I'm skeptical about that view. Better inference is actually a positive because it enables us to run more agents, and the control plane along with the myriad of other agentic tool calls and compute actions will require CPUs. It might not be obvious, but traditional computing excels at concrete computations and things that operate within a finite state machine. Inference is great for nondeterministic and stochastic correlations where the outcome can be comparatively imprecise or non-idempotent. $Intel(INTC)$ , $Advanced Micro Devices(AMD)$ , and $ARM Holdings(ARM)$  all stand to benefit as agent usage grows. $Cloudflare, Inc.(NET)$  Cloudflare noted the other day that agent traffic already exceeds human traffic. If we fast forward, it will increase by orders of magnitude, exponentially raising compute demands. That's why all the hyperscalers and adjacent mega-caps see the need to invest in custom silicon via ASICs—they all want to accelerate their specific workloads in preparation for this.