Decision Guide

On-prem or cloud AI? The Swiss SME decision guide

Three criteria decide: data sensitivity, team size, acquisition budget. Honest, without a tech-sales bias.

The two worlds

Cloud AI: ChatGPT Enterprise, Microsoft Copilot, Claude Enterprise, Google Gemini Enterprise. Quickly introduced, low initial effort. But: your data leaves the operation, often via US servers. License costs per user per month. revDSG compliance only via a DPA contract.

On-prem AI: A local LLM on your own hardware, such as a Lenovo ThinkStation PGX, with an open-source stack (Ollama + Open-WebUI), optionally with an orchestration layer like the Xinity Engine. Higher initial investment. But: your data stays in-house, no per-user licenses, full revDSG compliance structurally instead of contractually.

Criterion 1: Data sensitivity

Data typeRecommendation
Generic office content (mails, standard documents)Cloud
Personal data in regulated industries (fiduciary, finance, healthcare)On-prem (or cloud only with a very strict DPA)
IP-sensitive codeOn-prem
Client correspondenceOn-prem
Marketing content without customer dataCloud

Criterion 2: Team size + usage profile

TeamCloud license / month (example)On-prem (one-time)Break-even
5 users~CHF 300/monthLenovo list price + setupAfter 12–18 months
20 users~CHF 1,200/monthidenticalAfter 6–12 months
50 users~CHF 3,000/monthidenticalAfter 3–6 months

Note: cloud prices are illustrative based on typical ChatGPT Enterprise license costs. Current prices in the architecture conversation.

Criterion 3: Compliance and governance requirements

  • Audit trail obligation? → On-prem structurally simpler (or cloud + Xinity-style orchestration)
  • EU AI Act readiness? → On-prem or an EU-sovereign stack
  • Internal company rule "data does not leave the house"? → On-prem, period
  • No special requirements? → Cloud is usually enough

Hybrid architectures are the rule

What is written above sounds like an "either-or". In practice, hybrid setups are the norm: office workflows in cloud AI, sensitive workflows on-prem. We think one architecture per use case, not per company.

When Xinity comes into play

You want not only Swiss hardware but also an EU-sovereign software stack? Then we integrate the Xinity Engine. A Vienna-based open-source orchestration layer (Apache 2.0) with an OpenAI-compatible API, model routing, and audit trails. Recommended for multi-team setups with elevated compliance requirements.

The Swiss special case: revDSG

The revised Swiss Federal Act on Data Protection (revDSG), in force since 1 September 2023, requires a data processing agreement (DPA) and information to the data subjects when using cloud AI with US providers. On-prem avoids this topic structurally and completely. The data does not leave the operation.

Frequently asked questions

Is cloud AI illegal in Switzerland?

No. With a DPA contract, information to the data subjects, and a risk assessment, cloud AI is usable in a revDSG-compliant way for many use cases. It is not "illegal".

Then why go on-prem at all?

For sensitive data, IP, regulated industries, multi-team setups, or an internal rule that "data stays in-house", on-prem becomes structurally more secure, and above a certain team size more economical.

How fast does cloud AI run vs on-prem?

Cloud setup: 3–8 days. On-prem with a ThinkStation: 2–6 weeks including hardware delivery. The time factor favours cloud, the sensitivity factor favours on-prem.

Can we switch later?

Yes, in both directions. Cloud → on-prem is more common, because data sensitivity typically grows, not shrinks.

What is Xinity?

A Vienna-based open-source orchestration layer for on-prem LLMs. OpenAI-compatible, Apache 2.0. We integrate it on request.

What do you typically recommend?

Hybrid: cloud for standard office use cases, on-prem for sensitive ones. The question is not "on-prem or cloud", but "which use cases belong where".