Growing pains: how distributed AI training changes the network between datacenters

Distributed AI training across multiple datacenters is increasing demand for high-bandwidth, low-latency inter-site networks; Cisco modelling puts aggregate bandwidth needs at roughly 14x a conventional DCI baseline. Cisco estimates that connecting two 100MW AI sites could require 12,000–32,000 coherent optical ports, versus about 1,000–2,000 for conventional DCI, and promotes its deep-buffer Silicon One P200 and coherent optics for these deployments. Scale-across AI networking remains early, with operators testing different architectures and deployments expected to develop over several years.

Analysis

Scale-across could shift part of AI infrastructure spend from power-constrained campuses toward regional fiber, coherent optics and high-buffer routing. The second-order benefit is not just networking revenue: sites with available power may become viable sooner, while longer-distance synchronization can reduce GPU utilization if latency and congestion are poorly managed. Operators may therefore need more compute to deliver the same training throughput, but that demand could be offset if distributed jobs prove inefficient.

Cisco is positioned to sell a co-designed stack, but this is a sponsored account of an early market, not evidence of material orders or durable share. Hyperscalers may favor in-house designs or multi-vendor sourcing; Arista Networks and Ciena are plausible competitive checks. CoreWeave’s cross-cloud flexibility may help differentiate its service, while also making performance more dependent on interconnect quality and counterparties. The article does not establish customer commitments, pricing, or revenue contribution.

Near term, little justifies repricing the group on this signal alone. Over 1–3 months, watch for named deployments, networking order commentary and evidence that inter-site fabrics are incremental rather than shifted capex. Over 6–18 months, successful deployments could broaden the addressable market for optics and routing; failed utilization or latency targets would undermine the case. The main contrarian risk is treating a larger network bill as a net positive: if networking complexity erodes training efficiency, hyperscalers could slow distributed expansion or redesign workloads.

Key Decisions for Investors

• No immediate trade on the sponsored article. Keep CSCO on a catalyst watchlist; consider a small long only after independently corroborated hyperscaler or AI-cloud scale-across wins and evidence of incremental networking revenue. Risk: the architecture remains experimental or customers build internally.
• Monitor CSCO against Arista Networks and Ciena for customer wins, product qualification, and order commentary rather than assuming Cisco’s co-design claims confer an advantage. A lack of named deployments or no discernible networking demand in subsequent results would falsify the near-term thesis.
• Track CRWV’s interconnect announcements as both a potential service differentiator and an execution risk. Do not infer improved economics without disclosed utilization, performance, and customer-demand evidence; weaker deployment or operating metrics would argue against treating cross-cloud training as a valuation catalyst.