Yes, for a single site and a modest number of departments, provided you treat it as the first node of a platform rather than as a big workstation. A DGX Station GB300 can serve open-weight models through an OpenAI-compatible API, and StorageReview measured about 9,280 output tokens per second with gpt-oss-120b at 128 concurrent requests on the ASUS ET900N G3. That is enough capacity to put several internal applications behind one endpoint. What makes it a private AI cloud rather than a shared box is the software around the GPU: a gateway that issues keys and budgets, sign-in tied to your directory, logs you can audit, and a way to stop one department from starving the rest. All five GB300 stations on SourceIT are price on request.
What does a private AI cloud need besides the GPU?
Six layers. The Station supplies the first. The other five are software and process, and most of the work lives there.
| Layer | What it does | Typical tools on one Station | What to decide |
|---|---|---|---|
| Model serving | Runs the models and exposes an API | vLLM or NVIDIA NIM, both with OpenAI-compatible endpoints | Which models, and at what precision |
| Gateway | One front door: keys, budgets, rate limits, spend tracking | LiteLLM proxy (per-key and per-team budgets with reset periods); Kong AI Gateway (token limits are an Enterprise feature) | Who gets a key, and what happens when a budget runs out |
| Identity | Links every request to a person or service | Your existing directory and SSO in front of the gateway and apps | Whether shared service accounts are allowed |
| Logging and audit | Records who asked what, which model answered and what it cost | Gateway spend logs plus a log store with access controls | Retention period, and who may read prompts that contain personal data |
| GPU partitioning and scheduling | Stops one workload from taking the whole GPU | MIG (up to 7 isolated instances); Kubernetes GPU operator or NVIDIA Run:ai quotas | Static slices or a shared pool with quotas |
| Data and retrieval | Gives models access to company documents, with permissions | A vector database such as pgvector, Milvus or Qdrant, plus document-level access filters | Which repositories are indexed, and how access rights carry over |
Sources: vLLM and NVIDIA NIM documentation; LiteLLM proxy documentation; Kong AI Rate Limiting Advanced documentation; NVIDIA MIG User Guide (Oct 2026) and Run:ai scheduler documentation; pgvector, Milvus and Qdrant project documentation (all accessed 5 Nov 2026). Throughput from StorageReview's ASUS ExpertCenter Pro ET900N G3 review. SourceIT has not tested these systems.
How do departments reach it?
Through one internal API rather than through the machine itself. vLLM "provides an HTTP server that implements OpenAI's Completions and Chat API", and NIM provides "native OpenAI-compatible API endpoints". Most internal tools, chat front ends and agent frameworks that already work with OpenAI's API can therefore be pointed at your own endpoint by changing a base URL and a key. Staff never need a login on the Station. That matters for security and for uptime: the fewer people with shell access, the fewer accidental changes to a machine everyone depends on. NVIDIA Sync, NVIDIA's tool for connecting to the Station, works over SSH or Tailscale but not through bastion or jump hosts, so plan administrative access separately from user access. Our post on remote access and shared use covers the access patterns in more detail.
How do you stop one department using all of it?
At two levels. At the API level, a gateway enforces budgets and rate limits per key or per team. In LiteLLM, for example, once a key "crosses it's max_budget, requests fail", and a budget_duration such as 30 days resets the budget automatically. At the hardware level, Multi-Instance GPU can split the GB300 into as many as seven isolated instances, from 1g.31gb up to the full 7g.251gb profile in NVIDIA's October 2026 MIG User Guide, with separate memory, cache and fault isolation. MIG suits fixed allocations, such as reserving a slice for an always-on customer-service model. A shared pool with quotas suits bursty work: Run:ai, for instance, gives each project a "deserved" quota and lets it borrow idle capacity, which is then reclaimed when the owner needs it. Can Seven AI Teams Share One DGX Station? explains MIG's trade-offs, and How Should Companies Charge for AI Usage? covers turning that usage into internal charges.
