Distributed AI training reshapes inter-datacenter networks

Large AI models are increasingly trained synchronously across clusters in multiple datacenters, with Google, Microsoft, AWS, Meta and CoreWeave already using such setups. Synchronized traffic bursts demand coherent optics and up to 14x the bandwidth of conventional DCI, or the network becomes a training bottleneck.

• Google trained Gemini synchronously across clusters in multiple locations
• By 2030 the largest frontier training runs could draw 4–16 GW
• Cisco estimates bandwidth needs can reach 14x a conventional DCI baseline
• Inter-site links require coherent optics carrying 400–800 Gbps