Muse is the spark. The fire is the workload shape. Chat inference is GPU-bound. Agentic inference is a loop: reason, tool, parse, orchestrate, policy check, reason again. Intel and Georgia Tech put CPU tool processing at 50-90% of end-to-end latency. AMD says 7 of 8 stages in real agent pipelines run on the CPU. That is why the mix is shifting from about 1 CPU per 8 GPUs in training toward 1:1 in agents, and sometimes 4:1. Arm’s number is 30 million CPU cores per gigawatt today versus 120 million in the agent era. Muse makes it concrete. Every user gets a persistent Secure VM plus Sentinel. That is a long-lived cloud computer, not a chatbot session. Intel already said it can fill only about half of CPU demand. Lisa Su already called agentic sandboxes the fastest-growing slice of server CPU TAM. Muse did not create the shortage. It made the Street treat it as structural. Inference was already 2:1 versus training and still climbing. Muse is the first consumer-scale proof that those extra loops live on CPUs. $INTC $AMD $ARM