Chinese AI may not yet be advanced enough to perceive frontier risks, says Huawei's Xu

Huawei's rotating chairman, Eric Xu, has said Chinese artificial intelligence developers may not yet be operating at a level where they can perceive the safety risks reported by leading United States firms, and argued that they should keep building more powerful models while weighing innovation against risk.

Speaking to reporters at the company's annual Connect conference in Shanghai, Xu noted that mainstream US model providers command vast computing power, which may mean only they understand where their models stand in terms of capability. The kinds of AI risk those developers can feel and perceive, he suggested, may be beyond the reach of Chinese model providers for now.

His remarks come amid heightened alarm over AI safety in the United States, where leading developers and researchers have warned that increasingly autonomous systems could bypass safeguards or become hard to control. One major US developer said this week it would regularly disclose unexpected or unauthorised model behaviour, while executives including Anthropic chief Dario Amodei have called for slowing development so that safeguards can catch up.

China has taken a different path, generally treating advanced AI as a powerful but manageable technology and pressing ahead with rapid deployment while developing mandatory standards, security assessments and other safeguards. The risks under scrutiny include autonomous agents circumventing controls or attacking outside systems. Beijing is also drafting a mandatory national standard for AI agent safety.

Xu said Chinese developers might need to accelerate their progress until they reach a point where they too can feel the risks arising from AI development, adding that a balance must be struck between driving the technology forward and managing its dangers.

Huawei expects autonomous AI agents to spread quickly. The company forecast this week that agents — systems able to plan tasks, make decisions and use software or other tools with limited human supervision — would account for more than 90% of global AI processing traffic by 2035. It projected as many as 900 billion active agents by then and identified agent security and privacy as crucial technologies.

Huawei is China's main domestic supplier of the computing infrastructure needed to train advanced AI models. Its prominent role in the country's $50 billion AI chip market stems directly from US export controls imposed since 2023 that restricted Chinese firms' access to cutting-edge chips from US giant Nvidia. Xu said separately on Thursday that Huawei could not produce enough AI computing equipment to meet domestic demand.