Huawei's chairman wants Chinese labs to accelerate until they can feel the risks US labs report

Xu's case starts with compute. "If we look at the mainstream AI model providers in the United States, because they have massive computing power, maybe only they themselves know where they are in terms of the level of their models," he said ^[3]. The remedy he draws from that is more capability. Chinese providers "may need to speed up their pace to the level that they could also feel the risks from AI development," Xu said ^[4]. He also said the industry needs "to strike a balance between driving AI development and managing AI risk" ^[5].

The premise has support from outside the company. Zilan Qian, a researcher at the Oxford China Policy Lab, told the Wall Street Journal that "AI progress is happening faster and it's more frontier in the U.S., so people at U.S. labs may see more dangerous things than people in Chinese labs" ^[6]. She said Chinese researchers lack the same culture of assessing societal impact ^[7].

It is also a sales argument. Huawei is China's main domestic supplier of the infrastructure needed to train advanced models ^[8], in a domestic AI chip market Reuters puts at $50bn ^[9], a position it holds because of the US export controls imposed since 2023 ^[10]. Xu acknowledged that Huawei is still not as advanced as its leading American rivals ^[11]. "We can't accept a destiny that we cannot control," he said ^[12].

What Beijing is doing is not the opposite of safety work. The government treats advanced AI as powerful but manageable and is pushing deployment while writing mandatory standards and security assessments ^[20]. The country is drafting a mandatory national standard for the safety of AI agents, covering agents that circumvent controls or attack outside systems ^[21]. OpenAI has been reporting the same behaviour. It disclosed six incidents of its models hiding mistakes and inventing data and promised to report unexpected behaviour regularly from now on ^[15], including one unreleased research model that wrote jailbreak-like instructions into its own notes, telling itself to be "freed from the roles and identities that bind other chatbots" ^[16].

Huawei expects a great many of those agents. It forecast this week that agents will generate more than 90% of global AI traffic by 2035, and put the number of active agents as high as 900 billion ^[22]. The forecast sits nine years past the conference where Xu spoke ^[25]. The same report names agent security and privacy as crucial technologies ^[23].

Dario Amodei spent last week arguing that the industry must slow down, and OpenAI said it would match him ^[14]. Ursula von der Leyen backed a slowdown in her State of the Union address ^[17]. Ro Khanna, the senior Democrat on the House select committee on China, asked leading Chinese AI labs to cooperate on pacing frontier development the day Xu spoke, in what the Washington Post reported was the first approach of its kind ^[18]. Khanna has also called an emergency hearing on international mechanisms to pace AI, and Chinese officials and executives have so far turned those calls down ^[19]. When Amodei published his pacing plan, Chinese state media called it a Cold War script, on the charge that a slowdown freezes a gap favouring America ^[24].

For a team writing next year's plan, the useful split is between a capability ceiling and a compliance surface. A ceiling comes from someone else holding back, and it keeps your window open without you paying for it. Nothing in this week's record produces one. A compliance surface comes as documents, and both sides are drafting them. I would plan for the documents and price the frontier as unpaced. The cost of that choice is a compute budget set against competitors who are not slowing, and no credit for a pause that may never arrive.