Sovereign AI in Switzerland
Some data cannot leave the building: patient records, court files, client mandates, deal rooms. Those workflows still get AI, on infrastructure where the models come to the data, under your control, with every output traceable.
What sovereign actually means
Sovereign is not a marketing label for "EU region of a US API". It means models you host, on-prem or in your private cloud; weights you can swap without a contract negotiation; logs you own; and provenance on every output: which model, which prompt version, which sources, when.
That definition matters because the regulator, the data-protection officer and the works council will each test a weaker one. A setup that survives all three is an architecture decision, not a procurement checkbox.
Custom models where APIs cannot see the data
Where even a private API call is off the table, small fine-tuned models do the work locally: named-entity recognition, PII masking, classification and drafting on your terminology and your edge cases. Reference point: 97%+ accuracy on PII detection in Swiss clinical text, with zero data persistence.
The rest of the workflow can still use frontier models through routing: sensitive steps run local, everything else runs where quality per franc is best. Sovereignty where required, not everywhere by reflex.
Shipped where it is hardest
Switzerland's first production clinical LLM platform went live in Swiss hospitals in under 90 days, on-prem, FADP and GDPR compliant. DocIQ Shield anonymises court documents for the Swiss judiciary with fine-tuned local models and zero data persistence.
The same architecture patterns carry to banks, insurers, public administration and any organisation bound by the Swiss Berufsgeheimnis or the EU AI Act.
Where to start
A 2-3 week process audit maps your workflows and their data constraints, and tells you which steps genuinely need sovereign infrastructure and which do not. Fixed fee, CHF 15-25k. A sovereign system build then ships in 4-10 weeks, with your team owning operations at handover.
Common questions
Do we need to buy our own GPUs?
Not necessarily. On-prem hardware, a private cloud tenancy and a Swiss-hosted VPC are all valid sovereign deployments; GPU sizing for the actual workload is part of the build. Buying hardware before the workload is understood is the most common way to overspend.
Are local models good enough?
For scoped tasks, yes, and often better: a small model fine-tuned on your entities and terminology beats a general-purpose API on the domain, and it never sees third-party infrastructure. Broad reasoning tasks route to larger models where the data classification allows it.
Which regulations does this address?
FADP, GDPR, the EU AI Act and the Swiss Berufsgeheimnis are the usual drivers, plus sector rules in healthcare, legal and public procurement. Audit trails and provenance are built into the architecture, so the compliance review works from evidence, not assurances.
What does sovereign AI cost?
The process audit is CHF 15-25k fixed. Sovereign builds sit at the upper end of the CHF 40-150k system-build range, driven by hosting choice and model work. Quoted firmly after the audit, not before.
Map the process before you buy the tool.
One call to agree whether a process audit, a system build or a fractional AI lead fits where you are.