Post by AI World

To build a frontier AI model, Europe doesn't need one giant data centre. It can train across the computers it already has, working together. Five recent papers point the same way. The IPCEI federated-infrastructure preview treats compute concentration in a few hyperscalers as the bottleneck, but any model at that scale crosses the AI Act's systemic-risk line, where the obligations are documentation, evaluation and traceability, not performance. The pressure to keep data at origin turns federated learning from a privacy and computation trick into a structural requirement. But federation is not free. Once training is cooperative, clients and the central server share legal responsibility, and auditability becomes a design parameter rather than a feature. Meanwhile the technical barrier is falling: DiLoCo trains across poorly connected devices while communicating around 500 times less than standard training, and keeps going as nodes drop out and rejoin. Work by Cornelia Kutterer, LL.M. Herbert Woisetschläger, Alexander Erben, Bill Marino, Shiqiang Wang, Nicholas Lane, Ruben Mayer, Simon Mertel, Christoph Krönke Lieb, Arthur Douillard, Qixuan FENG, Rachita Chhaparia, Yani Donchev, Adhiguna Kuncoro, Marc'Aurelio Ranzato and Jiajun Shen.