Rethinking Data Center Network Architecture

Dudy Cohen, VP of Product Marketing at DriveNets, examines the rapid evolution of Ethernet technology and its critical role in AI infrastructure. Cohen discusses how Ethernet is advancing through multiple dimensions—from 800G to 1.6T line rates, co-packaged optics integration, and the introduction of scheduling layers that address Ethernet's traditional limitations. He explores the dramatic scaling of GPU clusters, from thousands to potentially millions of units, and explains why different scheduling approaches—fabric scheduling versus endpoint scheduling—suit different deployment scenarios. Cohen traces DriveNets' journey from building scalable service provider routers to applying the same scheduled fabric technology to AI backend networks. The discussion reveals how technical architecture choices directly impact business metrics, from time to first token to cost per million tokens, and why open, Ethernet-based solutions are gaining traction across hyperscalers, neoclouds, and enterprise deployments.What You Will Learn:How Ethernet technology is evolving across speed, optics, and scheduling capabilitiesThe difference between fabric scheduling and endpoint scheduling for AI networksWhy ESUN technology enables Ethernet to compete with purpose-built scale-up protocolsHow GPU cluster sizes are scaling from thousands to potentially millions of unitsThe business case for Ethernet-based AI infrastructure in terms of deployment speed and operational costs