At an Agentic-era AI infrastructure forum on July 19, Chinese Academy of Engineering academician Zheng Weimin said rapid expansion of compute capacity does not guarantee efficient token production. He warned the real scarcity is system capability to stably produce low‑cost, high‑quality tokens; tokens are transitioning from a technical metric to a core production input and inference systems are moving from single‑machine optimization toward distributed, cached, heterogeneous and service‑oriented architectures.(金十数据)