Sovereign AI in Switzerland: where does your data go?
When an employee pastes a contract into an AI assistant, where does that text go? Four possible paths, what Swiss hosts promise, and a decision rule for an SME.

What does “sovereign AI” actually mean?
An AI service is sovereign when you know where your data is processed, which law applies there and who can demand access to it. The location of the servers is only one of those three criteria. It is the nuance most offers gloss over: an international provider can host your data in Switzerland and still answer to another jurisdiction, as Swisscom itself points out in its own material on digital sovereignty.
For a Swiss SME the practical question is simpler. When an employee pastes a contract, an accounting document or a client file into an AI assistant, where does that text go, who hosts it, and what happens to it afterwards? The answer changes completely depending on the model you pick. This article sorts the options by data path, not by brand.
The four paths your data can take
Whatever the tool, your text follows one of four paths. The table compares them on the three useful questions: where, under which law, with how much control.
| Path | Where the text is processed | Law and jurisdiction | Your control |
|---|---|---|---|
| 1. Consumer service or API from a US provider, default region | Most often in the United States | US law, the provider’s terms of use | Low: you depend on its settings and contracts |
| 2. Same provider, European or Swiss region selected | Datacenter in Europe or Switzerland, depending on the offer | The location changes, the company stays American | Medium: good for latency and some contracts, not for jurisdiction |
| 3. Swiss host running open models (Infomaniak, Swisscom, for example) | Datacenters in Switzerland | Swiss law, Swiss company | High on data, quality limited to what open models offer |
| 4. Open model on your own hardware | On your premises | Yours | Maximum, but you run the machine |
This grid avoids a common trap: mistaking path 2 for path 3. Choosing a Swiss region at a US provider settles one part of the problem (location) and leaves the other untouched (the jurisdiction of the company operating the service).
Why jurisdiction matters as much as location
The US CLOUD Act (2018) allows US authorities to require a US provider to hand over data it controls, even when that data is stored outside the United States. This is the mechanism that explains why “hosted in Switzerland” is not enough on its own when the operator is a US company. The US Department of Justice publishes its own overview of the act, linked in the sources.
That does not mean your data will be requested: for most SMEs the risk is low. It means a management mandate, professional secrecy or a confidentiality clause with your own clients may be incompatible with this path, and it is for you to find that out beforehand, not afterwards. This article is not legal advice: for your own case, have it checked by whoever handles your data protection duties. The Swiss data protection framework is covered in our article on what to check before sending data to an LLM.
What Swiss hosts promise, and what to verify
Two Swiss players currently offer language models hosted in the country. Here is what they state publicly, to be read as vendor claims rather than contractual guarantees.
- Infomaniak says its AI services run in its datacenters in Switzerland, that requests are neither stored nor used to train models, and that it relies on open models compatible with the OpenAI API. Billing is usage-based, per token.
- Swisscom presents an AI platform hosted in Switzerland and gives access to the Swiss model Apertus. Its own page on sovereignty insists that storage location is not enough and that the applicable law must also be examined.
Before signing anything, three checks deserve to be made in writing: how long requests are retained (zero, thirty days, more), the commitment not to use them for training, and the list of subcontractors who can access them. A promise on a sales page is not a contract clause.
Apertus: the Swiss model, and its real place
Apertus is an open large language model developed by Swiss universities, including EPFL, and presented in September 2025. Its weights can be downloaded and run on your own servers, which makes it a natural candidate for path 4. Its designers describe it as a reference for transparency, not as a rival to US budgets: press coverage reported performance comparable to Llama 3 from 2024, while other players have released more advanced versions since.
In practice, Apertus meets a requirement for provenance and openness. On its own, it does not meet a requirement for top quality on complex reasoning. If your need is to classify, extract or summarise documents, an open model of this class is often enough. If you expect polished writing or reasoning over long files, test before you decide.
Six questions to ask any AI provider
Whatever path you choose, these questions give you a written, comparable answer. A provider that hesitates to answer is already telling you something.
- In which countries are requests processed, including for backup and support?
- Which legal entity operates the service, and under which law?
- Are requests retained, for how long, and who can read them?
- Are they used to train or improve the models?
- Which subcontractors are involved, and are they listed in the contract?
- What happens at the end of the contract: deletion, timeline, proof?
How to choose: a five-step decision
- Sort your data into three levels. Public (your website text), internal (procedures, notes), sensitive (client files, accounting documents, contracts). The sensitive level is the one that dictates the path.
- List your written constraints. Mandate, professional secrecy, confidentiality clauses from your own clients. If a clause forbids any offshore subcontracting, paths 1 and 2 are ruled out for that data.
- Match a path to each level. Nothing forces you to use a single tool: public data can go to the best online model, sensitive data to a Swiss host or to your own hardware.
- Test on 30 to 50 of your real documents. Measure the error rate with each model. A quality gap that changes nothing for your task does not justify paying more.
- Document the decision. One page is enough: which tool for which type of data, who approved it, on what date. This is the document you will be asked for if a client raises a question.
A worked example: an eight-person fiduciary
Take a fictional but realistic case: an eight-person fiduciary wants 300 files summarised each month, about 20 pages each. That is roughly 6,000 pages, so in the order of a few million tokens a month depending on text density. At that volume, the usage-based API cost of a Swiss host generally stays modest next to the reading time saved: human time dominates the equation, not the AI bill.
The real trade-off is therefore not price, it is the path. A client’s documents are covered by professional secrecy: path 3 or 4 is required, and the question becomes “Swiss host or on-site machine?”. Our article on local AI in your business details the hardware, the hidden costs and the threshold at which an on-site server pays off.
When sovereign AI is not the right choice
In some cases paying for sovereignty is wasted money, and we would rather say so.
- Your data is public or harmless. Rewriting a page of your website or drafting an announcement exposes nothing. A leading online service is then the best choice, and the simplest.
- You need the best reasoning available. On long, complex tasks the best online models keep a lead over open models. If quality comes first and the data is not sensitive, sovereignty costs more than it brings.
- Nobody can run the solution. A poorly updated on-site server is a risk in itself. Without in-house skills, a Swiss host is often safer than a machine forgotten in a utility room.
- The problem lies elsewhere. A sovereign model does not fix a shared password or an employee using a personal account. Start with an inventory of real usage.
Summary: the rule in one sentence
Choose the path according to the most sensitive piece of data that will go through the tool, not according to the model’s brand. Public or internal: the best online service. Sensitive: a Swiss host under contract or your own hardware, after a test on your real documents.
To find which path suits your processes, the five-question express diagnostic helps place your data and your constraints. To see how these choices fit into a complete system, read the method: a foundation, then modules, which includes a local AI option.
Frequently asked questions
What is sovereign AI?
It is an AI service where you know where data is processed, which law applies and who can demand access to it. Server location is only one of those three criteria.
Is a model hosted in Switzerland by a US provider sovereign?
Not necessarily. The location is settled, but the company operating the service remains subject to its home law, including the US CLOUD Act. Check the jurisdiction, not only the datacenter.
Is Apertus as good as ChatGPT?
Not on complex reasoning. Press coverage reported performance comparable to Llama 3 (2024), which is often enough to classify, extract or summarise. Test on your own documents before deciding.
Should I always choose AI hosted in Switzerland?
No. For public or harmless data, a leading online service is simpler and often better. Pick the path according to the most sensitive data that will go through the tool.
What should I check before signing with an AI provider?
In writing: the country of processing, the company operating the service, how long requests are retained, no training on your data, the list of subcontractors and deletion terms at the end of the contract.


