Networking silicon, not GPUs, is the hidden bill in every AI cluster

SANTA CLARA, Calif. — The public conversation about AI hardware still begins and ends with the GPU. The private conversation begins with the switch. Scale-out training jobs die in the fabric long before they die in the math.

Networking vendors said customers who under-bought optics last year are over-correcting this year, pulling in 800G and 1.6T ports on timelines that would have seemed reckless in the last cloud cycle.

That spend does not photograph well. It also does not go away when a cheaper accelerator arrives. Every new chip generation still has to move gradients across a hall.