Rakesh Chopra (SVP and Fellow, Hardware Architecture, Cisco) presented Cisco's case for rethinking AI networking at the OCP Global Summit 2025 in San Jose. He argued that power — not bandwidth — is the fundamental constraint on AI cluster scaling, and that training frontier models like GPT-5 already exceeds a single data center's power envelope, necessitating a new "scale-across" network tier. Cisco announced the 8223 router series and the Alibaba co-designed QuantumCat, both powered by the new Silicon One P200 chip, which delivers routing-level capabilities at switch-level power efficiency through a fully shared deep packet buffer architecture.

Key Takeaways

• Power is the binding constraint on AI scaling: a three-tier clos fabric connecting 33 million GPUs with today's switches would require 44 GW, and GPT-5 training at 50,000–100,000 GPUs already exceeds a single data center's power capacity.
• A new scale-across network tier — 16,000 ports at 13 petabits/sec — is required to connect multi-datacenter AI training clusters; modular chassis systems cannot meet this bandwidth, so fixed pizza-box leaf-spine topologies are the correct architecture.
• Cisco Silicon One P200 is a 51.2T full-duplex routing ASIC with a fully shared deep packet buffer and run-to-completion engine; compared to the 2023 Cisco 8804, the 8223-64EH replacement is 70% smaller, 85% lighter, and 65% lower power (real system power including optics).
• Cisco announced the Cisco 8223-64EF (OSFP800) and 8223-64E (QSFP-DD800) routers sampling to customers at OCP 2025, both SONiC-first, with IOS-XR and NX-OS to follow; Alibaba co-announced QuantumCat (128×400G, 4RU, ALiNOS) using P200 as silicon-only.
• Deep buffers are necessary, not harmful, for AI scale-across workloads: AI workloads are synchronous so job completion time is driven by the slowest path (not average jitter); at scale-across distances, failures require deep buffers to absorb reactive traffic during rerouting.
• Microsoft's David Maltz confirmed a 10+ year SONiC-on-routing collaboration with Cisco and that Microsoft deploys Silicon One across its entire network from top-of-rack to regional spines serving 1–4+ GW power zones.
• Alibaba's Eddie Ruan confirmed existing large nationwide Q200 DCI deployment and endorsed P200 for extending into core network chassis replacement, with the QuantumCat device physically demonstrated at the OCP booth.