docs: Clarify how NICo provides tenant-network addressing and intra-VPC connectivity (routed L3 overlay)

[benhuntley]

How would you describe the priority of this documentation request

Medium

Describe the future/missing documentation

What to clarify: The Network Prerequisites → "Tenant Networks" section documents the /31-per-interface allocation, but doesn't explain how NICo actually delivers connectivity: that a tenant network is a routed L3 overlay, and how instances reach one another across it. Adding this makes the existing /31 detail make sense and sets correct expectations for ISVs building on NICo.

How NICo does it (to document):

• Each instance interface gets a /31 point-to-point link to its DPU; the instance is advertised to the VPC as a /32 host route over the VXLAN/EVPN overlay.
• DPUs are the VTEPs and interconnect over the overlay, so any instance reaches any other by routing — instances are routable but not L2-adjacent and don't share a flat subnet.
• Implication for ISVs: treat the VPC as a routed network (run the CNI in routed/overlay mode), don't assume L2 adjacency. (Surfaced when SpectroCloud's K8s cluster-formation assumed flat L2.)

Existing page: https://docs.nvidia.com/infra-controller/documentation/getting-started/prerequisites/network

Code of Conduct

• I agree to follow this project's Code of Conduct

[benhuntley]

Also capture the Kubernetes/ISV angle (SpectroCloud, Reliance): ISVs should run the CNI in routed/overlay mode (instances are not L2-adjacent). The related K8s questions are already answered elsewhere — VIPs via the BGP-Peering page, the NICo-vs-ISV responsibility split in Scope & Boundaries, control-plane HA is the ISV's responsibility, and no CAPI provider today.