
Heavy Networking HN850: AWS Takes a Unique Approach to AI Fabrics (Sponsored)
Nov 6, 2026
JR Rivers, AWS Distinguished Engineer and VP of Network Core, and Stephen Callaghan, Senior Principal Networking Engineer, explore the infrastructure powering AI on AWS. They discuss synchronization-bound workloads, training and inference diversity, AWS’s multi-planar RDMA fabric, custom hardware, automated failure recovery, and how EFA and SRD handle massive flows. They also examine UltraClusters, cross-region scaling, and evolving inference networks.
AI Snips
Chapters
Transcript
Episode notes
AI Networks Are Governed By The Slowest Packet
- AI networking is primarily synchronization-bound, because the slowest packet can hold every GPU back.
- AWS therefore targets tail latency while supporting training, post-training, reinforcement learning, and research workloads.
AI Fabrics Must Be Fast Without Becoming Brittle
- AI infrastructure must support more than GPU-to-GPU transfers, including checkpointing, hydration, redundancy, scaling, and upgrades.
- A useful fabric must remain operationally resilient rather than merely fast and low-latency.
AWS Built A Fabric For Continuous Upgrades
- AWS designed its multi-planar spine-leaf fabric around a 10-petabit capacity target under 10 microseconds.
- The network now spans thousands of switches and must be upgraded continuously without a shutdown window.
