The Next Frontier AI Challenge has a clear objective: Europe should not simply replicate existing AI systems but develop new technological paradigms. To qualify, teams had to present an approach that promises an improvement of at least one order of magnitude in at least one essential KPI – such as data efficiency, energy consumption or the ability to learn continuously.

The selected teams are pursuing a wide range of approaches. These include Physical AI, from World Models for robotics and Embodied AI to algorithms that can invent industrial hardware. Other teams are developing AI systems for time series that operate directly on live streams of measurements and signals from real-world systems, or pursuing continual learning with systems that adapt their World Models in real time with every interaction. Further teams are working on automating scientific discovery and improving the efficiency of large language models through foundation models trained on the weights of other neural networks rather than on text or images.

The finalists include spinouts from renowned European universities such as Oxford, ETH Zurich, the University of St. Gallen and the University of Tübingen. Several founders previously built AI at companies such as Google DeepMind, OpenAI and Hugging Face – and have now chosen to continue developing their technologies in Europe.