Pool first, fab later

A region without a frontier fab has a move left: treat the chips it already has as one resource instead of many. Europe's path to training the largest models runs through that move, and the bottleneck worth solving is not silicon. It is wiring.

RAND Europe's newest report calls the mechanism by name. Distributed training, the family of techniques that let geographically separated datacentres cooperate on a single training run, is a relief valve on Europe's two soft bottlenecks: power, where a single site cannot host a frontier build, and politics, where cross-border rules fragment the existing stock. The chip gap with the US and China is named in the same breath and left open. Distributed training is not a chip.

The portable lesson: when a region cannot add a fab, it can pool what it has. Harmonize the cross-border rules, prepare the physical sites, and let fragmented clusters behave as one training resource. The constraint migrates from silicon to coordination. Europe is betting that coordination is easier to legislate than chips are to fabricate. The chip gap stays where it was, but the training option does not depend on closing it.