Yes, if they are 20 people using AI the way most offices do, with a well-chosen model, and you accept slower answers at the busiest moments. Twenty employees rarely send requests at exactly the same second, and a GB10 system serves several requests at once far more efficiently than one at a time. Published tests show GPT-OSS-120B on a DGX Spark rising from about 57 tokens per second for one request to about 353 in total across 32 simultaneous requests. That is roughly 11 tokens per second each at full load, still faster than most people read. Where one GB10 struggles is 20 people pasting long documents at once, or a dense 70B model. Systems such as the ASUS Ascent GX10 (from S$8,100) and NVIDIA DGX Spark (S$8,800) handle a team; a department-wide service belongs on more units or a GB300. Prices at SourceIT, listed 28 Oct 2026.
What do published concurrency tests show?
| Model and test | 1 request | 8 at once (total) | 32 at once (total) | Source |
|---|---|---|---|---|
| GPT-OSS-120B, short prompts (512 tokens) | 57 tok/s | 159 tok/s | 353 tok/s | llama.cpp DGX Spark benchmark, Feb 2026 |
| GPT-OSS-120B, long prompts (8,192 tokens) | 52 tok/s | 126 tok/s | 226 tok/s | llama.cpp, Feb 2026 |
| GPT-OSS-20B, short prompts | 80 tok/s | 272 tok/s | 681 tok/s | llama.cpp, Feb 2026 |
| Llama 3.1 8B FP8, SGLang | 20.5 tok/s | Not stated | 368 tok/s | LMSYS, Oct 2025 |
Total generation throughput, DGX Spark. StorageReview's vLLM tests on a Dell Pro Max GB10 (March 2026) reached up to 671 tok/s on GPT-OSS-120B and 1,585 tok/s on GPT-OSS-20B at up to 64 simultaneous requests. All GB10 systems use the same chip and memory, so results transfer between brands.
What does that mean for each employee?
Divide total throughput by the number of requests in flight. With GPT-OSS-120B and short prompts, eight simultaneous requests each get about 20 tokens per second and 32 get about 11. With long 8,192-token prompts, the figures fall to about 16 and 7. As a rule of thumb, people read at around 5 to 8 tokens per second, so even the busiest case is usable for chat, though a long answer takes a while to finish.
The smaller GPT-OSS-20B changes the picture: 681 tokens per second shared by 32 is over 20 each. For drafting, Q&A and helpdesk-style questions, a 20B to 30B model gives a 20-person team a much faster experience, keeping the 120B model for the harder questions.
How many of your 20 people are actually using it at once?
Fewer than you think. A person asking a question, reading the answer and typing a follow-up spends most of that time reading and typing, not waiting for the model. Twenty staff with the assistant open will often mean only a handful of requests in flight. The exceptions are predictable: the start of the day, the end of a deadline, a training session where everyone tries it together, or an automated job that fires many requests at once. Measure your own pattern during a pilot before buying more hardware.
Twenty users is not twenty simultaneous requests. Size for the peak, not the headcount.
What slows a shared GB10 down most?
- Long prompts. Every pasted document must be processed before the answer starts. GPT-OSS-120B processes about 2,444 tokens per second on DGX Spark at a 2,048-token prompt (llama.cpp), so a 30-page document of around 20,000 tokens takes roughly 10 to 15 seconds to read on its own, and longer in a queue.
- Dense large models. Llama 70B generates about 3 to 4 tokens per second for one user on DGX Spark (LMSYS, Ollama), before anyone else joins.
- Long contexts held in memory. Each conversation's KV cache takes memory. Many long conversations at once can exhaust the 128 GB before compute runs out.
- The wrong software. Simple single-user tools serve requests one after another. For a team, use a server built for batching, such as vLLM, SGLang or TensorRT-LLM, or llama.cpp's server with parallel slots.
- Agents. Automated workflows can fire dozens of requests in a burst. Give them their own model or a schedule.
What about a RAG assistant for 20 people?
A document assistant adds retrieval and longer prompts. One published test is a useful reference: Classmethod deployed NVIDIA's RAG Blueprint on a DGX Spark in August 2026, measured 16 simultaneous connections and estimated it would serve about 10 users comfortably. That suggests one GB10 is a good fit for a team-sized document assistant and that a 20-person rollout should plan for more capacity or careful model choice. Our post Can NVIDIA GB10 Analyse Thousands of Company Documents Without Sending Them to the Cloud? covers the indexing side.
One GB10, two, or a GB300?
| Scenario | Suggested setup |
|---|---|
| Up to about 20 staff, chat and drafting, 20B to 120B MoE model | One GB10 with a batching server |
| 20 staff with heavy document work, or agents plus people | Two GB10s, each running its own copy of the model, behind a simple load balancer |
| 50 to 100+ staff on one endpoint, or 400B-class models | A DGX Station GB300 |
Why two separate GB10s rather than a cluster? Because for many users, separate copies serve more requests. In Openzeka's July 2026 vLLM tests with Qwen3.6-27B, four separate DGX Sparks served about 1,203 tokens per second in total to 64 users, against about 508 for the same four units joined as one cluster. Clustering is for models too big for one box, as we explain in Two NVIDIA GB10 Systems, One AI Model: Does Performance Actually Double? For when a GB300 makes sense, see GB10 vs GB300.
How to pilot it
- Start with a 20B to 30B MoE model as the default and offer a 120B model for harder tasks.
- Serve it with a batching engine and a chat interface with single sign-on.
- Log requests per minute, prompt length and time to first word.
- Run for a month and look at the busiest 15 minutes, not the average.
- Add a second GB10 if peak waits are too long; consider a GB300 if usage keeps growing.
The wider build is covered in Is NVIDIA GB10 Powerful Enough to Build Your Own Corporate ChatGPT? and Can NVIDIA GB10 Run an Entire Company's Private AI Assistant?
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. We can commission units with a batching server and your chosen model. Browse the NVIDIA DGX Spark collection or the AI workstations and servers 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
How many users can one DGX Spark or GB10 support? Around 20 typical office users of a chat assistant is realistic with a mixture-of-experts model and a batching server. Heavy document work, long prompts or dense 70B models reduce that considerably.
How fast is GPT-OSS-120B with 32 users on DGX Spark? llama.cpp's February 2026 benchmark shows about 353 tokens per second in total with short prompts, roughly 11 per user, and about 226 in total with 8,192-token prompts.
Which software should serve a team? One designed for batching, such as vLLM, SGLang or TensorRT-LLM, or llama.cpp's server with parallel slots. Single-user tools handle requests one at a time.
Is it better to cluster GB10s or run them separately for more users? Separately, for models that fit in one unit. Openzeka measured four separate DGX Sparks serving about 1,203 tokens per second to 64 users, against about 508 when clustered.
When do we need a GB300 instead? When a department of 50 or more people shares one endpoint, when models exceed what one or two GB10s hold, or when IT needs a managed, partitionable system.
SourceIT Approved
Researched and verified by the SourceIT team for Singapore buyers.
