A private GPT: what does private actually mean when you buy one?
The phrase is used for at least four different architectures. Only one of them means what most buyers assume.
« Private GPT » is now a common offer to Swiss SMEs, and it is not a defined term. At least four different setups are being sold under that name, and they differ enormously in what actually happens to your text.
The four things it might mean
1. A business account with a large provider. Your prompts go to the provider as before, but under business terms: no training on your content, a data processing agreement, retention limits. This is the most common meaning and it is a real improvement over personal accounts. Your data still leaves Switzerland.
2. Your own application over someone else’s model. A Swiss vendor runs the interface, stores your documents in Switzerland, and calls a large model for the reasoning. Your files sit here; the text of each question travels. Where it travels and under what terms is the question worth asking.
3. An open model on rented infrastructure. An open-weight model running in a Swiss or European data centre. Nothing goes to a US provider. You are trusting the operator of that data centre instead, which may be an easier thing to verify.
4. A model on hardware in your own building. Nothing leaves. This is what most people picture when they hear the phrase, and it is the rarest of the four because it needs hardware, maintenance and a reason.
The question that separates them
One sentence, and any honest vendor answers it immediately: when I type a question, which companies and which countries does the text of that question pass through before I get an answer?
Not « where is the data stored », which is a different and easier question. Documents can sit in Zurich while every question about them is answered in Virginia. Both statements are true at once, and only one of them is usually in the brochure.
Which one you need
Most SMEs are well served by the first or second option, provided the terms are right. Option four is for genuine constraints — professional secrecy, a client contract that forbids third-party processing — not for general unease.
Decide which constraint you actually have, then buy the cheapest architecture that satisfies it. Buying option four to solve a feeling is an expensive way to feel better.