Wafer output to 4x on AI interconnect demand — optics is becoming the bottleneck

As a leading indicator, the read is that the front line of AI infrastructure has shifted from single-GPU performance to full-cluster interconnect — and Nvidia's own investment activity confirms it. Vertical integration of the optical fabric supply chain (extending the Mellanox playbook) combined with external investment in optics vendors runs in parallel, accelerating Nvidia's transition from a chip company into an infrastructure-platform company.

The fork is whether (a) this transition reinforces Nvidia's dominance or (b) the Kyber delay and rising competitor stack push the ecosystem toward pluralism. In (a), the optics-vendor investment shows up as utilization gains within several quarters; in (b), hyperscalers explicitly build out alternative paths to reduce Nvidia dependence. Meta's storage rebuild and Google's TPU expansion are already visible as early moves consistent with (b).

Watch three things: (1) the timing of new optical interconnect landing in Blackwell versus Rubin generations, (2) hyperscaler order mix between Nvidia GPUs and in-house silicon (TPU, Trainium, Maia), and (3) downstream effects of the Kyber delay — mentions in Nvidia quarterly guidance and reports of alternative rack adoption. If (2) tilts further toward in-house silicon, the concentration trend around Nvidia's platform will visibly ease.