Instead of relying on commercial AI services, EPFL has built its own open-source AI platform, running entirely on campus infrastructure. One year after its launch, it is helping researchers, teachers and staff use powerful AI tools securely, efficiently and while keeping control of their data.For the past year, EPFL has implemented an in-house alternative to commercial AI clouds: an open-source, on-premise AI inference platform built and operated by the Research Computing Platform (RCP), which will soon merge with SCITAS to form EPFL's Integrated Computing Platform (ICP). Since going live in August 2025, the service has become a quiet success story, giving EPFL collaborators a sustainable, efficient, and sovereign way to run AI workloads without depending on external providers.The idea is simple. Instead of every lab, course, or administrative team creating its own AI infrastructure or sending data to third-party clouds generating important costs, RCP provides a shared, Kubernetes-based platform that runs entirely on EPFL's own hardware that is already being used for the research community. Users interact with the inference service through standard, OpenAI-compatible APIs, while the platform handles the heavy lifting: dynamic scaling, secure multi-tenant operation, and enterprise-grade reliability.A Shared AI Service for the EPFL CommunityA year in, adoption speaks for itself. A broad community of internal service providers now relies on the platform for production use, spanning research, education, and administration alike, from large language model experimentation and clinical studies to AI teaching assistants and internal chatbots (https://chat.rcp.epfl.ch/). In addition to this chat interface, users can also request their own API key via the RCP Portal (https://portal.rcp.epfl.ch/) and integrate it into their preferred workflow tools such as coding agents and assistants. User feedback has been consistently positive, and continues to drive improvements to the service.Open Models, Sovereign Infrastructure, Broader ImpactAs a public academic institution, EPFL is committed to openness while maintaining control over its data, infrastructure, and costs, in the most sustainable and responsible way possible. That's reflected in the platform's open access to leading open-weight models with a catalog now spanning more than 130 models across language, vision, embedding, reranking, and speech-to-text, including the Mistral, Qwen, and Llama model families. Naturally, it also includes the latest version of Apertus, the fully open large language model developed by EPFL, ETH Zurich, and CSCS as part of the Swiss AI Initiative.This growing ecosystem has recently expanded beyond EPFL. The AI Inference platform is now officially integrated as an inference provider within the Swiss AI Research Platform operated by the Swiss National Supercomputing Centre (CSCS). This enables users of the Swiss AI Research Platform to seamlessly access Apertus through infrastructure hosted and operated by RCP (more info). This bridge between the two platforms is the result of a close collaboration between EPFL and CSCS teams, marking an important milestone towards a more interconnected sovereign AI infrastructure for Switzerland.The service has also been able to provide the necessary infrastructure for initiatives such as red teaming with UNICC; and in outreach, the expertise developed in-house is being used to advise industry partners to set up their own sovereign inference solutions, increasing EPFL’s impact on society.By sharing infrastructure across research, education, and administrative needs, and bringing together internal service providers across different sectors, EPFL has built something increasingly rare: a broad, vibrant community around a shared, transparent, and sovereign foundation for AI that keeps competencies, data, and control in-house, at scale, for everyone on campus.