It can be the hardware for one, but not the department. A DGX Station GB300 has the memory (748 GB coherent: 252 GB HBM3e on the GPU plus 496 GB LPDDR5X on the CPU) and the software stack to run a company's private document search, internal chat assistant, model fine-tuning and a few internal AI apps at the same time. NVIDIA itself describes it as a "Deskside Server for AI Teams". What the box does not include is the people, the data pipelines, the governance or a second machine for when it fails. Treat it as a well-equipped lab that still needs staff, rules and a backup plan.
What could one DGX Station run for a company?
| Workload | What it involves | What the Station brings | What you still have to supply |
|---|---|---|---|
| Enterprise search and Q&A (RAG) | Ingest documents, embed them, retrieve and answer with citations | NVIDIA's RAG Blueprint covers multimodal ingestion (text, tables, charts, infographics, audio), embedding and reranking models and an LLM | Clean source documents, access rules per department, a vector database (Elasticsearch by default, Milvus optional) |
| Document intelligence | Extract fields from invoices, contracts, forms and scanned PDFs | OCR, table-structure and page-element extraction NIMs in the same blueprint | Validation steps and a human to check exceptions |
| Internal chat assistant | A private ChatGPT-style tool for staff | vLLM, SGLang and Ollama playbooks; HP adds HP Z Runtime for serving models on the LAN | Sign-in, logging, usage policy and training for staff |
| Model development and fine-tuning | Adapt open models to company data | PyTorch, JAX and NGC containers built for Arm | Someone who knows how to do it safely, plus evaluation data |
| Agents and internal apps | Automate multi-step tasks against business systems | NVIDIA's NemoClaw and OpenShell agent playbooks | Approval steps, least-privilege accounts and audit logs |
From NVIDIA's DGX Station playbook catalogue (build.nvidia.com/station, accessed 4 Nov 2026), the NVIDIA RAG Blueprint on GitHub (accessed 4 Nov 2026) and the HP ZGX Fury G1n product listing. NVIDIA's RAG Blueprint page states no DGX Station-specific support or hardware minimums. SourceIT has not deployed these workloads on a DGX Station.
Can one box serve the whole company at once?
For many mid-sized companies, the throughput is there. In StorageReview's review of the MSI XpertStation WS300 (August 2026), the system served the GPT-OSS-20B model at 22,161 tokens per second across 128 concurrent requests, which averages about 173 tokens per second per request. Its ASUS ET900N G3 review measured 1,766 tokens per second across 32 streams with DeepSeek V4 Flash at FP8, about 55 per stream. These are reviewers' measurements on specific models and settings, not a promise for yours, but they show the Station is built for many users rather than one.
To keep departments from interfering with each other, the GPU can be split into up to seven isolated MIG instances, for example giving HR and Finance their own slice; see Can Seven AI Teams Share One DGX Station?. NVIDIA's Brev playbook can also register the Station as "a shared GPU workspace your team can access from anywhere".
A DGX Station can hold your company's AI workloads. It cannot hold your company's AI judgement. The department is the people and rules around the box.
What does the "department" part cost?
The hardware is the visible cost. Public pricing puts DGX Station systems around S$150,000 depending on OEM, storage and options: public US pricing for most OEMs runs roughly US$93,000–125,000 (Supermicro direct, Exxact and ServeTheHome reports), which at USD/SGD 1.2802 is about S$119,000–160,000 before GST, and Dell's configurations list higher. On top of that:
- People. Someone has to patch the OS, manage users, monitor the GPU and keep models updated. Morgan McKinley's 2026 Singapore guide puts DevOps engineers with 5–10 years' experience at S$80,000–150,000 a year. Even a part-time share of that role is a real cost. See You Bought a S$150,000 AI Supercomputer. Now Who's Going to Manage It?
- Software licences. NVIDIA AI Enterprise is a separate subscription on every OEM version, and some production NIM features depend on it.
- Storage and backup. Company documents, vector indexes, model weights and checkpoints all need a home outside the Station; see A S$150,000 AI Workstation With Just a Few Terabytes of SSD Storage?
