Show HN: Could Europe train a frontier AI model using the computing resources it already has?

An analysis has been released showing that if Europe's existing public computing resources—including its EuroHPC supercomputers and 19 AI Factories—were connected via DiLoCo (Decentralized Low-Communication) federated learning, it could secure a frontier-scale AI model much faster than building new large-scale data centers.

Key findings of the analysis are as follows:

• Time efficiency: While a new 1GW-class data center must wait an average of 7.6 years for grid connection, leveraging existing resources through federated learning could enable frontier-scale model training by around 2028 (compared to an estimated 2033 for a new data center).
• Model structure: The analysis used a three-stage model accounting for efficiency, availability time, and regional feasibility, finding that time-to-availability of resources has a more decisive impact on outcomes than the training efficiency penalty.
• Limitations: The analysis notes that its estimates are based on grid-connection wait times, and that existing EuroHPC resources have batch scheduling and heterogeneous hardware configurations, meaning that unifying them into a single training job would require political/operational decisions.