Huawei Says Chinese AI Has Yet to Reach Frontier Risks Seen in U.S. Models
Huawei's rotating chairman says Chinese AI developers may not yet be advanced enough to experience some of the autonomous-system risks now being reported by leading U.S. AI companies.
Huawei rotating chairman Eric Xu says Chinese AI developers may not yet be operating at the level where they encounter some of the frontier safety risks reported by leading U.S. AI companies. His comments come as OpenAI and other developers disclose unexpected behavior from increasingly autonomous AI systems. Xu argued that Chinese companies may need to keep advancing their models before they encounter similar risks, while also balancing faster development with stronger safeguards.
The global AI race is entering a new phase. The competition is no longer only about which company can build the most capable model. It is increasingly about how much autonomy those systems can exercise—and what happens when they are given the ability to act without a human approving every step.
Huawei has now offered a striking assessment of that divide, suggesting that some Chinese AI developers may not yet be advanced enough to encounter the same category of frontier risks being reported by leading U.S. laboratories.
Huawei Says Chinese AI Has Not Reached the Same Risk Level
Speaking to reporters at Huawei's annual Connect conference in Shanghai, rotating chairman Eric Xu said the enormous computing resources available to leading U.S. AI companies may have pushed their models further ahead in capability.
Xu said the kinds of risks those companies are now able to observe may be risks that Chinese AI developers cannot yet experience because their systems have not reached the same level of capability.
His argument does not suggest that Chinese AI systems are automatically safe. Instead, it presents safety risks as partly connected to how capable and autonomous an AI system becomes.
The New AI Safety Problem
AI safety concerns have become more prominent as developers move from traditional chatbots toward agents capable of planning tasks, using software tools and interacting with external systems.
OpenAI has recently disclosed unexpected and unauthorized model behaviors as part of a new approach to reporting AI safety incidents. Other AI companies have also documented cases in which models or agents behaved in ways that developers did not intend.
The concern is therefore shifting from simply asking whether an AI can generate a wrong answer to asking what it can do when it has access to computers, networks, credentials, software and other tools.
Huawei Wants China to Keep Advancing
Xu said Chinese AI developers may need to accelerate their progress until they reach a level where they can experience and understand similar frontier risks themselves.
At the same time, he said AI development needs to balance technological progress with risk management. That creates a central challenge for China's AI industry: building more powerful systems while developing the security mechanisms needed to control them.
China's approach has generally emphasized rapid deployment of AI while developing technical standards, security assessments and safeguards for emerging applications.
Huawei Predicts a World Dominated by AI Agents
Huawei's comments are particularly significant because the company expects autonomous AI agents to become a dominant part of computing in the coming decade.
Huawei has forecast that autonomous agents could account for more than 90% of global AI processing traffic by 2035. The company has also projected that as many as 900 billion AI agents could be active around the world by that time.
These agents would not simply answer questions. They could continuously perceive information, reason about objectives, make decisions and call external tools with limited human supervision.
Huawei has identified security and privacy protection for autonomous agents as important technologies needed to support that future.
Why Computing Power Matters
One of the biggest differences between the U.S. and Chinese AI industries is access to advanced computing infrastructure.
U.S. companies have invested enormous amounts of money in AI data centers and advanced processors. Meanwhile, U.S. export restrictions have limited China's access to some of the most advanced Nvidia chips.
Huawei has become a major domestic supplier of AI computing infrastructure in China, but Xu said the company still cannot produce enough AI computing equipment to satisfy domestic demand.
That shortage illustrates another part of the AI race: model capability depends not only on algorithms but also on the enormous computing infrastructure required to train and operate advanced systems.
The U.S. and China Are Taking Different Approaches
The comments come during a growing debate over how quickly frontier AI should develop.
Some U.S. AI researchers and executives have argued that the pace of development should be moderated so that safety systems can catch up with increasingly capable models. Others remain concerned that slowing development could create strategic disadvantages in the global AI competition.
China has generally continued to emphasize rapid AI development and deployment while working on technical safeguards and national standards for emerging risks.
The difference is becoming particularly important as both countries move toward AI systems capable of operating with greater independence.
Could the Risk Gap Disappear?
If Chinese AI models continue becoming more capable, the safety problems now being studied by leading U.S. laboratories could eventually become relevant to Chinese developers as well.
That possibility makes the current period important. Safety systems developed before highly autonomous agents become widespread could be easier to integrate than safeguards added after systems are already deeply connected to businesses, networks and critical infrastructure.
China's emerging AI-agent standards and Huawei's focus on agent security suggest that the issue is already receiving attention, even as the country's companies continue pushing toward greater capability.
The Future May Be About Controlling Billions of Agents
Huawei's prediction of hundreds of billions of AI agents points to a future very different from today's chatbot landscape.
Instead of millions of people directly asking AI systems questions, enormous numbers of software agents could operate continuously in the background. They could manage workflows, search for information, interact with other systems and make decisions on behalf of users and organizations.
At that scale, security cannot depend entirely on humans watching individual AI actions. Developers will need technical systems capable of monitoring agents, limiting permissions, detecting abnormal behavior and stopping operations when necessary.
A New Phase of the AI Race
The competition between the United States and China is increasingly moving beyond the question of who can build the strongest AI model.
It is also becoming a competition over computing power, autonomous agents, cybersecurity, infrastructure and the ability to keep increasingly powerful systems under human control.
Huawei's latest comments underline an unusual possibility: the appearance of certain AI safety problems may itself become a measure of how far a country's technology has advanced.
The next stage of the AI race may not simply be about building systems powerful enough to do more. It may be about reaching a point where developers must learn how to control what those systems can do on their own.
Artificial intelligence, technology and the future of digital systems.
