AI Inference Traffic Pushes Telecom Networks Toward a New Planning Problem

Light Reading says AI inference could create longer and more upstream-heavy telecom traffic than video-era networks were built to handle. Analysts cited in the report point to hyperscalers influence over optical transport, data-center interconnect and subsea capacity planning. Operators face a visibility gap because no comprehensive public study maps AI traffic patterns, even as the sector spends more than $600 billion in capex.

The impact sits in capacity, compute costs and supply chains: one deployment or bottleneck can change how companies buy chips, cloud contracts and data-centre space. Readers should track whether the announcement turns into available infrastructure, not just a product claim.

The Network Shift

Light Reading reports that AI inference could force telecom networks to handle a traffic pattern different from the video and browsing loads that shaped current designs.

The article cites a projection that inference may account for one-quarter of network traffic by 2035.

The issue is not only volume.

AI agents and applications can create longer sessions, heavier upstream demand and lower tolerance for degradation when the connection between an agent and its model becomes part of the service itself.

Hyperscaler Gravity

The report links that change to the growing role of hyperscalers in transport, data-center interconnect, edge infrastructure and subsea capacity.

MTN Consulting says telecom infrastructure is becoming more tied to cloud infrastructure as AI and cloud workloads move through operator networks.

TeleGeography data cited by Light Reading says hyperscalers now account for about 75% of total subsea bandwidth, up from almost zero in 2010, and are involved in more than two-thirds of planned new cables.

Google alone is planning eight Asia-Pacific cable projects.

Operator Incentives

Brian Washburn of Omdia told Light Reading that hyperscalers want to install their own optical equipment in partner carrier facilities so they can run private networks under their own control.

That can create demand for carrier sites and interconnection, but it also reduces visibility into traffic volumes.

Matt Walker of MTN Consulting said US operators see near-term AI training traffic positively, while agentic AI creates harder planning questions because growth could appear quickly in unexpected parts of the network.

Strategic Watchpoints

The planning risk is that operators may need to spend before they have reliable demand signals.

Walker said no comprehensive public study maps AI traffic volumes, patterns or growth, even as the industry spends more than $600 billion in capex this year.

The test for operators is whether AI traffic can become a priced connectivity opportunity rather than a cost imposed by cloud-controlled services.

Key signals include private optical deployments, edge capacity demand, subsea ownership patterns and any new traffic disclosures from hyperscalers or equipment vendors.