According to Phil Wong, technology principal at KPMG US, developers are increasingly locating large AI campuses outside established hubs because of land and power constraints
In sum – what to know:
Inference growth — KPMG expects agentic AI and inference workloads to drive demand for high-speed, low-latency connectivity while changing traffic patterns across cloud and AI infrastructure.
Power constraints — Reliable power remains the biggest obstacle to AI infrastructure expansion, ahead of supply chain delays and labor availability.
Connectivity follows — While most AI investment is currently directed toward compute, demand for network infrastructure is expected to grow alongside the shift from AI training to inference.
As enterprises move toward agentic AI, inference workloads are expected to reshape network traffic patterns and increase demand for high-speed, low-latency connectivity, according to Phil Wong, technology principal at KPMG US.
Wong told RCR Wireless News the next phase of AI adoption will create more traffic between traditional enterprise cloud environments and AI-specific compute infrastructure, with some inference workloads potentially moving closer to end users.
“As we move into the Agentic AI era, traffic coming from inference workload will drive demand for high speed, low latency connectivity (i.e., fiber). Agentic AI works best when combined with data, context, and memory. There will be increased traffic between traditional cloud, where enterprise data and systems of record are stored and AI-specific compute. We could also see inference traffic spread more towards the edge of the network, closer to the end users, especially if physical AI takes off,” he said.
The expansion of AI infrastructure beyond traditional data center markets is also creating new connectivity requirements. According to Wong, developers are increasingly locating large AI campuses outside established hubs because of land and power constraints, driving demand for new fiber routes.
“As more larger scale data center developments move outside of the traditional data center markets because of land and power constraints, you will see the need to build new, high bandwidth fiber routes going to these locations. The challenge for fiber operators is whether they can get good ROI from these routes that may not pass through traditional population and business center,” Wong added.
Despite growing demand for connectivity, Wong said power availability remains the industry’s biggest bottleneck. “Currently, access to power, on or off-grid, is the biggest challenge, followed by supply chain delays and availability of labor. These challenges lead to longer deployment timelines and capital spend. In some cases, hyperscalers have cancelled already committed capacity because of the delays and prospect of ballooning costs,” he said.
He also noted that AI infrastructure spending today is primarily directed toward compute capacity, although connectivity requirements will continue expanding as workloads evolve.
“Most of the current capex is focused on building compute (GPU) capacity. However, for every gigawat of new compute, there is a corresponding requirement for connectivity, and that would rise as workload shift from training to inference and agentic AI. Hence, the demand will still be there in the long run. ROI for specific routes will need to be evaluated as some of the deployments will be farther away from traditional business and population centers, which means operators may not be able to capture additional revenue for the incremental infrastructure,” Wong said.
Looking ahead, Wong identified reliable power as the primary factor that will determine how quickly AI infrastructure can scale over the next several years.
While he expects AI infrastructure investment to remain strong, Wong also anticipates greater attention to AI efficiency as token consumption and inference costs increase.
“From an end user demand perspective, we expect demand growth to continue at a crisp pace in the near term. The demand of compute and storage is expected to continue to increase substantially as adoption of AI/Agentic AI continues across enterprises and for consumers. Agentic AI with reasoning, multi-modal processing, and physical AI are all going to drive explosion in token consumption and hence AI-related compute and storage infrastructure. However, we do expect enterprises and providers will start to manage the consumption more proactively as the token usage and costs increases, models and how usage is being orchestrated will get better. Technologists will find ways to optimize current agentic architecture for token consumption,” he said.
The interview with KPMG’s Phil Wong is part of a recent 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.
