Cisco Scale-Across: Converged Silicon and Optics for Distributed AI Training
Summary
Key Takeaways
Cisco introduces Scale-Across to address the power wall (up to 100MW per site) forcing AI training clusters to span multiple data centers. Traditional DCI fails for GPU synchronization requiring lossless, deterministic performance over tens of kilometers.
Core components: Cisco Silicon One P200 (51.2Tbps, powering Cisco 8223 and N9000) with hyperscaler design wins; 400G ZR/ZR+ and 800G ZR/ZR+ coherent pluggables (750k 400G DSP ports, 40k 800G DSP ports shipped); Cisco Open Transport 3000 open line system (multi-rail, C+L band, multi-Pb long-haul) and NCS 1014 metro OLS (coherent probe, dynamic gain equalization).
Key technologies: deep-buffered routers (not switches) to absorb fiber latency (100MB in-flight on 100km link); proactive congestion control leveraging AI workload predictability; hardware-based security at every layer. Co-design of silicon and optics is emphasized for power efficiency.
Why It Matters
Cisco's Scale-Across is a strategic move to contain Arista and Nvidia in AI networking. By tightly coupling Silicon One routers with its coherent optics, Cisco aims to lock users into its optical supply chain—once deployed, customers cannot easily substitute third-party modules, losing flexibility and bargaining power.
Hidden cost trap: deep-buffered routers (e.g., N9000) are more expensive than standard AI switches (Arista 7800R3, Nvidia Spectrum-4). Proactive congestion control over 100km links is unproven at scale—tail latency and PFC/ECN bottlenecks are ignored. The open line system adds optical amplification sites, increasing power and operational complexity.
Cisco avoids discussing Head-of-Line Blocking when multiple high-priority flows share deep buffers. This is a control plane shift—proprietary silicon-optics integration transfers network control from open standards to Cisco's ecosystem.
PRO Decision
【Vendors】 Arista Networks should attack Cisco's lock-in by promoting standard Ethernet with open optical disaggregation, supporting third-party coherent modules. Nvidia should highlight its mature low-latency InfiniBand (Quantum-2) for cross-DC and offer pre-validated reference architectures with optical partners.
【Enterprises】 CIOs and architects must conduct zero-trust audits: demand interoperability proof for Cisco's optics (OIF/OpenROADM), evaluate TCO including optical amplification power, and require independent benchmarks on tail latency and throughput stability under proactive congestion control. Prefer multi-vendor decoupled solutions for cross-cloud portability.
【Investors】 Cisco's AI order growth ($3.6B forecast) is driven by hyperscalers, but its lock-in strategy may limit penetration in mid-tier and sovereign clouds. Track Arista and Nvidia's AI networking revenue growth—if Cisco's integrated silicon-optics fails in open ecosystems, its long-term share will erode.
Get 3-5 key AI infrastructure signals weekly →
💬 Comments (0)