A private AI cloud is a GPU plus a front door, a ledger and a set of house rules. Without the last three, it is just a shared computer.
What should be logged, and who should see it?
Enough to answer three questions later: who sent the request, which model and data sources were used, and what it cost. Gateways track spend per key automatically. LiteLLM needs a database for this. The harder decision is about prompt content. Prompts often contain personal data, and under PDPA section 24 an organisation "must protect personal data in its possession or under its control by making reasonable security arrangements". Logs of prompts are personal data in their own right. Keep them only as long as you need them, restrict who can read them, and include them in your backups and access reviews. Singapore's CSA "Securing Agentic AI" addendum (June 2026) also recommends logging agent actions to your security monitoring, which is far easier when every request passes through one gateway.
What can one Station realistically support?
A meaningful internal workload, with clear limits. The 9,280 tokens per second figure is aggregate output across 128 short requests. With longer prompts (8,192 tokens in, 1,024 out) StorageReview measured 5,023 output tokens per second at 128 concurrent requests, or about 39 tokens per second per stream, which is comfortable reading speed. The limits are structural:
- One node, no failover. If the Station is down, every department's AI is down. Keep a fallback path, such as a cloud API for non-sensitive work, and read Your DGX Station Could Be a Single Point of Failure.
- Model choice. Open-weight models still trail the best closed models on broad benchmarks. On Artificial Analysis's index the top open-weight model scored 33 against 55–60 for the closed frontier. Some tasks will still justify a commercial API.
- Operations. Someone has to patch, monitor and upgrade the serving stack. See Who's Going to Manage It?
Where does the private AI cloud stop and the public cloud start?
Most companies end up hybrid. The private platform handles work where data control matters or volume is high and predictable: internal document search, drafting with confidential material, batch processing. A commercial API remains the better choice for the hardest reasoning tasks, for spiky public-facing demand and for teams that need a frontier model the private platform does not run. Putting both behind the same gateway lets you route by sensitivity and cost, and gives finance one place to see total AI spend. Our posts on sovereign AI and building a private document assistant look at which workloads belong where, and Is GB10 Powerful Enough to Build Your Own Corporate ChatGPT? covers the smaller starting point.
Buying in Singapore
SourceIT lists all five DGX Station GB300 systems, the HP ZGX Fury G1n, MSI XpertStation WS300, ASUS ExpertCenter Pro ET900N G3, GIGABYTE W775-V10-L01 and Supermicro ARS-511GD-NB-LCC, all price on request. The supporting pieces are in stock (listed 5 Nov 2026): the Synology RackStation RS822RP+ at S$2,086.80 for document storage and backups, the D-Link DXS-1210-10TS 10GbE switch at S$1,680.00, and the APC Smart-UPS SRT 3000VA at S$4,165.00. Browse the full range in our NVIDIA DGX Spark and DGX Station collection. 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 existing ChatGPT-based tools connect to a DGX Station? Often, yes. vLLM and NVIDIA NIM both expose OpenAI-compatible endpoints, so many tools only need a new base URL and API key. Features specific to OpenAI's hosted platform will not carry over.
Do staff need to log in to the DGX Station? No. Put a gateway in front of the model server and give users or applications API keys. Keep shell access for a small administration team.
How many departments can share one Station? MIG allows up to seven isolated GPU instances, and a gateway can enforce budgets per team. The practical limit depends on models and traffic. StorageReview measured about 5,023 output tokens per second at 128 concurrent long-prompt requests on one GB300 system.
Is a private AI cloud automatically PDPA compliant? No. Keeping processing on your own hardware removes the overseas-transfer question for that workload, but you still need access controls, logging, retention rules and the other PDPA obligations.
What is the biggest weakness of a one-Station private cloud? It has no redundancy. Plan backups, a documented recovery process and a fallback route, such as a commercial API for non-sensitive tasks, before departments depend on it.
SourceIT Approved
Researched and verified by the SourceIT team for Singapore buyers.
