium rate again if data isn't shared. Since NSCC's June 2026 policy update, a new Tier 2 route also gives institutions a path to national compute beyond RIE-funded programmes.

A sound justification (our framing) shows: the projects that will use the node each week; the work that can't wait for an allocation window or can't leave the lab; the work that will still go to NSCC once it's been prototyped locally; and the support plan (owner, OEM warranty, spares). Our post on running the Station as a 24/7 batch node and on charging departments for AI usage help with the utilisation side.

Which OEM version suits a lab?

All six OEM versions use the same NVIDIA GB300 board. ServeTheHome found three Stations within about 2% of each other on CPU tests, and StorageReview found the ASUS within 2% of the MSI on 24 of 28 model-serving tests. The differences that matter to a lab are physical and contractual: weight (about 27 kg for the ASUS and HP to about 40 kg for the Supermicro), rack options (HP includes rails for 5U; Supermicro sells a tower or 5U version), cooling extras (GIGABYTE has a leak-detection tray with automatic shutdown; Dell has a redundant pump), and the OEM warranty and onsite terms, which you should confirm in writing for Singapore. Dell specifies 8 A at 240 V for the 1,600 W system; ask facilities to confirm the circuit, and see our office installation checklist.

Buying in Singapore

SourceIT lists DGX Station GB300 systems from ASUS, Dell, HP, MSI, GIGABYTE and Supermicro, supplied through each OEM's channel: the ASUS ExpertCenter Pro ET900N G3, Dell Pro Max with GB300, HP ZGX Fury G1n, MSI XpertStation WS300, GIGABYTE W775-V10-L01 and Supermicro ARS-511GD-NB-LCC. All six are price on request, which is a supplier condition rather than a SourceIT choice. Universities buy through their own procurement systems (NUS and NTU use SAP Ariba), so we can quote against your sourcing event, validate the configuration and help with the end-user information suppliers ask for (see KYC and EUC requirements explained). For writing the specification, see DGX Station GB300 procurement for universities and government agencies. To get 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

Can one DGX Station GB300 serve a whole research group? Yes, for development, fine-tuning and inference. NVIDIA says the GPU can be split into up to seven MIG instances, but changing the layout affects all users, so the owner should set it on a schedule.

Does a lab-owned Station replace NSCC? No. NSCC's ASPIRE 2B has 1,536 H200 GPUs for large multi-node work. A lab node suits daily development and data that should stay in the lab; the two work best together.

Is NSCC compute free for university researchers? No. NSCC charges for allocated resources on a cost-recovery basis. RIE-funded and Tier 2 projects pay the lowest rate, and industry- or agency-funded work pays more.

Do we need a graphics card to use the Station? Only if someone will sit at it with a monitor. The GB300 has no display output, and a card over 300 W cuts the GB300's power share to 950 W, so remote-access labs usually skip it.

Who should own the machine day to day? A named person with Linux experience and a deputy, backed by department or central IT, with a published booking rule and a data-handling rule for each project.