AMD doubled its server CPU TAM forecast to $120B by 2030 in its Q1 2026 earnings call, up from $60B it put down six months ago. The reason is not a better processor. Agentic AI systems spawn CPU tasks at the same rate GPUs spawn model calls, and those orchestration workloads (tool calls, memory retrieval, decision trees, inter-agent coordination) run on CPUs, not accelerators. Lisa Su called it directly: the CPU:GPU ratio in AI deployments is converging from 1:4 or 1:8 toward 1:1, and at enough agent scale, you can have more CPUs than GPUs.
The mechanism is straightforward. A GPU runs the model. An agent wraps that model in a loop, calls external tools, manages state, routes to other agents, and synthesizes results. Every step in that loop outside the model call runs on a general-purpose CPU. As agentic workloads scale, CPU compute is not a host overhead; it is proportional to the work being done. A gigawatt AI campus in 2026 was sized with CPUs as thin orchestration nodes. That sizing assumption is now wrong.
Hardware teams designing AI infrastructure who used 2024 CPU allocation models are building racks that will be CPU-bottlenecked within two years. Budget a fresh sizing exercise against your agent-to-model ratio before the next procurement cycle. AMD and Intel both gain from this shift, but the bigger move is at the hyperscalers: with CPU headcount back as a first-class variable in the rack, every Arm server CPU program just got a louder business case. Expect the next wave of custom-silicon CPU announcements out of AWS, Google, and Microsoft to lean harder into agent-orchestration workloads, not just general-purpose host duty.