The Fabric and Optics Supply Chain: Scarcity Inside the Cluster
Zoom from the ocean to the rack. Inside an AI training cluster, tens of thousands of accelerators must exchange data constantly, and the switches and optics that wire them together are their own supply chain with their own bottlenecks. Optical transceivers, switch silicon, and the fiber inside the building can gate a cluster as surely as a chip shortage. My read: the intra-cluster fabric is the least-visible and fastest-moving constraint in the stack, because it scales faster than the accelerator count itself, and that superlinear growth is where the next scarcity forms.
Knowledge check
Why does intra-cluster network demand tend to grow faster than the number of accelerators, and why does that matter for the optics supply chain?
Show the answer
Because keeping every accelerator in high-bandwidth communication scales with the number of interconnections, which grows faster than the accelerator count itself, so each increment in cluster size demands a disproportionate increase in switches and optical transceivers, straining the optics supply chain and making fabric components a distinct scarcity point.