Can NVIDIA GB10 Run an Entire Company's Private AI Assistant?
It can for a small company, and it should not for a large one. A single NVIDIA GB10 system, such as the ASUS Ascent GX10 (from S$8,100) or NVIDIA DGX Spark (S$8,800), has the memory to run a capable open model, an embedding model and a document index together, entirely on your premises. For a firm of 10 to about 30 people that is a realistic private assistant. Beyond that, one box runs into five practical limits: how many people ask questions at once, how much memory the whole stack needs, whether staff only see documents they are allowed to, what happens when the box is down, and who looks after it. This post explains each limit and when to add capacity. Prices at SourceIT, listed 28 Oct 2026.
What size of company can one GB10 serve?
| Company size | Realistic setup | Main constraint |
|---|---|---|
| Up to about 10 staff | One GB10: chat, drafting and document Q&A | Choosing the right model |
| 10 to about 30 staff | One GB10 with a batching server; mixture-of-experts model as default | Peak-hour waits; permissions |
| 30 to 100 staff | Two or more GB10s running separate copies, or a pilot team only | Concurrency, availability, administration |
| 100+ staff | A DGX Station GB300 or server-class system, possibly with GB10s for developers | Needs managed, redundant infrastructure |
Our guidance based on published GB10 concurrency tests and NVIDIA's DGX Spark playbooks; your usage pattern matters more than headcount, so pilot first.
Limit 1: how many people ask at once?
A GB10 batches requests well: llama.cpp's DGX Spark benchmark shows GPT-OSS-120B producing about 353 tokens per second in total across 32 simultaneous requests, against 57 for one. Split 32 ways, that is still readable. But long pasted documents, dense 70B models and automated agents eat into that quickly. The full arithmetic is in Can One GB10 Support 20 Employees Using AI at the Same Time?
Limit 2: does the whole stack fit in 128 GB?
A company assistant is not one model. It is usually a main language model, an embedding model to index documents, a vector database, sometimes a reranker or a coding model, plus memory for every conversation in progress. NVIDIA's own "Build and Deploy a Multi-Agent Chatbot" playbook for DGX Spark shows how quickly this adds up: a GPT-OSS-120B supervisor of about 63 GB alongside a coding model and an embedding model uses about 120 GB of the 128 GB by default. For a shared assistant, a 20B to 30B main model leaves far more room for users. See What NVIDIA's 120B and 200B Claims Really Mean for the memory arithmetic.
Limit 3: will staff only see what they should?
This is the limit that has nothing to do with hardware and the one most likely to stop a rollout. If board papers, HR files and client contracts all go into one search index, the assistant can quote them to anyone. A company-wide assistant needs document permissions carried into retrieval, sign-in through your existing Microsoft or Google accounts, and logs you can audit for PDPA purposes. We cover the build in Is NVIDIA GB10 Powerful Enough to Build Your Own Corporate ChatGPT? and indexing in Can NVIDIA GB10 Analyse Thousands of Company Documents Without Sending Them to the Cloud?
A private assistant for the whole company is an IT service, and IT services need a plan for the day the box is off.
Limit 4: what happens when it is down?
One GB10 is one point of failure. It is a desktop, not a server: there is no redundant power supply and no BMC for out-of-band management. During a DGX OS update, a power cut or a hardware fault, the assistant is gone. For a small team that is an inconvenience. For a company that has built processes around the assistant, it is an outage. Two units running separate copies of the model give you both more capacity and a fallback. Add a UPS, as covered in Would You Trust a S$200 UPS to Protect a S$10,000 Workstation?
Limit 5: who looks after it?
Someone has to update the model when a better one appears, patch DGX OS, watch the logs, add new document sources, handle "it gave me a wrong answer" reports and decide what goes in the index. For an SME that is often a few hours a month from an IT person or managed-service partner. Budget for it from the start. The hidden costs of AI hardware are set out in The Hidden Costs of Owning an AI Workstation Nobody Talks About.
What can a company assistant on GB10 do well?
- Answer questions from HR policies, IT guides, SOPs and product manuals.
- Draft and polish emails, proposals and reports.
- Summarise meetings, documents and long email threads. Our post Can a GB10 Transcribe 100 Hours of Meetings? covers the audio side.
- Help staff write and review code or spreadsheets.
- Run scheduled jobs overnight, such as inbox triage, as covered in Can a GB10 Run AI Agents 24/7?
All without prompts or documents leaving your network, which is the reason most Singapore firms consider it. For the cost comparison with cloud services, see Local AI vs Cloud API Cost.
A sensible rollout
- Start with one department and one document set.
- Use one GB10 with a batching server and single sign-on.
- Measure peak usage, waiting times and wrong answers for a month.
- Add a second GB10 for capacity and resilience before opening it to the whole company.
- Move to a GB300-class system when usage outgrows two or three units; see GB10 vs GB300.
Buying in Singapore
SourceIT supplies the ASUS Ascent GX10 (1 TB S$8,100, 2 TB S$8,300, 4 TB S$10,635), the NVIDIA DGX Spark (S$8,800), the HP ZGX Nano G1n (S$11,499) and other GB10 systems, plus DGX Station GB300 towers on request, with local warranty and a GST invoice, and can commission units with your chosen model and serving software. Browse the NVIDIA DGX Spark collection or the ASUS Agent Computer page. 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. Government buyers can purchase through GeBIZ.
Frequently asked questions
Can one GB10 run a private AI assistant for my whole company? For a company of up to about 30 staff with typical usage, yes. Larger companies need more units or a GB300-class system for capacity, resilience and management.
How much memory does a company AI assistant need? More than the model alone. NVIDIA's multi-agent chatbot playbook for DGX Spark, with a 120B supervisor plus coding and embedding models, uses about 120 GB of the 128 GB by default.
What happens if the GB10 fails? The assistant stops. A GB10 is a desktop without redundant power or a BMC, so use two units for important services and protect them with a UPS.
How do we stop staff seeing confidential documents through the assistant? Carry document permissions into the retrieval layer, use single sign-on and keep audit logs. Do not index sensitive files into a shared index without access control.
Does a private assistant keep data in Singapore? Yes. The model, documents and conversations stay on hardware in your office, which helps with PDPA obligations, though you still control what is indexed and how long logs are kept.
SourceIT Approved
Researched and verified by the SourceIT team for Singapore buyers.




