Highlights of SRv6 at OCP Global Summit, October 2026
The OCP Global Summit 2026 (opencompute.org), 12–15 October 2026 in San Jose, put SRv6 at the center of how hyperscalers build AI scale-out, scale-across, and backbone networks. Meta, OpenAI, Cisco, Alibaba, and Google shared production deployments and open-ecosystem work spanning FBOSS/SAI, SONiC, and host-driven source routing.
This page highlights the SRv6 sessions from the summit.
Highlighted presentations:
OpenAI and Cisco: SRv6 Deployments in Scale-Out and Scale-Across Networks
Guohan Lu, OpenAI, and Clarence Filsfils, Cisco Fellow
Guohan Lu and Clarence Filsfils show how SRv6 uSID replaces ECMP in AI backends: the NIC encodes a deterministic path and switches forward strictly on that list. This is deployed in production at OpenAI. MRC over static SRv6 source routing is used to train frontier models, so jobs ride out network failures without waiting for control-plane convergence—including a 75K-GPU production incident where a T1 stopped forwarding, MRC removed the path, and throughput recovered. The session also covers SONiC uN/uA support and scale-up requirements for native strict source routing.
Alibaba and Cisco: SRv6 Traffic Engineering in AI DCI/WAN
Yuanchao Su, Alibaba Cloud, and Clarence Filsfils, Cisco Fellow
AI workloads now span clusters over a regional network, where capacity is scarce and latency matters. Yuanchao Su and Clarence Filsfils present SRv6 as the end-to-end TE fabric from DC through DCI/WAN, reusing the mature SONiC SRv6 stack on DC-class hardware. The talk compares host-based and aggregation-node steering and walks through Alibaba’s ultra-scale eCore deployment for AI scale-across.
Meta: SRv6 powered backbone networks
Jasmeet Bagga and Andrey Golovanov, Meta
Meta is deploying SRv6 traffic engineering on regional backbone rings that share compute and storage inside a latency envelope. The team already runs MPLS SR-TE at scale and still chose SRv6—IPv6 Traffic Class carries the six Olympic QoS classes their MPLS encoding cannot. The data plane uses uSID to encode up to six uSIDs in the IPv6 destination address (DA), each entry an adjacency (uA) through the fabric. A per-plane FBOSS/SAI controller programs the network; first RBB sites with SRv6 TE are in production on SAI with Cisco Silicon One.
Google: Virgo — campus-as-a-computer, with SRv6 for bare-metal
Sunil Ahluwalia and Jeongkeun “JK” Lee, Google
Virgo is Google’s dedicated accelerator-to-accelerator scale-out fabric beside Jupiter. Sunil Ahluwalia and JK Lee describe a two-layer, multi-plane design that treats the campus as a computer. Routing combines adaptive switching with SRv6 source routing for bare-metal customers running their own host stack, plus congestion signaling, high-speed telemetry, and Falcon transport.
OpenAI: Clustermapper — deterministic path diagnosis using SRv6
Ryder Lewis and Guohan Lu, OpenAI
Once an AI backend uses SRv6 for strict source routing, the same path identifiers can diagnose the fabric. Ryder Lewis and Guohan Lu present Clustermapper, OpenAI’s tool for deterministic path testing, and call on switch/ASIC vendors, SONiC contributors, and operators to qualify SRv6 steering and make path testing part of fabric acceptance.
Alibaba: Beyond Traditional BGP
Eddie Ruan and Dennis Cai, Alibaba
Traditional BGP in a large AI Clos explodes routing state. Eddie Ruan and Dennis Cai present HPN Pivot: infrastructure routes stay in BGP, service routes are distributed by a small set of cluster-proximate virtual route reflectors, and BGP Group GR keeps forwarding up even if every RR disappears. The work continues in the SONiC Routing working group.