The routing protocol stack of an AI fabric is a stack of complementary decisions, each addressing a different dimension of the traffic engineering problem:
eBGP unnumbered eliminates link addressing overhead -- tens of thousands of fabric links managed with zero IPv4 address configuration. ASN design controls failure semantics -- same-ASN spine tiers prevent suboptimal routing at the cost of declaring connectivity lost when no valid path exists, which is the correct behaviour for training jobs that cannot tolerate degraded performance. BGP Link Bandwidth Community and DPF add traffic engineering above the baseline: weighted ECMP that adapts to capacity changes, and logical fabric partitioning that gives different tenants and different workloads their own dedicated routing planes.
RIFT and IS-IS Flex Algo represent the next generation: protocols that carry link-quality and topology metadata natively, eliminating the need for BGP extensions and out-of-band quality signalling. RIFT is the choice for pure-scale fat-tree deployments willing to invest in a newer protocol stack. IS-IS Flex Algo is the evolutionary choice for environments already running IS-IS who want multiple routing planes without migrating to a new protocol.
SRv6 enables path programming at the highest level of precision -- the controller dictates the exact sequence of switches each flow traverses. The flexibility is enormous; the operational burden of maintaining segment lists is correspondingly significant.
Multi-tenancy ties all of this together: EVPN-VXLAN provides the data-plane isolation between tenants, BGP DPF or IS-IS Flex Algo provides the routing-plane isolation, MIG + SR-IOV provide the server-level isolation, and GBP adds intra-tenant microsegmentation. In combination, these mechanisms transform a shared physical fabric into a platform where multiple organizations can train confidential models on the same switches, at the same time, with complete assurance that their data never crosses the isolation boundary.
The routing protocol is not a detail. In an AI fabric, it is the instrument that determines whether a training job completes on schedule, whether a tenant's gradient updates stay private, and whether a link failure propagates as a clean failover or as a cascading congestion event that costs hours of recompute.