Architecture Shift
Impact: Major
Strength: High
Conf: 85%
Cisco Defines 'Scale-Across' Networking Paradigm, Converging Silicon and Optics for Geographically Distributed AI Clusters
Summary
Cisco's blog outlines its 'scale-across' networking architecture to address geographically distributed AI training clusters. Centered on Cisco Silicon One P200 routing systems and 800G ZR/ZR+ coherent pluggable optics, the approach co-designs silicon and optics to deliver deterministic, low-latency, lossless connectivity for GPU clusters spanning tens to hundreds of kilometers.
Key Takeaways
Cisco categorizes AI infrastructure evolution into four bandwidth tiers: traditional DCI, front-end networks, scale-up, and scale-out, and formally introduces 'scale-across' as a new category.
Scale-across connects geographically dispersed GPU clusters, handling a small number of extremely high-bandwidth, loss-intolerant, synchronous, long-lived flows, distinct from traditional DCI. Cisco estimates a 100MW AI data center houses 60k-70k GPUs, requiring thousands of coherent optical ports for cross-site connectivity, far exceeding the 1k-2k ports for traditional DCI.
The solution converges Cisco Silicon One P200-powered routing systems (e.g., 8223, N9000, offering 51.2 Tbps) with its coherent pluggable optics portfolio (over 750k 400G DSP and 40k 800G DSP ports shipped). It also introduces the Open Transport 3000 Series open line system for multi-rail, multi-band operation to optimize power and scalability for scale-across networks.
Scale-across connects geographically dispersed GPU clusters, handling a small number of extremely high-bandwidth, loss-intolerant, synchronous, long-lived flows, distinct from traditional DCI. Cisco estimates a 100MW AI data center houses 60k-70k GPUs, requiring thousands of coherent optical ports for cross-site connectivity, far exceeding the 1k-2k ports for traditional DCI.
The solution converges Cisco Silicon One P200-powered routing systems (e.g., 8223, N9000, offering 51.2 Tbps) with its coherent pluggable optics portfolio (over 750k 400G DSP and 40k 800G DSP ports shipped). It also introduces the Open Transport 3000 Series open line system for multi-rail, multi-band operation to optimize power and scalability for scale-across networks.
Why It Matters
This is a classic **control layer shift**. The control point is moving from intra-data center switching/network architecture to the cross-data center (scale-across) co-design of silicon and optics and system-level optimization. Value is shifting from individual device performance/density to end-to-end reliability, power efficiency, and deterministic latency for the entire geographically distributed AI cluster. Cisco aims to seize the definition and control point of this new networking tier, elevating competition from single dimensions like chips or optics to the system-solution level.
PRO Decision
[Vendors] Networking and optical vendors must assess capability gaps in coherent optics, high-capacity routing, and system-level power optimization. They must engage in defining the technical standards and reference architecture for 'scale-across' to avoid marginalization in the next wave of AI infrastructure procurement.
[Enterprises] Enterprise architects planning large-scale AI clusters should treat 'geographic distribution' and 'cross-domain networking' as core design constraints for early evaluation. Focus on vendors' concrete implementations in silicon-optics co-design, deep buffering, and proactive congestion control, not just port bandwidth.
[Investors] Focus on companies with unique combinations in high-speed coherent pluggables, open line systems, high-capacity low-power networking silicon, and system integration. Cisco's direct linkage of 'scale-across' to its AI order growth (forecast >$3.6B for FY26 Q4) signals this market is entering a substantive capex cycle.
[Enterprises] Enterprise architects planning large-scale AI clusters should treat 'geographic distribution' and 'cross-domain networking' as core design constraints for early evaluation. Focus on vendors' concrete implementations in silicon-optics co-design, deep buffering, and proactive congestion control, not just port bandwidth.
[Investors] Focus on companies with unique combinations in high-speed coherent pluggables, open line systems, high-capacity low-power networking silicon, and system integration. Cisco's direct linkage of 'scale-across' to its AI order growth (forecast >$3.6B for FY26 Q4) signals this market is entering a substantive capex cycle.
💬 Comments (0)