Bandwidth and latency problems are no longer edge-case complaints. According to A10 Networks' State of AI Infrastructure Report 2025, cited in DE-CIX's analysis, bandwidth issues affected 59% of organisations in 2025 — up from 43% the year before. Latency concerns rose from 32% to 53% over the same period. Storage bottlenecks are beginning to surface too, with 41% of organisations planning infrastructure upgrades.
The backdrop: AI workloads are not uniform. Training pipelines demand raw throughput — they move large volumes of data into models and tolerate some latency. Inference is a different problem entirely. Real-time decisions, sensor-driven systems, and interactive user applications are sensitive to every additional millisecond. As AI moves into production and into edge environments, the gap between what the network can deliver and what the application requires becomes consequential.
Hybrid and multi-cloud deployments dominate. Roughly 42% of organisations combine on-premise infrastructure with cloud, and 35% lean heavily on public cloud. But distributing workloads across clouds, data centres, and edge locations creates routing complexity that traditional connectivity models were not designed to handle.
Direct peering — connecting networks without passing through multiple transit providers — is one mechanism for reducing that complexity. About half the organisations surveyed are already using peering in some form; the other half are not. For AI specifically, the case rests on reducing the number of network hops and gaining more control over traffic paths. Cloud on-ramps and interconnection platforms such as DE-CIX allow organisations to move data between Azure, GPU clusters, and other resources more directly than default transit routing allows.
Brandon Ross, senior interconnection consultant at DE-CIX, concludes that the limiting factor for AI will not be compute or model sophistication alone, but how efficiently data can move between distributed systems. With 44% of organisations already naming IT infrastructure as the top barrier to AI scaling, and bandwidth and latency pressures rising sharply, the network is moving from background concern to strategic constraint.
