Hits $120B CPU TAM by 2030
- Morgan Stanley said on April 22 that agentic AI could add $32.5 billion to $60 billion to server CPU demand by 2030. - That pushes total server CPU TAM above $100 billion, with CPU orchestration responsible for 50% to 90% of agent workflow latency. - Chipmakers are already repositioning around that shift, with Intel, AMD, and Arm pitching CPUs as core AI infrastructure.
Server CPUs are back in the AI story — not as the star, but as the part you stop ignoring once systems get real. The big change is that AI workloads are getting more agentic, which means they do more than generate one answer and stop. They plan, call tools, move data around, keep context alive, and coordinate across a mess of software and hardware. Morgan Stanley argued on April 22 that this shift could add $32.5 billion to $60 billion of incremental CPU demand by 2030, taking total server CPU TAM to more than $100 billion. (aninews.in) ### What changed? The old AI spending story was simple — buy more GPUs. That still matters, but the newer story is about the system wrapped around the GPU. In agentic setups, the CPU handles orchestration, sc(aninews.in)s a chain of actions instead of one model invocation, the control plane starts to matter a lot more. (aninews.in) ### Why do CPUs suddenly matter more? Because latency is no longer just “how fast did the model run.” It is also “how long did the system spend deciding what to do next.” Morgan Stanley’s estimate is the key (aninews.in)ser experience can still be bottlenecked by general-purpose compute. (aninews.in) ### What does “more than $100 billion TAM” really mean? Not that CPUs are replacing GPUs. The catch is almost the opposite. GPUs stay essential for model training and a lot of inference, but the surrounding s(aninews.in)dening story, not a substitution story. (aninews.in) ### Are chip companies already leaning into this? Yes — very openly. Intel said in April that its multiyear AI infrastructure collaboration with Google reinforces the role of Xeon CPUs in orchestration, data (aninews.in)tional CPU for AI systems and host nodes paired with GPUs. (intc.com) ### What about Arm? Arm is pushing even harder on the thesis. Its new AGI CPU is marketed as production silicon built specifically for AI infrastructure at scale, with the pitch centered on orchestrating compute, managing accelerators, and coordinating thousands of (intc.com)as. That is not a generic server pitch — it is a direct bet on the “CPU as AI control plane” idea. (arm.com) ### Where does AMD fit? AMD is clearly framing data-center CPUs as part of enterprise AI infrastructure too, even if the cleanest public number here comes from Morgan Stanley rather than an AMD filing we could verify directly. AMD’s current positioning around EPYC stresses host CPU roles, memory capacity, and I/O bandwi(arm.com)amd.com) ### Is this really a CPU renaissance? Probably, but with an asterisk. This is not a return to the old world where CPUs dominated everything. It is a more balanced one where AI clusters need accelerators for raw model compute and stronger CPUs for coordination. Think of GPUs as the engine and CPUs as the traffic system — if the roads jam, a faster engine does not save the trip. (aninews.in) ### Bottom line? The real AI buildout is turning into a full-stack infrastructure story. If agentic AI keeps spreading, the winners will not just be the companies selling the biggest accelerators. They will also be the ones selling the CPUs, memory, networking, packaging, and system designs that keep those accelerators busy. (aninews.in)
