The quiet reshaping of the AI chip supply chain

For the past two years, the bottleneck in artificial intelligence has quietly shifted. It used to be about who had the cleverest model architecture. Increasingly, it's about who can actually get chips manufactured, and in what volume.

Fabrication capacity is booked years in advance, and expanding it means multi-billion-dollar plants that take the better part of a decade to come online. That mismatch, model progress measured in months against manufacturing capacity measured in years, is reshaping which companies can compete regardless of how good their research is.

This piece traces where that capacity is concentrated, who's investing to diversify it, and why the next constraint on AI progress may have less to do with algorithms and more to do with concrete, silicon and shipping lanes.