TrendForce projects $60B CPU TAM
- TrendForce said on April 24 that Intel now sees AI inference shifting server buildouts toward more CPUs, with CPU-to-GPU mixes tightening from roughly 1:8 in training toward 1:1. - TrendForce also pointed to Arm’s estimate that agent-era AI data centers could need about 120 million CPU cores per gigawatt, up from roughly 30 million in traditional AI facilities. - The backdrop is a broader Wall Street and industry push to reprice CPUs as agentic AI infrastructure, not just GPUs, with Morgan Stanley modeling up to $60 billion of incremental CPU demand by 2030. (trendforce.com)
AI servers are no longer being sized only around graphics chips. TrendForce said this week that inference and multi-agent workloads are pushing data centers to buy more central processing units, or CPUs. (trendforce.com) The change showed up in Intel’s latest comments to investors. TrendForce, citing Intel’s earnings-call transcript, said the old mix of about one CPU for every seven to eight graphics processing units, or GPUs, in training clusters has already tightened to about one CPU for every three to four GPUs in inference. (trendforce.com) Intel Chief Executive Lip-Bu Tan said customers now see CPUs doing more of the control work in AI systems, including orchestration, control-plane tasks, and managing multiple agents and data flows. TrendForce said Intel thinks that ratio could move toward parity at 1:1, and in some cases tilt further toward CPUs. (trendforce.com) That is a different job from model training. Training is the phase where giant clusters chew through data to build a model, while inference is the phase where deployed models answer prompts, call tools, and handle live requests. (trendforce.com) (aninews.in) In agentic systems, the CPU often acts like the traffic controller. It routes requests, keeps context in memory, coordinates tool use, and handles lower-latency tasks that do not need a full GPU pass for every step. (aninews.in) (trendforce.com) TrendForce said Arm estimates traditional AI data centers use about 30 million CPU cores per gigawatt, while the AI-agent era could raise that to about 120 million cores per gigawatt. That is a fourfold jump in CPU demand per unit of power capacity. (trendforce.com) The vendor lineup is shifting with that demand. TrendForce’s April 8 research note said Nvidia used its March 16 GTC event to debut a standalone Vera CPU rack, and Arm on March 25 introduced its own Arm AGI CPU with air-cooled and liquid-cooled rack variants. (trendforce.com) TrendForce’s report also flagged tight CPU supply and said Intel and Advanced Micro Devices were planning price increases by the end of the first quarter of 2026. The same note highlighted AMD’s EPYC Venice, Nvidia’s Vera CPU, Arm’s AGI CPU, AmpereOne, and Amazon Web Services’ Graviton5 as products to watch. (trendforce.com) Wall Street is starting to put numbers on the shift. Morgan Stanley said on April 22 that agentic AI could create $32.5 billion to $60 billion of incremental CPU total addressable market by 2030, inside a total server CPU market that could exceed $100 billion. (aninews.in) The same Morgan Stanley note said agentic workloads could add 15 to 45 exabytes of dynamic random-access memory demand by 2030, equal to 26% to 77% of 2027 annual DRAM supply. That points to a broader hardware bill for AI systems, not a swap that leaves GPUs behind. (aninews.in) Intel’s own numbers show why investors are paying attention. TrendForce said Intel’s Data Center and AI group posted first-quarter revenue of $5.1 billion, up 22% from a year earlier, as demand for server-grade Xeon 6 processors improved. (trendforce.com) The near-term question is no longer whether AI needs GPUs. It is how many more CPUs, memory chips, and full racks cloud providers will need as AI moves from one-shot generation to systems that plan, call tools, and keep working. (trendforce.com 1) (trendforce.com 2)
