Open-source models hosted in Switzerland: when they beat a US AI subscription

In brief. An open-source model hosted in Switzerland is preferable to a US AI subscription in four situations: your data is covered by professional or medical secrecy, the task is repetitive and well defined, your volumes make pay-per-use billing expensive, or you do not want to depend on a single supplier. In other cases (varied uses, small volumes, no team to run an infrastructure), a managed service can remain the best choice. The two approaches can also be combined.

This guide is aimed at management and IT leads who have to decide. It contains no prices and no ranking of models: both change too fast, and the only comparison that matters is the one you will make on your own documents.

What we are talking about

A US AI subscription is an online service (ChatGPT, Copilot, Claude and others) or a pay-per-use API. The model belongs to the vendor and runs on its infrastructure. You send your texts and receive an answer.

An open-source model is a model whose weights are published and which you can install yourself: the Llama, Mistral, Qwen or Gemma families are examples. "Open source" covers different licences, some with restrictions on use. They must be read before any deployment.

Hosted in Switzerland means that the model runs on machines located in Switzerland: yours, or those of a Swiss host.

The two worlds are not watertight. Some proprietary models are offered in cloud regions located in Switzerland, and some open models are offered as a managed service. The real question is therefore not "American or Swiss", but: who operates the model, where the computing takes place, and who can access the content.

What "hosted in Switzerland" means in practice

The expression can remain vague. For it to mean something, four points must be set in writing.

1. Where inference runs. On which machines, in which data centre, operated by whom.
2. Where the data is. The documents, but also the search indexes, temporary files, backups and results.
3. Who has access. A named list: internal administrators, provider, host. With which rights, and with what traceability.
4. What is logged. Accesses and decisions must be. The content of documents should be logged only to the extent necessary, with a defined retention period.

One point not to forget: hosting in Switzerland does not by itself answer the question of which law applies to the operator. If the infrastructure belongs to a foreign group, your legal department must assess what that implies. A model run on your own servers or by a Swiss host simplifies this analysis; it does not remove it.

Five criteria for deciding

1. The data

This is the first criterion to examine. Ask three questions.

• Is the content covered by professional secrecy? Art. 321 of the Swiss Criminal Code covers in particular lawyers, notaries, doctors and their auxiliaries. Other secrets exist, for example banking secrecy.
• Is it sensitive personal data within the meaning of the Federal Act on Data Protection (FADP), such as health data (art. 5, let. c)?
• Is there a disclosure abroad? Art. 16 FADP makes it subject to conditions: an adequate level of protection recognised by the Federal Council or, failing that, appropriate safeguards. And art. 9 FADP allows subcontracting only if no legal or contractual duty of confidentiality prohibits it.

This does not mean a US service is forbidden. It means it requires an analysis and contractual safeguards, where a model run in-house requires much less. For a business letter or an internal note without personal data, the question hardly arises.

2. Cost

The two cost models are different by nature.

• A subscription or an API is a variable cost: per user or per volume of text processed. It is low at the start and grows with use.
• A self-hosted model is mostly a fixed cost: hardware or rented computing capacity, set-up, operation. It is higher at the start and varies little with volume.

The tipping point depends on your volumes. A process that handles documents all day, such as file sorting, reaches this point much sooner than an assistant used from time to time. In the calculation, do not forget human time: running a model takes someone.

3. Quality

The large proprietary models remain ahead on open-ended tasks: long reasoning, nuanced drafting, unexpected questions. On narrow tasks (classifying, extracting fields, summarising a short document, checking a list of criteria), medium-sized open models can give sufficient results.

"Can" does not mean "always". The only reliable method is to build a sample of your documents, with the expected answers, and run two or three models on it. Look at languages too: a model comfortable in English may be less so in French, in German or when faced with a mix of the two.

4. Maintenance

With a managed service, the vendor takes care of everything, including changing the model. This is comfortable, and it is also a risk: a model that is withdrawn or modified can change the behaviour of your process without you having decided it.

With a self-hosted model, nothing changes without you. In return, you take on security updates, monitoring, capacity and testing before each version change. For a process that must be auditable, this stability is a real advantage.

5. Team skills

Running a model requires infrastructure skills and a minimum of know-how in evaluation. If you do not have these skills and do not wish to acquire them, three options exist: a managed service, a Swiss hosting arrangement where operation is entrusted to a provider, or a maintenance contract. What matters is knowing who is on call the day the system stops responding.

Comparison table

When open source hosted in Switzerland is the right choice

• Sorting sensitive files, such as minor liability claims at a hospital or sorting KYC files and alerts.
• Processing documents in volume, such as sorting invoices and expense reports at a fiduciary firm.
• Internal assistant on files covered by professional secrecy.
• Process that must remain stable and auditable for several years.

When a managed service remains preferable

• Varied office uses in an environment where your content is already stored, for example Microsoft 365.
• Complex drafting, analysis or reasoning on content of low sensitivity.
• Small volumes, or a need to start within a few days.
• No resources to run an infrastructure, and no wish to entrust its operation to someone else.

The limits to be aware of

• It is not free. The absence of a licence fee does not remove the hardware, the operation or the testing.
• It is not compliance. Hosting in Switzerland simplifies the legal analysis, but it settles neither the information of individuals, nor security, nor access governance.
• The model makes mistakes too. An open model has the same weaknesses as the others: errors, invented answers, sensitivity to wording. Safeguards remain necessary.
• Licences vary. Some limit commercial use or impose conditions. To be checked for each model.
• Hardware has its constraints. Capacity, procurement lead times, power consumption: it is better to size on a real case than on a hunch.

A simple method for deciding

1. Rank your use cases by data sensitivity and volume.
2. For sensitive or high-volume cases, test two or three open models on a sample of your documents.
3. Compare with a proprietary model on the same sample, if your internal rules allow anonymised test data to be submitted to it.
4. Cost both options over three years, operation included.
5. Decide case by case. A mixed architecture is a perfectly sound option.

Frequently asked questions

Is an open-source model worse than a proprietary model? On open-ended tasks, this can be the case. On targeted tasks, the gap narrows and can become negligible. Only a test on your documents can tell.

Do we need to buy servers? Not necessarily. Computing capacity can be rented from a host in Switzerland. Buying is justified for continuous use or for particular control requirements.

Is our data used to train the model? With a self-hosted model, no: the weights are frozen and your texts do not leave your infrastructure. With a managed service, the answer depends on the contract and the settings; it must be checked.

Can we change model later? Yes, if the integration was designed for it. We always separate the business rules, the data and the model, so that the model can be replaced without redoing everything.

Can the two approaches be combined? Yes: a managed service for office work, a model hosted in Switzerland for processes that touch sensitive data.

Going further

Choosing the model is a step in our AI consulting: we test several options on your documents and give you a reasoned recommendation, with its limits. If you prefer a ready-to-use platform, see Swiss GPT. Contact us for a first conversation.

References

• Federal Act on Data Protection (FADP) – Fedlex
• Swiss Criminal Code, art. 321 (professional secrecy) – Fedlex
• AI and data protection – Federal Data Protection and Information Commissioner (FDPIC)