Zhou Hongyi, founder of 360 Security Technology, said on October 23 that the company has built a sovereign AI operating system and that it is already in use at some government and enterprise organizations. The idea is that a business should control its own models, compute and data rather than renting them.
Zhou made the comments while discussing DeepSeek's recent open-sourcing of DeepInfer, its internal inference framework, which schedules CPUs and GPUs to work at the same time and cuts waiting and repeated computation. He called the release a sign of technical confidence and of open-source inclusiveness. His wider point was that as open models and inference software mature, companies no longer need to train large models from scratch and can run AI on their own servers, building agents that understand their own business.
Zhou defines sovereign AI as the user having authority over the AI it relies on. He splits it into five levels: nations, institutions, enterprises, individuals and families, and devices. He argues that the enterprise level matters most in the agent era. Many companies use AI that is effectively rented, with models, compute and platforms controlled by someone else. If the provider raises prices, they must accept it, and if the service changes, their business can stall. Once agents start doing real work, he says, the question of who owns the data they process and the knowledge they build up decides where a company's core assets end.
According to 360, the operating system covers compute resources and scheduling, a sovereign large model, an enterprise AI knowledge base, an agent factory, Agent Teams, a human-agent collaboration platform, and security governance. Companies can choose their own models, schedule their own compute and post-train on their own data. Data stays inside the company's network, and the knowledge, skills, experience and task records built up during use belong to the company.
Security is built in to address the risk of agents acting beyond their permissions. The system controls and audits agent permissions, tool calls and execution, and key operations require human approval, so the final say stays with the company.
360 did not name customers or give deployment numbers, and the description so far is architectural rather than a set of technical specifications.
SEE ALSO: Alibaba, Tencent and 360 Enter the AI Office Agent Race · OpenAnolis Builds Anolis OS 27 Agent-Native