Skip to content

SRv6 for AI Backend#

OCP Global Summit 2025 digest Program Core pick 13–16 October 2025

Abstract Session recording Materials

Speakers: Changrong Wu (Microsoft), Abhishek Dosi (Microsoft)

  • Why: Using WAN traffic-engineering machinery (SRv6) inside an AI backend fabric is a significant architectural bet — it directly addresses the failure of BGP+ECMP to meet AI training's path-control requirements, and it's Microsoft in production, not a proposal.
  • What: Microsoft's use of Segment Routing over IPv6 in Ethernet-based AI backend networks for fine-grained network path control, maximizing fabric utilization and delivering resiliency — including how SRv6 enables continental-scale GPU clusters. (from abstract only)
  • How: SRv6 segment lists provide explicit per-flow path steering and traffic engineering in the backend, replacing the coarse hashing of BGP+ECMP; a video and the speakers' segment-routing.net materials are linked.
  • Where applicable: Microsoft Azure AI backend production. The SRv6-in-the-DC idea is broadly applicable to any routed Ethernet fabric needing engineered paths, but note the scale assumptions (continental GPU clusters) behind the requirements.