What does it take to run large language models on your own infrastructure? And why does GPU memory capacity sometimes matter more than raw computing power? Lukáš Hejtmánek and Ivana Křenková from CERIT-SC at Masaryk University attended NVIDIA AI Days 2026, where they shared insights into developing and operating AI services for the Czech research community.
The tenth edition of NVIDIA AI Days took place on 5 November 2026 in Prague as part of the nationwide AI Days festival (Dny AI). This year's programme focused primarily on AI infrastructure, modern data centres and practical experience with deploying artificial intelligence.
In his presentation, Scientific AI Cloud: Lessons from Real-World Operations, Lukáš Hejtmánek introduced the technological infrastructure of CERIT-SC and e-INFRA CZ, which provides access to powerful open-weight language models through both chat and API interfaces. The service currently has 4,300 chat users and 2,800 API users.
However, the presentation went beyond showcasing computing power. It focused on lessons learned from day-to-day operations, including efficient GPU utilisation, the importance of GPU memory and caching, computational optimisation, and the technical challenges and limitations of current AI tools.
Sharing these practical insights is an important contribution by CERIT-SC to the wider technology community. It demonstrates that running AI models on dedicated infrastructure is not simply about investing in powerful hardware, but also about developing sustainable and reliable solutions capable of supporting thousands of users in their everyday work.
Romana Ďuračiová presented Foldify at Dell Technologies Forum, a web application she began developing during her bachelor's studies at the Institute of Computer Science at Masaryk University.
At Brno.AI Meetup #11, Terézia Slanináková from CERIT-SC at Masaryk University presented the AI inference services operated within e-INFRA CZ, demonstrating that these are not just emerging technologies but services already used by thousands of researchers and other users.
