GPU servers in a county AI rack must survive a single leaf failure. How should vNIC policies be aligned?

GPU servers in a county AI rack that must survive a single leaf failure need vNIC policies that bind hosts to fabric A/B aligned with redundant leaf pairs. UCS high-availability LAN design pairs fabric-A and fabric-B paths so a leaf or FI-side fault does not strand the GPU node or its RDMA sessions mid-job. Domain 3.2.d requires that redundant vNIC alignment for AI servers in municipal racks. Binding GPU servers to fabric A/B with vNIC policies pinned to redundant ports on one leaf keeps both paths on the same failure domain; if that leaf fails, both ports go down together and do not provide the redundant leaf-pair alignment required to keep the server online. Binding GPU servers to fabric A/B with vNIC policies aligned to redundant spine pairs builds redundancy upstream; if a leaf fails, spine-pair alignment alone does not provide the redundant leaf-pair alignment required to keep the server online, so a leaf failure can still strand the server. Binding GPU servers to fabric A/B with vNIC policies split across disjoint leaf pairs spreads paths across unrelated leaves without coordinated failover and does not create the redundant leaf-pair failover needed to survive a leaf loss. Exam caveat: host multipathing or teaming mode must match the fabric A/B design; policy alone without correct cabling to both leaves will not protect against leaf loss. Operational check: map GPU vNICs to fabric A/B toward redundant leaf pairs, validate link state on both paths, fail a single leaf in a maintenance window, and confirm the county AI servers keep RDMA and management connectivity on the surviving fabric path.