Principal analyst Sean Kinney’s recent research note said that agentic AI is increasingly a “consumer of network capabilities,” with networks having to connect not only applications, but also infrastructure that must be discoverable by machines that will have to discover, provision, secure, and consume what is ostensibly “infrastructure.” That’s because agentic agents are designed to achieve a high-level goal, not just execute a predefined command or generate text.

That fact will lead to an explosion of internal “East-West” traffic (machine-to-machine data movement inside a data center), changing physical and operational requirements in support of continuous, adaptive, and multi-step actions. In other words, the ways in which data centers are designed, powered, and secured will be fundamentally changed by the evolution from static, single-turn request:response constructs to extremely dynamic, multi-turn and autonomous execution. 

Arm has said that data centers must soon support up to 4x the CPU core density within the same power footprint to support the agentic AI era. Agentic workloads, it posits, will also introduce more unpredictable control flow, irregular memory access patterns, synchronization requirements, and I/O intensity.

That will necessitate tighter CPU-accelerator coupling, with racks engineered to support tightly coupled hybrid systems that integrate CPUs, GPUs, high-bandwidth memory (HBM), and advanced CXL fabric links. As summarized by Semiconductor Engineering, data centers will need to pivot from GPU-centric number crunching to CPU-driven orchestration, where managing long-running reasoning loops and context will be just as important as raw compute.