Report: CPUs strained by agents

Artificial intelligence agents are pushing cloud demand beyond graphics chips and into ordinary server processors, with Amazon reportedly expanding CPU capacity and still hitting limits. (wccftech.com) A central processing unit is the general-purpose chip that runs operating systems, databases, networking and task scheduling, while a graphics processing unit is the specialized chip that does the heaviest model math. Agent systems use both: the graphics chip generates answers, and the central processor handles the surrounding work that moves data, calls tools and coordinates steps. (intel.com) Futurum Group said on February 24 that agentic workflows and reinforcement learning workloads are driving “massive general-purpose compute” for simulation, orchestration and disaggregated inference. The firm said CPU-to-GPU ratios in artificial intelligence clusters are moving back toward 1:1 as those workloads expand. (futurumgroup.com) That changes the bottleneck inside data centers built for the 2023 and 2024 graphics-chip crunch. Instead of waiting only on accelerators, cloud operators now also need more host processors to run memory management, storage, networking and the software layers around each model request. (futurumgroup.com) Amazon has been signaling unusually strong demand for its own infrastructure. In his April 9, 2026 shareholder letter, Chief Executive Andy Jassy said Amazon’s chips business, including Graviton, Trainium and Nitro, had an annual revenue run rate of more than $20 billion and was growing at triple-digit year-over-year rates. (aboutamazon.com) Jassy also wrote that two large Amazon Web Services customers asked whether they could buy all of Amazon’s Graviton capacity in 2026, and Amazon declined because it still needed to serve other customers. CNBC reported the same day that Amazon plans up to $100 billion in capital spending this year, with most of it tied to artificial intelligence projects. (aboutamazon.com) (cnbc.com) Other chipmakers are describing the same shift. Reuters reported on April 9 that Intel and Google expanded their partnership around artificial-intelligence-focused central processing units as changing artificial intelligence use renewed demand for traditional computing chips. (reuters.com) Advanced Micro Devices has also said demand is arriving faster than expected. Data Center Dynamics reported in March that Chief Executive Lisa Su told investors server CPU demand had exceeded expectations, driven primarily by agentic artificial intelligence applications, and that supply chains were tightening. (datacenterdynamics.com) The immediate effect is not that graphics chips stopped mattering. It is that each artificial intelligence deployment now needs a larger supporting cast of central processors, and cloud providers are spending on both at the same time. (intel.com) (futurumgroup.com) That leaves the industry with a broader capacity problem than the first wave of the generative-artificial-intelligence boom. The new question is no longer only who has enough graphics processors, but who has enough of the ordinary chips that keep the agents moving. (wccftech.com) (futurumgroup.com)