Agentic AI reshapes network demands: KPMG says inference workloads drive high-speed connection demand

Industry Background

As generative AI shifts from model training to large-scale inference deployment, network infrastructure is facing new pressures. Phil Wong, U.S. Technology Leader at KPMG, noted in an interview with RCR Wireless News that the rise of Agentic AI will fundamentally change traffic patterns between data centers. Agentic AI not only requires frequent interactions with enterprise data systems in the cloud, but inference workloads may also spread to the network edge, thereby driving strong demand for high-bandwidth, low-latency fiber connections.

Market Impact

This shift has profound implications for investors and operators of AI infrastructure. Currently, capital expenditure is heavily concentrated on compute (GPU) capacity, but Wong emphasized that every additional gigawatt of compute power comes with a corresponding connectivity requirement—and this ratio will increase as workloads shift from training to inference. This means network equipment vendors, fiber operators, and cloud service providers will face new growth opportunities. At the same time, power supply has become a more severe bottleneck than supply chain delays and labor shortages, with some hyperscale data centers already canceling committed capacity due to power delays and cost surges.

Competitive Landscape

Beneficiaries: Fiber operators and high-speed interconnect hardware vendors (such as Ciena, Nokia, etc.) are expected to secure new route construction orders. Cloud platforms focused on inference optimization and edge computing (e.g., AWS Local Zones, Azure Edge) may gain a first-mover advantage.

Under Pressure: Infrastructure operators located in traditional data center hubs may face demand outflow. AI projects in regions with insufficient power supply will be delayed or canceled, impacting related equipment suppliers.

Followers: Telecom operators and network service providers need to assess the return on investment for deploying fiber to remote AI campuses, which often do not pass through traditional population and business centers, challenging their revenue models.

Enterprise Insights

• Enterprises should focus on the following aspects:
• Inference Cost Management: As token consumption surges, enterprises need to proactively optimize Agentic architectures, reduce unnecessary inference calls, and adopt more efficient models.
• Network Planning in Advance: If planning to deploy large-scale AI applications, coordinate fiber and edge node resources with network service providers early to avoid connectivity bottlenecks affecting latency-sensitive businesses.
• Power Strategy: Data center site selection should prioritize power availability and evaluate green power supporting solutions to reduce long-term operational risks.

Future Outlook

Next 12 Months: Investment in inference infrastructure will accelerate, fiber construction orders will increase, but power constraints may delay the progress of new projects.

Next 24 Months: The widespread adoption of Agentic AI in enterprises will drive network traffic growth, giving rise to a new generation of inference-dedicated network architectures. Operators will begin to introduce differentiated pricing models for remote AI campuses.

Next 3 Years: The entry of physical AI (such as robotics, autonomous driving) may further push inference traffic to the edge, completely reshaping network topology.Next 3 years: Physical AI (e.g., robotics, autonomous driving) entering the market may further push inference traffic to the edge, completely reshaping network topology. The combination of efficient models and edge computing will partially alleviate bandwidth pressure, but the core network still needs continuous expansion.