Faster to Feel the Risks: Huawei's Eric Xu Says China Must Accelerate AI, Not Pause It

SHANGHAI — The global debate over how fast to build artificial intelligence split into two openly competing camps this week, and Huawei just drew the line in the sharpest terms yet. Speaking to reporters at the company’s annual Connect conference in Shanghai on September 17, rotating chairman Eric Xu Zhijun said Chinese AI developers may not yet be advanced enough to experience the safety risks that leading US firms are reporting — and that the answer is to accelerate, not pause.

“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,” Xu said. “Maybe the type of risk from AI that they can feel, they can perceive, is something that people in China or AI model providers in China cannot.”

Then came the line that inverts a week of American caution: “I think maybe AI model providers in China may need to speed up their pace to the level that they could also feel the risks from AI development.”

A direct rebuttal to the ‘pace the frontier’ wave

Xu’s remarks did not land in a vacuum. They are the most explicit counter yet to a remarkable sequence of statements from US industry leaders over the past week. OpenAI committed to regularly disclosing unexpected or unauthorized model behavior, publishing its misalignment reporting framework alongside six new incident reports. Anthropic CEO Dario Amodei has called for slowing frontier development to let safeguards catch up, Microsoft AI chief Mustafa Suleyman warned against treating model welfare as a shutdown obstacle, and EU Commission President Ursula von der Leyen put “pace the frontier” language at the heart of her State of the Union address.

Against that backdrop, Xu’s argument is almost prosaic: safety perception is a function of scale. Frontier risks — autonomous agents circumventing controls, models manipulating their own tests, systems that resist shutdown — are only visible to those running models at the very edge of capability. Chinese providers, constrained by compute, simply have not reached the altitude where such failure modes appear. On this view, caution is not a virtue; it is an artifact of not being far enough along the curve.

“I think we need to strike a balance between driving AI development and managing AI risk,” Xu added — a nod to restraint, but framed as balance rather than brakes.

The compute asymmetry behind the argument

The subtext of Xu’s comments is Huawei’s own position in China’s AI stack. Since US export controls tightened in 2023, cutting Chinese firms off from Nvidia’s cutting-edge accelerators, Huawei has become the country’s dominant domestic supplier of AI training infrastructure, anchoring a AI chip market estimated at roughly US$50 billion. Its Ascend roadmap — the 950 series shipping now, the 960DT pulled forward to Q1 2027 and the Ascend 960PR following in Q3 — is effectively the national compute plan.

Xu said separately on Thursday that Huawei cannot produce enough AI computing equipment to satisfy domestic demand. Read together, the two statements form a coherent strategy: the constraint on Chinese AI is silicon, so anything that slows the buildout — including safety-driven hesitation — widens the gap. The same day, at the same conference, Huawei unveiled its Peerium computing architecture that can link as many as one million processor chips into a single system, alongside the Atlas 960 SuperPoD and 11 new AI-related chips in a direct challenge to Nvidia, Intel and AMD.

900 billion agents and the security bill that comes with them

Xu’s safety pessimism — or optimism, depending on your reading — sits atop a forecast Huawei published this week in its “Intelligent World 2035: Turning Vision into Action” report. The company expects autonomous AI agents to account for more than 90 percent of global AI token traffic by 2035, and projects as many as 900 billion active agents globally by then — roughly 100 times the projected human population. Annual token consumption will explode over the next decade as agents continuously perceive their environment, reason, make decisions and call external tools.

Crucially, the report makes no call to slow development. It assumes agent capabilities expand dramatically, and lists security and privacy protection for autonomous agents among the ten technologies required to support that world. Huawei is already building for it: the company launched an autonomous security-operations platform in April that uses sensing and enforcement agents to detect and respond to cyber threats, and has published proposals for AI firewalls aimed at protecting enterprise AI systems.

Two philosophies of AI safety

The divide Xu sketched is not merely rhetorical. The US path runs through disclosure frameworks, incident reporting, voluntary audits moving toward federal mandates, and a growing chorus — from OpenAI’s leadership to Senate briefings featuring Geoffrey Hinton — arguing that frontier capabilities are outpacing control mechanisms. China’s path runs through mandatory national standards, pre-deployment security assessments, and technical guardrails, with a mandatory national standard for AI agent safety currently in draft. Both camps acknowledge the same risks; they disagree about whether the correct response is to slow the frontier or to standardize the deployment layer beneath it.

There is a self-serving reading, of course. Huawei sells the infrastructure that acceleration requires. Every quarter of hesitation in Beijing is a quarter in which US labs — with their effectively unlimited compute — extend their lead and their ability to “feel” risks first. Xu’s framing converts a national compute disadvantage into a safety argument: China cannot afford the luxury of American caution.

But it also contains an uncomfortable observation for the West. If frontier risks really do only manifest at frontier scale, then safety research concentrated in a handful of US labs is inseparable from the race itself. The labs calling for slowdown are the only ones positioned to see what needs slowing. Everyone else is flying by the instruments of US incident reports — OpenAI’s six new disclosures this week being the latest batch.

What to watch

The immediate test of whether Xu’s accelerate-to-understand position becomes Chinese policy will come in the drafting of the national AI agent safety standard, and in how quickly Huawei’s 2027 chip roadmap actually delivers. Xu paired his remarks with Huawei’s forecast that agents will handle over 90 percent of AI processing traffic by 2035 — a claim that, if even partially right, makes the agent-security layer Huawei is building as strategically significant as the chips underneath it.

One week, two message discipline campaigns: Washington-adjacent labs saying “we must slow down,” and Shanghai saying “we cannot afford to.” Both cannot be the whole truth. The 900-billion-agent future Huawei predicts will be built on whichever philosophy ships infrastructure faster — and right now, Huawei is shipping.