What NFD40 Taught Me About AI Networking
Router Jockey, Tuesday, July 7th, 2026
Networking Field Day 40 takeaways on why AI cluster networking differs fundamentally from cloud architectures.
Tony Mattke's NFD40 wrap-up argues AI networking is distinct from traditional cloud networking.
AI clusters rely on collective operations where every GPU communicates at once, so statistical load balancing fails and jitter tolerance is near zero; standard ECMP can stall entire collectives. A 100,000-GPU cluster needs 125,000 server ports and more switch racks than compute racks.
He separates scale-up, scale-out, and scale-across as distinct problems needing different silicon and protocols, and notes differentiation has shifted from hardware to control planes and observability.