Zhou Hongyi talks about DeepSeek open source: The AI era requires "sovereign AI"

Recently, DeepSeek announced that DeepInfer, the internally used AI inference framework, will be fully open source, becoming another landmark move in the open source community. Zhou Hongyi, founder of 360 Group, said that behind this is "a kind of technological confidence, and also a concept of open source and inclusiveness"; looking at a deeper level, this matter has a more important meaning - technological sovereignty in the AI ​​era. He also revealed that 360 has built a sovereign AI operating system that has been put into use in some government and enterprise units.

It is understood that DeepInfer optimizes scheduling to allow the CPU and graphics card to work at the same time, reducing waiting and repeated calculations, allowing the same machine to handle more tasks. In Zhou Hongyi's view, the value of such open source achievements goes beyond the technology itself: as open source models and inference technology continue to mature, companies do not have to train large models from scratch, but also have the opportunity to deploy AI on their own servers, create agents that truly understand their own business, and take competitiveness into their own hands.

Zhou Hongyi believes that the core of sovereign AI is that users have control over AI. He divided sovereign AI into five levels: country, institution, enterprise, individual and family, and device.

Zhou Hongyi further pointed out that in the era of intelligent agents, AI sovereignty is particularly important to enterprises. The AI ​​currently used by many companies is "rented" - the model, computing power and platform are all in the hands of others, just like being built on a faucet controlled by others: the other party can only accept the price increase, and as soon as the service is adjusted, the business may stop. More importantly, after AI begins to work for the enterprise, who owns the data, accumulated knowledge and experience processed by the agent every day will directly determine the boundaries of the enterprise's core assets. "The stronger the AI, the more companies are controlled by others, which is absolutely not okay." In his view, what sovereign AI needs to solve is who controls the model, computing power and data, and who owns the knowledge and experience.

It is understood that the sovereign AI operating system that 360 has built covers computing resources and scheduling, sovereign large models, enterprise AI knowledge base, intelligent agent factory, Agent Teams (human-machine collaboration platform) and security governance capabilities, supporting enterprises to independently select models, schedule computing power, and conduct post-training based on their own data. The data remains within the enterprise's own network, and the knowledge, skills, experience and task execution records accumulated during use belong to the enterprise. In response to the potential risk of the agent's unauthorized operation, the system has built-in security capabilities to control and audit the agent's permissions, tool calls, and execution processes. Key operations are manually controlled, and the final control is in the hands of the enterprise.

Sovereign AI is becoming a global issue. McKinsey estimated in a 2026 report that by 2030, 30% to 40% of global AI spending may be affected by sovereign claims, corresponding to a market size of approximately US$500 billion to US$600 billion. For enterprises, future AI competition will not only focus on how smart the models are, but also who can truly take the initiative in cost, deployment and data control.