For most of the current AI buildout, GPUs and high-bandwidth memory have absorbed nearly all the attention as the scarce resources holding back deployment. A report published by IEEE Spectrum on August 16 describes a different constraint now emerging: server CPUs. The core issue is architectural. When a large language model handles a single-turn query, nearly all the work, the matrix multiplication at the heart of inference, runs on the GPU, and the CPU sits mostly idle. Agentic systems break that pattern. Every time an agent calls a tool, parses a result, decides which tool to call next, checks a safety guardrail, or hands work off to a sub-agent, that step typically executes on the CPU rather than the accelerator. Chipmakers cited in the piece estimate that a majority of the stages in a realistic agentic pipeline, as much as seven of eight steps in one vendor internal testing, run entirely on CPU cycles. As enterprises move from prototype agents to production deployments that spawn thousands or millions of concurrent agent instances, each with a chain of sub-agents making API calls through protocols like MCP, that CPU-side work multiplies fast. The report also flags a subtler compounding effect: tokenization. Every tool-call result an agent receives has to be re-tokenized against the running conversation, and that job runs mostly single-threaded rather than in parallel, so it slows disproportionately as context windows stretch into the hundreds of thousands or millions of tokens. The practical fallout is already visible: AWS has reportedly told engineering teams to conserve CPU capacity amid a surge in wait times, Intel says it has sold out server CPU capacity through year-end and is reallocating client-chip production toward servers, and AMD has doubled its server CPU revenue forecast. For teams building or scaling agentic products, the takeaway is that capacity planning can no longer treat CPU headroom as an afterthought behind GPU procurement; under sustained agentic load, an undersized CPU fleet can leave expensive GPUs idle while they wait for orchestration work to catch up.