A municipal AI team needs DSCP/CoS markings from UCS GPU servers to survive into the Nexus AI fabric. What alignment is required?

A municipal AI team that needs DSCP/CoS markings from UCS GPU servers to survive into the Nexus AI fabric must align leaf QoS policy with UCS vNIC QoS so markings applied on the server are trusted and preserved. End-to-end class consistency is the Domain 3.1.c requirement: relying on leaf access list re-marking to assign fabric queues from incoming server DSCP values overrides the server's intentional markings and can drop RDMA traffic into the wrong fabric queue, losing PFC/ETS treatment intended for clinic or city training classes. Enabling priority flow control on spine ports does not transport UCS markings toward fabric leaf queues; PFC only pauses frames on a link and does not program the leaf to trust or preserve server DSCP/CoS markings. Mapping GPU data flows to the UCS Fibre Channel system class applies the wrong domain—Fibre Channel system classes do not provide Ethernet DSCP/CoS queuing for wired UCS RDMA vNICs across Nexus leaves. Exam caveat: trust boundaries differ by platform (trust DSCP vs CoS, rewrite vs preserve); the exam point is deliberate UCS-to-leaf alignment, not relying on leaf access list re-marking to assign fabric queues from incoming server DSCP values. Operational check: set UCS vNIC QoS to mark the RDMA class, configure the adjacent Nexus leaf to trust/preserve those markings into the AI class-maps, capture packets or interface QoS counters during a GPU flow, and confirm the municipal fabric still classifies the traffic as intended RDMA.