GPUs waste most of their capacity running looping, tool-calling AI agents

AAI agents loop continuously while calling tools, branching, backtracking, and maintaining context across many steps, which is very different from traditional model inference. Current GPUs reach only 30 to 40 percent utilization on these workloads because agent work alternates between memory-bound model calls, I/O-bound tool use, and CPU-bound orchestration. The need is purpose-built silicon for agent execution, with fast context switching and persistent memory across the whole execution graph.