This interview was published on June 2, 2026 in IT-Markt, conducted by Coen Kaat. Translated from German.
Digital sovereignty is forcing IT leaders to rethink. With AI too, companies want to reduce their dependence on large foreign corporations. Romano Roth, Chief of Cybernetic Transformation at Zühlke, explains when a local AI infrastructure pays off and what to consider.
In which scenarios does running AI applications on-premises or in local data centers make more sense than using large international providers?
Romano Roth: Especially in Switzerland, where the revised Data Protection Act (revDSG) and regulators like FINMA set strict requirements, data sovereignty is a central argument for local operation. We see this particularly with financial services providers and in healthcare. At the Children’s Hospital Zurich, for example, the AI platform deliberately runs on-premises because patient data must not leave the building. For industrial real-time applications, on-site operation is also often the better choice.
Is it realistic and sensible to run AI applications for enterprise solutions completely locally?
Technically, it is more realistic today than ever. Open-source models can be operated securely behind the company’s own firewall with platforms like Ollama or Nvidia NIM. However, completely local operation requires in-house expertise for operations, monitoring, and updates. We recommend a risk-based approach: process sensitive data locally, use cloud services for non-critical tasks. Pure ideology in either direction is rarely economical.
What is essential when building a local AI infrastructure?
Scalability: the platform must grow with the requirements without every model becoming an infrastructure project. That requires a modular architecture that provides AI components as self-service. And the people: local AI infrastructure is useless without skilled professionals who operate and evolve it. Many companies underestimate the organizational change towards the Cybernetic Enterprise.
By which criteria should one decide between standard AI models and tailored solutions to maximize ROI?
Clear, measurable goals are decisive: What accuracy does the chatbot need? What latency does the voice bot have? How high must the detection rate in quality control be? Our study with ETH Zurich (633 companies) shows: the biggest business impact emerges when AI is tailored to a company’s own data and processes. Start with standard models, specialize for key applications, continuously measure and adapt.
What opportunities does this create for resellers and system integrators?
Local AI infrastructure needs corresponding ecosystems. At the Children’s Hospital Zurich, for example, Cisco, Netcloud, 44ai, and Zühlke work together, from the hardware to the network to the AI. This is how solutions emerge that no single player can deliver alone. What is needed are partners with domain expertise who do not just install AI but embed it into business processes. Those who demonstrate these competencies become strategic partners instead of interchangeable suppliers.
Which developments in AI infrastructure or data strategy should IT partners keep an eye on now to remain competitive in the coming years?
AI agents are changing the rules of the game. AI no longer just delivers suggestions, it increasingly acts autonomously. In our projects we already see a productivity increase of 30 percent. This requires new infrastructure for orchestration, governance, and monitoring of agent systems. Add to that techniques like RAG and federated learning, where data stays where it originates. Those who remain flexible and avoid vendor lock-in are prepared for this change.
Answers from the other panel participants#
- Daniel Bachofner, Netapp: “Completely local is technically possible, but rarely optimal.”
- Massimo Fumagalli, VAR Group: “Clear governance, compliance, and training are decisive for using AI safely and effectively.”
- Daniel Henneke, HPE: “A hybrid or edge-first solution is often more advantageous.”
- Steffen Märkl, Cloudera: “Private AI is not a question of location alone, but an architecture decision.”
- Christoph Schnidrig, AWS: “Local hardware becomes outdated faster than the ROI is achieved.”
- Roland Stritt, Fast LTA: “Those who run inference locally decide for themselves who gets which answers.”
- Dominik Wotruba, Red Hat: “Completely local AI applications are quite feasible for inference with pre-trained models.”