Europe’s public AI compute, spread across EuroHPC supercomputers and national AI factories, can launch a frontier‑class model sooner than a new gigawatt data center. The study shows federated low‑communication training could deliver a model by 2028, versus 2033 for a single campus. The analysis rests on exaflop totals already in place for European researchers today.

Model authors built a three‑layer framework: layer one estimates DiLoCo training efficiency, layer two tracks site activation and cumulative compute, and layer three scores regions on time, cost, and carbon. The headline result hinges on layer two; if federated sites energize before a gigawatt campus, the federation wins for European AI capabilities and public infrastructure efficiency today.

Data sources include the AWS “up to seven years” grid‑queue claim and IEA 2‑to‑10‑year ranges, tied to seven European regions. EuroHPC flagship counts and AI factory accelerators supply the compute inventory, while sensitivity tests confirm that communication penalties are secondary to activation timing. The report stresses political decisions shape usable capacity for future deployment efforts.

Authors note that no European operator has yet energized a 1‑GW point load, and that distributed training beyond 10 B parameters remains unproven. The model’s predictions rely on central estimates rather than observed data, and the addressable fraction of shared hardware depends on policy choices. The report is not peer‑reviewed, yet it offers a clear roadmap.