- Power and space. The Station draws up to 1,600 W and weighs about 27–41 kg depending on OEM, so it needs a suitable circuit, a UPS and a cool spot.
To see how those costs stack up against paying a cloud API for the same work, try the calculator below. Put a share of an engineer's salary and your software subscriptions into "Other running cost" for a fairer comparison.
Cloud API vs local AI hardware: your numbers
Pick a starting scenario, then change any figure. Prices in Singapore dollars.
Workload
Cloud API price
Local hardware
Cloud prices are list rates read on 18 Oct 2026, before volume, batch or caching discounts. Electricity uses the SP Group tariff for Oct–Dec 2026 and assumes full power draw for the hours entered. “Other running cost” is where staff time, support contracts and software licences such as NVIDIA AI Enterprise go; it is zero by default, which flatters local hardware. The DGX Station price is an indicative public figure; SourceIT quotes it on request. This assumes the hardware can actually serve the volume entered at the quality you need. Check that with the AI hardware calculator before relying on the result.
What governance does an in-house AI department need?
Running AI on your own hardware keeps data on premises, which helps with the PDPA's protection and transfer obligations, but it also makes you responsible for controls a cloud provider would otherwise run. Singapore's frameworks give a ready checklist. IMDA's Model AI Governance Framework for Agentic AI (22 January 2026) says "humans are ultimately accountable" and calls for human approval checkpoints. CSA's "Securing Agentic AI" addendum (17 June 2026) asks for least-privilege agent accounts, sandboxed code, outbound allow-lists and central logging of agent actions. CSA's Cyber Essentials (March 2025) adds an inventory of all AI tools in use, including "shadow AI". Our posts on whether your most powerful computer should be on the internet and where company data goes when using AI go further.
When is a DGX Station the wrong "department in a box"?
- You need guaranteed uptime. One Station has one GPU and one power supply. A company-wide service needs a fallback; see Your DGX Station Could Be a Single Point of Failure.
- Your usage is light. If only a handful of staff use AI a few times a day, a GB10 system or a cloud API may cost far less; compare with Can NVIDIA GB10 Run an Entire Company's Private AI Assistant? and Local AI vs Cloud API Cost.
- Nobody owns it. Without a named owner, the Station becomes an expensive experiment. Decide who runs it before it arrives.
Buying in Singapore
SourceIT supplies five DGX Station GB300 systems, all price on request: the HP ZGX Fury G1n, MSI XpertStation WS300, ASUS ExpertCenter Pro ET900N G3, GIGABYTE W775-V10-L01 and Supermicro ARS-511GD-NB-LCC. For the shared document store, the Synology RackStation RS822RP+ 4-bay NAS is S$2,086.80, and for power protection the APC Smart-UPS SRT 3000VA is S$4,165.00 (both listed 4 Nov 2026). Browse the NVIDIA DGX Spark and DGX Station and AI Workstations and AI Servers collections. 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
Can one DGX Station run enterprise search, a chat assistant and model training together? It can run several at once, and its GPU can be split into up to seven isolated MIG instances. Heavy training will compete with interactive use, so most teams schedule it or give it a dedicated slice.
How many users can a DGX Station serve? It depends on the model and request size. StorageReview measured 22,161 tokens per second across 128 concurrent requests with GPT-OSS-20B on an MSI XpertStation WS300, about 173 tokens per second per request. Larger models serve fewer users.
Do I need NVIDIA AI Enterprise? Not to start, since the Station ships with Ubuntu and NVIDIA's AI developer tools. NVIDIA AI Enterprise is a separate subscription that adds supported production software such as NIM microservices.
Does keeping AI on premises solve PDPA concerns? It helps with data transfer and control, but you remain responsible for security, retention and access. Singapore's IMDA and CSA agentic AI guidance also expects human accountability, least privilege and logging.
Who should own a company DGX Station? A named person or team with Linux, GPU and security skills. Without an owner for patching, user access, backups and model updates, a shared AI system quickly becomes unreliable.
