Fang Shaoxia of Original Semiconductor: Agent needs a dedicated computer

Zhidongxi Editor | Zhidongxi Editorial Department Zhidongxi reported on July 10 that from July 2 to 3, the 2026 China AI Agent Conference with the theme of "Paradigm Shift Reshaping the World" was successfully held in Hangzhou. The conference gathered 64 heavyweight guests. Through an opening ceremony, two forums, and seven closed seminars, the core topics of the Agent track were fully analyzed. The topics covered Harness, self-evolving Agents, Coding Agents, multi-Agent collaboration, and Skills. At the meeting, Fang Shaoxia, founder and CEO of Granite Semiconductor, delivered a speech on the theme of "Agent Computer: Redefining the Computing Paradigm and Silicon Base". He systematically elaborated on the concept of Agent Computer and proposed that the Agent era needs to redefine computing platforms, chip architectures and computing power evaluation standards. Fang Shaoxia proposed that AI is moving from "answering questions" to "completing work". The next generation computing platform Agent Computer will be born, and the evaluation standard of chips will also shift from "peak computing power" to "task completion efficiency under unit energy consumption and cost." Fang Shaoxia divides the evolution of AI into three stages: from "tool" to "assistant" and then to "employee". When AI enters the "employee" stage, it requires a dedicated computer, namely Agent Com puter. He believes that AI PC only adds AI functions to computers and cannot meet the needs of agents. There are four natural conflicts between humans and agents: resources, security, persistence and system logic. For Agent Computer, Original Semiconductor proposed four principles for chip design: first, memory priority, second, long context priority, third, Agent must be resident 24/7, and fourth, energy efficiency priority. Fang Shaoxia believes that the computing power carrier of Agent will move from large batch GPU clusters to heterogeneous systems that couple reasoning memory and scheduling. The real competition in the future is not how high the peak computing power is, but "if I throw you a complex task, can it be faster and more stable?"