You decided your company's documents should stay in the building, so you want a local AI. Now the buyer's question: how big a machine for 5 people, for 15, for 50? Staff count alone does not answer it.
Here is a three-step walk-through, with one worked example per team size. Memory sets the limit we can calculate, so we start there.
Step 1: pick the model class
The model decides the base memory, because everyone shares one copy of it.
- 27 to 31B models (for example Qwen3.8 27B) handle document questions, drafting, summaries and translation well. At 4-bit they need about 20 GB, at 8-bit about 32 GB.
- 120B-class mixture-of-experts models (MoE: a large model that uses only a few billion weights per word, so it stays fast) such as gpt-oss-120b suit stronger reasoning and longer, trickier tasks. At 4-bit they need about 67 GB.
If you have not decided, start with the smaller class and test it on your own documents. The model finder at ai.html lists more. The memory formula behind these figures is in our guide to GPU memory for local LLMs.
Step 2: count concurrent users, not staff
Fifty employees do not send fifty questions at the same second. Only some of them are in a conversation at any moment, and the rest are reading, working or in a meeting. So count the people who would be typing in your busiest hour. That is the number to size for. Ask each team lead, then add a margin.
Step 3: users times context equals KV cache
Each open conversation keeps a KV cache, the model's working memory, and it grows with the text. A chat with a 32,000-token context (roughly 24,000 words, a long report) takes much more than a chat with an 8,000-token context (a few pages). So the same card serves many short chats or fewer long ones. The selector counts how many conversations fit in memory; speed is checked separately.
Conversations that fit in memory, Qwen3.8 27B 4-bit, by context length
Estimate. Shorter context fits many more chats. A memory limit only; speed depends on the serving software.
Source: EMD Lab AI selector estimate, calculated from model files and configuration files. Not a measurement.
Example 1: a team of 5
Five people in a busy hour is a light load. The Studio preset (an RTX PRO 4500 with 32 GB) runs Qwen3.8 27B at 4-bit, with an estimated 35 tok/s for one user (tokens per second, roughly words per second). Memory holds 5 conversations at 32K, or about 17 at 8K. That covers five people who ask in turns, as long as most documents are short. Our memory guide shows the RTX 5090 option for the same job.
A 128 GB AI box is the other path, if you want a 120B model for a few people. It suits one to five people, but a box shares one slower memory bus, so it reads long prompts slowly. It is an "Ask us" item on our site, so tell us the job.
Example 2: a team of 15
Fifteen people usually means a department, and the busy hour might put a handful on the machine at once. Use one RTX PRO 6000 Max-Q (96 GB). Qwen3.8 27B at 4-bit fits about 33 conversations at 32K (109 at 8K), at an estimated 63 tok/s. Switch to gpt-oss-120b and it fits about 24 at 32K (96 at 8K), at an estimated 176 tok/s.
If your team also lives in shared files, look at the QNAP AI server. It puts the same 96 GB card in a network storage box with twelve SSD bays, so documents and the AI that reads them sit in one place. It is the storage-first option.
Example 3: a team of 50
For fifty people, use two RTX PRO 6000 Max-Q cards, 192 GB in one tower: the AI preset. The selector estimates room for about 102 conversations of gpt-oss-120b at 32K and 405 at 8K, or 74 and 246 for Qwen3.8 27B. That leaves headroom for the busy hour and for people who keep long chats open.
If the plan includes frontier-class models, such as the very large mixture-of-experts models that fit only around 2-bit, go to four cards, 384 GB, the Team preset. Our configurator estimates about 1,880 W at the wall for this build, so it needs its own circuit. Check your PLN connection and the breaker before you order.
Everything side by side
| Team | Machine | Model | Chats at 32K / 8K | Speed |
|---|---|---|---|---|
| 5 | Studio, RTX PRO 4500 | Qwen3.8 27B | 5 / 17 | about 35 tok/s |
| 15 | 1 × RTX PRO 6000 | Qwen3.8 27B | 33 / 109 | about 63 tok/s |
| 15 | 1 × RTX PRO 6000 | gpt-oss-120b | 24 / 96 | about 176 tok/s |
| 50 | AI, 2 × RTX PRO 6000 | gpt-oss-120b | 102 / 405 | about 176 tok/s |
What these numbers do not tell you
They are memory limits. They say how many conversations can stay open, not how fast the answers feel when many people press Enter together. That also depends on the serving software, on how long the prompts are and on your habits. So treat the table as the size of the job and let us size the final machine with you.
The short version
- Pick the model class first: 27 to 31B for documents and drafting, a 120B-class model for stronger reasoning.
- Size for the busy hour, not for headcount. Users times context equals the KV cache.
- 5 people: one 32 GB card. 15 people: one 96 GB card. 50 people: two 96 GB cards, and four for the largest models.
- Shorter context means many more conversations. Plan the power circuit early for the four-card build.