Module 5: Buying guide with prices in four countries · Lesson 5.5 · about 20 min
A server for a team or office
Sharing one local AI machine with a team over the office network. What size to buy, how many people it can serve and what to set up.
Module 5 · about 20 minutes · read first: 5.4 and 11.1. You’ll be able to size and set up one machine that a small team shares.
1. Why this matters
Buying one strong machine for five people often costs less than five strong laptops, and keeps data in one place you control. It needs a little planning: memory for the model, speed for the moments when requests overlap, and network safety.
2. How many people can one machine serve?
Think of the office kettle. Ten people share one kettle happily, because they rarely all want tea at the same moment, except at 11 o’clock. An AI machine is the same. Office use is bursty: people type, read and think between questions. If each person keeps the model busy about 5% of the working day, a machine that is fast for one person can serve 10 to 20 light users. Heavy users (agents, batch jobs, coding assistants) need more, and a serving engine that batches requests (handles several at once, Module 15).
With users, each busy a share of the time:
- n
- The number of people who share the machine.
- p
- The share of the working day each person keeps the model busy, as a decimal (5% = 0.05).
- n × p
- How many people are using it at the same moment, on average.
- (1 − p)n
- The chance that nobody is using it at a random moment.
- Physical meaning
- A shared machine is like one kettle for an office: most of the time it sits idle, sometimes one person uses it, and only at the busiest moments does anyone queue. Size for those busy moments, not for the number of names on the list.
“Ten users need ten times the machine.” Only if all ten ask at the same moment all day. Light office users overlap rarely, so one machine that is fast for one person often feels personal-speed to all of them. Heavy, constant users are the exception.
3. Sizing
- 2 to 5 light users, 12B to 27B model: a desktop with a 24 to 32 GB card, or a 48 to 64 GB Mac mini.
- 5 to 20 users, large model: a 128 GB box (Lesson 5.4), or a PC with two graphics cards.
- Many users or heavy agents: a server-class machine and a serving engine such as vLLM (Lesson 15.2).
4. Setting it up safely
- Give the machine a fixed address on the office network.
- Let the server listen on the network (Ollama:
OLLAMA_HOST=0.0.0.0; LM Studio: “Serve on local network”; Lesson 11.1). - Share a chat front end such as Open WebUI with user logins (Lesson 9.5).
- Never open the port to the internet. For remote staff, use the office VPN.
- Decide what is logged, tell staff, and keep logs only as long as needed.
Maths and principle
Principle: size for the busiest few minutes, not the average day.
Equation: with n users each busy a share p of the time, the expected number busy at once is n × p.
Analogy: a single office kettle. Ten people can share it because they rarely all want tea at the same moment, except at 11 o’clock.
Try it: 12 users each busy 5% of the time? (12 × 0.05 = 0.6 busy at once on average: usually zero or one, occasionally two.)
5. Worked example
Made-up numbers. Riverbend Accounting has eight staff. They buy a 64 GB Mac mini M5 Pro, run Ollama with a 27B model and Open WebUI with logins, and keep it on the office network only. Each person waits a second or two more at the 9 a.m. rush, otherwise answers feel personal-speed.
6. Practice
- Estimate n × p for your team.
- Why is a login on the chat front end important even inside the office?
- Which setting lets Ollama accept requests from other office computers?
Answers
- Below 1: one strong machine is plenty. Above 2: plan batching or a second machine.
- It keeps each person’s chats separate and shows who used what.
OLLAMA_HOST=0.0.0.0(inside the office network only).
New words: server, bursty use, batching, VPN (a private tunnel into the office network).
- Office use is bursty: n users each busy a share p of the time means about n × p busy at once.
- Below about 1 busy at once, one machine that is fast for one person serves the whole team.
- Heavy users (agents, coding assistants) need a batching serving engine such as vLLM, or more machines.
- Keep the server on the office network with logins; never open its port to the internet.
Check yourself
Sources
- Ollama server and network settings (OLLAMA_HOST), Ollama documentation (official page), https://docs.ollama.com
- vLLM, a serving engine built for many users, project documentation (official page), https://docs.vllm.ai
- Kwon, W. et al. (2023), Efficient Memory Management for Large Language Model Serving with PagedAttention (the paper behind vLLM), arXiv:2309.06180 https://arxiv.org/abs/2309.06180
Your ticks stay in this browser. Save them as a file on the course home.
Updated 2026-11-22. Prices and model names change fast: check the official source before you buy.