AI Agents: The Next Infrastructure Demand Driver

The first wave of generative AI infrastructure demand was driven by training and inference for large language models — workloads characterised by massive parallelism, high GPU density, and relatively predictable compute patterns. The second wave, which is now beginning to materialise, is driven by agentic AI: systems that operate autonomously over extended periods, making decisions, calling external tools, and managing complex multi-step workflows. The infrastructure requirements for agentic AI are meaningfully different from those of the first wave, and the data centre industry is only beginning to adapt.

The key distinction is latency and persistence. A language model responding to a user query can tolerate modest latency — a few hundred milliseconds is imperceptible to a human. An AI agent managing a complex business process, interacting with multiple external systems, and maintaining state over hours or days has very different requirements. It needs low-latency access to memory and storage, reliable connectivity to external APIs, and the ability to scale compute up and down rapidly as task complexity varies. These requirements favour edge deployments and distributed architectures over the centralised hyperscale campuses that have dominated the first wave of AI infrastructure investment.

The investment implications are significant. Edge data centres — smaller facilities located closer to end users and enterprise systems — are seeing renewed interest from hyperscalers and specialist operators alike. Networking infrastructure, particularly the software-defined networking and low-latency interconnect technologies that enable fast communication between distributed compute nodes, is becoming a critical bottleneck. And the memory and storage technologies that support persistent AI agent state — high-bandwidth memory, computational storage — are attracting substantial venture and growth capital.

At Lensbridge, our investment in PacketFabric — a leading Network-as-a-Service operator — was an early expression of our conviction that the networking layer would become as important as the compute layer in the AI infrastructure stack. The emergence of agentic AI reinforces that thesis and extends it to the edge. We continue to look for opportunities at the intersection of AI demand and infrastructure supply that are not yet fully reflected in market pricing.