Should your office rent AI from the cloud, or run a model on a machine in your own server room? The honest answer is that it depends, and for some offices the answer is "both". This post gives you a checklist so you can decide with facts, not with a sales pitch.
The honest case for cloud
We sell AI machines, and we still start here. The biggest cloud models are stronger than any model you can run on a single box, especially on hard reasoning, long multi-step analysis and tricky code. If your work needs the very best model today, cloud wins.
Cloud is also easy to start. You sign up, pay for what you use and nobody installs anything. For light use, a few people and occasional questions, the cost per answer is often lower than owning hardware. If that is your situation, use cloud and be happy.
One caution. Before staff paste company files into any cloud service, read the provider's terms: what happens to your inputs, how long they are kept, and whether they can be used for training. Plans differ, so check the plan you actually use.
Why offices still buy their own box
People who run a local LLM (a language model on their own hardware) rarely do it because it is cheaper per answer. They do it for these reasons:
- Data stays inside your network. When the model runs on your own machine, customer records, contracts and internal documents do not have to leave the building to get an answer. Indonesia's Personal Data Protection law (Law No. 27 of 2022, UU PDP) is part of the context here, and your own legal adviser can tell you what it means for your case. A box does not make you compliant by itself.
- A fixed cost with no fee per user. Ten people or fifty use the same machine. Nobody rations questions to save budget.
- It works offline. An outage at the provider, a slow link on the factory floor or a branch with weak internet does not stop it.
- Many staff share one machine. Several people can chat at the same time, each in their own conversation.
- Long documents and image tools run locally. Reading a 300-page contract or generating images does not depend on upload speed.
The checklist
Answer the questions below for your own office. Count how many point to "own box".
| Question | If yes | Points to |
|---|---|---|
| Will more than a handful of people use AI every day? | A fixed cost beats a fee per user | Own box |
| Are staff already pasting company files into public chatbots? | Give them a private option they can use in the same way | Own box |
| Are your documents long (contracts, tenders, manuals)? | Long context runs locally, with memory sized for it | Own box |
| Do you make images or video in-house? | Image and video tools run on the same GPU | Own box |
| Do you need it to work offline or on a weak link? | Only a local model keeps working | Own box |
| Do you need the strongest reasoning for hard problems? | The biggest models live in the cloud | Cloud |
| Is use light and occasional, with few people? | Pay for what you use | Cloud |
| Is there nobody to run a server? | EMD Lab installs the AI software before delivery, and staff use it in a browser | Either |
The last row matters. A box does not need an IT department. We install the model and a chat page before delivery, so your staff open it in a web browser like any website.
Hybrid is what many companies end up with
You do not have to pick one. A common setup: use the cloud for public work, such as marketing drafts, research on public topics and brainstorming. Use the local model for anything confidential, such as customer data, contracts and internal reports. Each tool does the job it is best at.
What one box actually serves
"Own box" is not one size. Two numbers decide how it feels: tok/s (tokens per second, roughly words per second) for one person, and how many conversations fit in the card's memory at the same time. The table shows estimates from the EMD Lab AI selector for two models at a 32K-token context, about 24,000 words of conversation and documents.
| Model and card | Memory needed | Speed | Conversations that fit |
|---|---|---|---|
| Qwen3.8 27B on RTX PRO 4500 (32 GB) | 21.5 GB | about 35 tok/s | 5 |
| Qwen3.8 27B on RTX 5090 (32 GB) | 21.5 GB | about 63 tok/s | 5 |
| Qwen3.8 27B on RTX PRO 6000 (96 GB) | 21.5 GB | about 63 tok/s | 33 |
| gpt-oss-120b on RTX PRO 6000 (96 GB) | 67.1 GB | about 176 tok/s | 24 |
Read it like this. A 32 GB card runs a 27B dense model at Q4 (a 4-bit quant, the usual compressed file) and serves a small group. The 96 GB card holds the same model with room for about 33 conversations, or a much larger model, gpt-oss-120b, with room for about 24. gpt-oss-120b does not fit in 32 GB of VRAM: part of it gets offloaded to system RAM and speed drops to about 20 tok/s. These are calculated estimates for a typical setup, not a promise for your documents.
Which box for which office
A small team or one department head: Studio
The Studio workstation comes with a 32 GB card (RTX PRO 4500 by default, RTX 5090 as an option). It runs 27B-class models for a small group and for drafting, summaries and document questions. It will not run the largest open models at full speed.
Storage-first offices: the QNAP AI server
If your files already live on a NAS and you want AI that works on them, the QNAP AI server puts storage and an RTX PRO card in one chassis. Staff open a browser page and ask questions. It is a good fit when the question is "our files, but smarter".
A department or a bigger team: the AI build
The AI workstation has two RTX PRO 6000 Max-Q cards for 192 GB of GPU memory. Use it when many people share a large model, or when you want long documents and image tools alongside chat.
Start from your own questions
Write down 20 real questions from your own work, and note how long the documents behind them are. Send them with your quote request: they tell us which model and how much memory the job needs, and once the box runs, you can compare its answers with your cloud tool. That list tells you more than any spec sheet.
The short version
- Cloud is stronger on hard reasoning and is often cheaper per answer. Start there if your use is light.
- Buy a box for control: private files, a fixed cost with no fee per user, offline use, many users at once, long documents and image tools.
- Many offices run both: cloud for public work, the local model for confidential work.
- Check your cloud provider's terms before staff paste company files into it, and ask your legal adviser about UU PDP.
- Size the box by memory and number of users. A 32 GB card suits a small group, a 96 GB card a department.
The speeds and conversation counts above are estimates. Find the model that suits your work in the model finder.