Every chapter so far has assumed that traffic finds its way between GPUs. Chapter 2 showed you why the traffic is so demanding. Chapter 5 showed you how to protect it with PFC and ECN. Chapter 6 showed you how to spread it across links. But underneath all of that, a routing protocol is making the fundamental decision: which path even exists? Which next-hop does this packet go to? How does the leaf switch know that the far GPU is reachable at all?
In traditional enterprise networks, routing is an afterthought -- OSPF runs, routes converge, done. In an AI fabric, the routing protocol is a precision instrument. The wrong ASN design produces suboptimal three-hop paths when a link fails, saturating an intermediate spine and killing a training job that had been running for six hours. The wrong load-balancing community leaves 800 Gbps links half-empty while others hit 100%. The wrong protocol choice makes DragonFly topologies unmanageable at scale.
This chapter takes you through the full routing stack that makes a modern AI fabric work -- from why eBGP won the data center routing war, through the advanced BGP capabilities that create logical fabric partitions for multi-tenant GPU clusters, to the emerging IGP protocols (RIFT and IS-IS Flex Algo) that are beginning to displace BGP where link-level awareness matters more than policy flexibility, and finally to SRv6 and the multi-tenancy isolation model that turns physical fabric into GPU-as-a-Service.