Rethinking local LLM inference as a full resource path across disk, RAM, PCIe, VRAM and compute—and why residency is only one part of the problem.