Private AI for your team

What the setup is

• One machine. A PC or a small box in your office, sized below by how many people ask at once. It runs open models such as Qwen3.8 27B and gpt-oss 120B, and everyone shares it, so adding a person adds no fee.
• A chat page for everyone. Each person signs in to a private chat page in the browser, with their own conversations; Open WebUI is the usual one. A server runtime such as vLLM or llama.cpp's server answers several people at once from one copy of the model.
• Answers from your own files. Contracts, manuals and notes are indexed on the same machine, so the assistant can find and quote them. They are read there and stay there.
• Where it sits. In your office, or on a server you rent from a provider that runs its own machines. Where your data stays compares them.

How much machine for how many people

Count the people who will ask at the same moment, not the whole team. The model is loaded once; each conversation running at that moment adds its own context, 0.69 GB for Qwen3.8 27B and 0.31 GB for gpt-oss 120B at 8k of context. That is why a few more people need only a little more memory.

Runs well means the model fits with room to spare and answers at 25 tokens per second or faster for one conversation, the rule of the AI computer picker, which shows every machine's speed. The 14 machines checked are the picker's, laptops left out: a team's machine stays on in the office. Rents are the cheapest single GPU, from a provider that runs its own machines, that runs the same job well, on demand before tax.

Where to buy them

• RTX 4090 PC: Sold mostly used now. The same 24 GB as a used RTX 3090, read a little faster.
• RTX 3090 PC (used): Sold used only. Check the seller, the fans and the warranty before paying.
• Olares One: Olares
• RTX 5090 PC: NVIDIA Marketplace
• Radeon AI PRO R9700 PC: AMD's list of shops
• RTX PRO 6000 workstation: NVIDIA Marketplace
• Ryzen AI Max+ 395 mini PC: GMKtec · GMKtec EU · Framework · Minisforum · Beelink · Corsair

Where your data stays

• In your office. The machine sits on your own network. Prompts, files and answers never leave the building, and the assistant keeps working when the internet is down.
• On a server you rent. A GPU server only your team uses, from a provider that runs its own machines; pick one with servers in your country or region. Your data leaves the office for that machine, so read the provider's logging policy layer by layer. For Qwen3.8 27B and 10 people at once it costs from $0.50 an hour (one RTX A6000 on Hyperstack), about $365 a month if it never stops; for gpt-oss 120B, from $1.80 an hour (one RTX PRO 6000 on Nebius). Prices read 19 October 2026.

Hours decide between them. Put the price you are quoted, your team's hours and your electricity rate into the build vs rent calculator: it returns the month owning becomes cheaper, or says it never does.

Questions

What is private AI for business?

An AI assistant your team uses the way it uses ChatGPT, running on a machine the business controls, so what people type and the files it reads stay with you. Open models such as Qwen3.8 27B and gpt-oss 120B run on one machine in an office.

How many people can share one AI server?

It depends on how many ask at the same moment, not on the size of the team. The model is loaded once, and each conversation running at that moment adds its own context: 0.69 GB for Qwen3.8 27B at 8k. A 32 GB machine runs it well for up to 10 at once, a 96 GB machine for 20.

Is it cheaper to rent a private AI server?

For a few hours a week, usually yes. A server that runs Qwen3.8 27B well for 10 people at once rents from $0.50 an hour (one RTX A6000 on Hyperstack, read 19 October 2026), about $365 a month if it never stops. A machine you own costs only its electricity by the hour once bought: the build vs rent calculator finds the month owning becomes cheaper, with your price, your hours and your electricity rate.

• Memory: the model's weights at 4-bit (Q4_K_M), one context cache of 8k per conversation and the runtime's overhead, the formula of the LLM checker, computed 3 November 2026. At one conversation every figure is the AI computer picker's.
• Machines: the memory and speed on each card's or chip's GPU page. No hardware prices: store pages carry today's.
• Hourly prices: each provider's own price list, read 19 October 2026, as in the cloud GPU price guide.