What does AI data sovereignty mean for Swiss businesses?

What AI data sovereignty is – and is not

Data sovereignty is an organisation's lasting control over its data: who can access it, where it resides and which law applies. AI raises the stakes, because prompts, documents, customer data and sometimes whole knowledge bases are handed to a model – often a cloud service located abroad.

It matters to separate related terms. Data residency only fixes the geographic storage location. Data localisation requires that data never leave a country. Sovereignty goes further: it also covers legal reach – for instance whether a foreign authority can compel disclosure even when the server sits in Zurich.

• Data residency: where does the data physically sit?
• Data localisation: may the data leave the country at all?
• Data sovereignty: who holds legal and technical control and access?

Why data sovereignty matters for Swiss businesses

• revFADP/nFADP: the revised Data Protection Act requires adequate protection when data is disclosed abroad; transfers to states without an adequate level of protection need additional safeguards.
• Professional and official secrecy: doctors, lawyers and fiduciaries (Art. 321 Swiss Criminal Code) and banks are bound by special confidentiality duties – an uncontrolled AI data leak can be a criminal offence.
• US CLOUD Act: US providers can under certain conditions be compelled to hand over data even when it physically resides in Europe.
• EU AI Act (extraterritorial): also applies to Swiss providers whose AI outputs are used in the EU, adding transparency and risk-management duties.
• Trust and competition: customers, hospitals and public bodies increasingly demand provable data hosting in Switzerland or the EU – sovereignty becomes a selling point.
• FDPIC (EDÖB): the supervisory authority expects documented processing and risk assessments, such as a record of processing and, for high risk, a data protection impact assessment.

Your options at a glance

There is no universally 'right' model – the fit depends on data class, budget and in-house know-how. These five approaches span the spectrum from full outsourcing to complete self-control.

• Swiss cloud / Swiss hosting: AI services from providers with Swiss data centres (e.g. Infomaniak, Exoscale, Swisscom). Upside: data and law stay in Switzerland. Check: which models are available and whether sub-processors abroad are involved.
• Sovereign/EU cloud: offerings with guaranteed EU data residency and partial shielding from third-country access. A good compromise between performance and control.
• Hyperscaler with a Swiss region: AWS, Azure and Google Cloud run Zurich regions. Data sits in Switzerland, but residual CLOUD Act risk and group structure still need weighing.
• Local/on-premise models: open-source LLMs in your own data centre or on your own hardware. Maximum control, no data leaving the house – at the cost of operations, hardware and maintenance.
• Hybrid: sensitive data local, non-critical tasks in the cloud. Often the most pragmatic route in practice.

Local and open-source models as a sovereign foundation

For maximum sovereignty, run models yourself. Open language models can run on your own infrastructure so that no data leaves the building. Apertus adds a fully open language model developed in Switzerland by ETH Zurich, EPFL and the CSCS supercomputing centre, offering transparency over training data and weights.

The price is effort: GPU hardware or a sovereign Swiss compute cluster, plus expertise for operations, fine-tuning and security. For many SMEs a hosted open-source model at a Swiss provider is the better middle path – open weights, but without running your own servers.

A practical decision framework

• Classify data: which data is especially sensitive (health, finance, secrecy-bound professions)?
• Clarify the legal frame: revFADP, professional secrecy, EU AI Act and sector rules (FINMA, healthcare).
• Assess the access model: not just storage location, but who can legally compel access.
• Vet the provider: data-centre location, group headquarters, sub-processors, contracts and certifications.
• Shape contracts: data-processing agreement, standard contractual clauses, deletion concepts and no use of your data for model training.
• Protect technically: encryption, pseudonymisation, confidential computing and access controls.
• Document: keep a record of processing and, for high risk, run a data protection impact assessment.

Common misconceptions

• 'Server in Switzerland = sovereign': not necessarily – the provider's group headquarters and applicable law count too.
• 'Open source solves everything': only if you run it yourself or at a sovereign provider; an open model via a foreign API still leaves you dependent.
• 'revFADP = GDPR': similar but distinct; Switzerland is governed by the revFADP/nFADP, not the GDPR.
• 'Small firms are not affected': professional secrecy and customer expectations apply regardless of size.