Data Sovereignty Is Not a Quality Promise: The Apertus Reality Check

What the Apertus reality check actually shows

Proton is integrating Apertus 1.5 into Lumo. At the same time, an SRF reality check shows that the open Swiss model is barely used in production a year after launch. The parliamentary AI 'Pia' relies on models from China, the US and France—not Apertus—because it adheres less strictly to source documents. That is no reason to dismiss Swiss sovereignty, but it is a reason to look more closely. The ki-podcast positions Apertus as Europe's answer to US and Chinese AI giants.

Data sovereignty is an architecture property—not a model label

Digital sovereignty is real. When the US government forced Anthropic to withdraw Fable 5 in June 2026, it showed how quickly a kill switch can disrupt operations. Europe depends on US AI models by more than 90 percent. But sovereignty does not come from a model being Swiss. It comes from where your data flows, who controls access in an emergency, and whether you can switch to another model. The Fable 5 shutdown shows how dependent Swiss and European companies are on US AI.

The market and our customers expect a finished product from AI models. Apertus is currently definitely still a raw diamond that needs to be polished.

Evaluate two axes separately

When choosing an AI provider or model, separate two questions: How sovereign is my architecture? And how reliable is the model on my documents? The Swiss flag does not automatically answer either. A Swiss-hosted US model can ensure data sovereignty, while a 'Swiss' model can run on US infrastructure.

• Where are data processed and stored—including for support and maintenance?
• Is training on your data contractually excluded?
• What is the hallucination rate on your own documents?
• How good is language fidelity for German and Swiss German?
• Is the model production-ready or just a demo?
• Is there a multi-model fallback if a provider is shut down?
• Who is liable in an emergency—your provider or you?

Apertus in perspective: underdog with real value

Apertus was never designed as a direct ChatGPT competitor, but as a knowledge and participation project. The first version was built with two full-time positions and a few dozen researchers, funded with 20 million Swiss francs over four years—US providers employ thousands and invest billions. It was trained on the Alps supercomputer in Lugano with around 15 trillion tokens and over 1000 languages. Yet Apertus partly runs on Amazon or Microsoft servers and was trained on Nvidia chips.

We are the absolute underdog here.

What this means for your AI choice

If you are choosing an AI partner now, do not start with the model. Start with your architecture: data sovereignty, failure scenario, liability. Then test concrete models on your own documents. A structured evaluation framework helps—one designed for Swiss SMEs covers exactly these criteria.