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

Large-scale AI training is no longer confined to a single datacenter, as hyperscalers and cloud providers increasingly spread GPU clusters across sites to access more power, space, and fewer planning constraints. For CIOs, this shifts the datacenter interconnect from a mostly asynchronous replication path to a critical part of the AI compute fabric, where latency, congestion, and packet loss directly translate into slower training, higher costs, and greater operational risk. IT organizations will need to design for predictable, high-bandwidth, low-latency synchronization across sites, not just connectivity and resilience.