Transforming Deployment Challenges Into Opportunities With Cisco AI Networking

Author: Muri Ganduru, Data Center Networking Business Lead, Cisco Collection: 2025 OCP Global Summit - Expo Hall Sessions

Summary

Cisco presents its approach to mainstreaming AI networking infrastructure beyond hyperscalers into neocloud, sovereign cloud, and enterprise environments. The presentation identifies three core challenges: (1) traditional networking infrastructure was not built for AI communication paradigms (GPU-to-GPU, hybrid frontend/backend/storage flows), (2) shadow AI organizations building infrastructure outside standard automation frameworks create fragmented management, and (3) security must be infused into AI network architecture rather than bolted on. Cisco positions the Nexus platform as a common operating model across frontend, storage, and multiple AI backend environments using differentiated silicon, systems, software, and management.

The key technical argument centers on hybrid workload support using Cisco's custom silicon. AI workloads often require traversing multiple network domains (backend to frontend to storage) for operations like data prefetching before training. Cisco's silicon enables deterministic dynamic path selection -- using flowlet load balancing for some traffic patterns and per-packet load balancing for others, driven by standard DSCP-based QoS policy rather than new proprietary mechanisms. The Nexus Dashboard provides a unified management plane for building EVPN fabrics (frontend/storage), legacy three-tier designs, and purpose-built AI fabric templates from the same interface, with granular link-level visibility for rapid problem diagnosis. Cisco also announced the Cisco 8283 and 8223 switches with 51.2 TB switching capacity at OCP.