Noah Taylor, of AFL, told RCR that the combination of fiber availability and power constraints presents a further challenge for AI infrastructure deployment.
In sum – what to know:
Traffic shifts – Inference, particularly real-time use cases, is creating more symmetric and upstream-heavy traffic while making ultra-low latency increasingly important.
Fiber constraints – Taylor says ribbon fiber lead times have stretched beyond 60 weeks in many markets, directly affecting AI build schedules.
Network redesign – Hyperscalers increasingly require multiple diverse and redundant connections, pushing networks from traditional hub-and-spoke designs toward more mesh-like architectures.
Artificial intelligence is changing the way data moves across networks, creating new demands for capacity, latency, and network resilience as AI workloads expand beyond traditional cloud environments.
According to Noah Taylor, head of market intelligence and growth strategy at AFL, AI is altering the traditional traffic model by generating substantially more upstream traffic, particularly as inference moves into real-time applications.
“AI is flipping the traffic model from the traditional heavy downstream (download) pattern to a much more symmetric—and often upload-heavy—profile,” Taylor said. “Inferencing at scale, especially real-time IoT use cases, is driving massive upstream traffic as data flows back to models for instant decisions. Bandwidth alone is no longer enough; ultra-low latency has become equally critical,” Taylor told RCR Wireless News.
That shift is putting particular pressure on long-haul networks connecting the growing number of AI data centers. Taylor said hyperscalers require high-capacity, low-latency connections between training clusters and the large datasets supporting them.
“Right now, the greatest pressure is on long-haul routes that connect the exploding number of AI data centers,” he said. “We’re seeing continued large-scale investment announcements in long-haul fiber exactly because hyperscalers need those high-capacity, low-latency links to connect the training clusters and massive datasets quickly,” Taylor added.
Metro networks could face greater pressure as AI inference increasingly moves closer to users and devices, while subsea networks are also affected by growing demand. For now, however, Taylor identifies terrestrial long-haul capacity and fiber plant as the more immediate constraint.
Higher-speed optical technologies are helping networks extract more capacity from existing fiber, but Taylor said they cannot eliminate the need for additional fiber. “400G, 800G, and emerging 1.6T transceivers are buying us valuable time and spectrum efficiency, but they cannot fully solve the problem alone,” he said. “The raw fiber count shortage is the limiting factor.”
That shortage is already affecting deployment schedules. Taylor said ribbon fiber lead times have extended beyond 60 weeks in many markets, creating a direct challenge for AI infrastructure projects.
The pressure is also changing how networks are designed. As AI data centers become increasingly interconnected, hyperscalers are seeking multiple diverse and redundant routes rather than relying on a smaller number of conventional connections. “AI is forcing a redesign of both physical and logical architecture around redundancy and resilience,” Taylor said.
“What used to be a hub and spoke model is quickly turning into more of a mesh network architecture design for outside plant,” Taylor said.
The combination of fiber availability and power constraints presents a further challenge for AI infrastructure deployment. “Fiber availability is of huge concern. “Power is also a strong contender and to be frank to solve both of these challenges in a traditional manner will require several years to accomplish,” he added.
Taylor said hyperscalers are therefore looking at alternative approaches to address those constraints, including higher-density fiber architectures and different approaches to power provision.
The economics of network deployment are changing alongside the physical requirements. Taylor said the scale of AI-related data center investment is allowing hyperscalers to place greater emphasis on performance and speed of deployment.
“AI is completely reshaping the economics,” he said. “Hyperscalers are spending roughly $700 billion on data centers—more than double the entire global telecom capex of ~$300 billion. They are willing to pay a premium for networks that deliver the highest bandwidth, lowest latency, and fastest time-to-revenue.”
For network providers, that means the challenge is no longer simply providing additional bandwidth. AI workloads are increasing the importance of fiber availability, low latency, route diversity, and the ability to deploy network infrastructure quickly enough to keep pace with data center construction.
The interview with AFL’s Noah Taylor is part of a report published by RCR Wireless News and RCRTech, titled Scaling Optical Networks for the Hyperscale and AI Era, which can be accessed by clicking here.
