Photo: Tara Winstead / Pexels
Eric Xu says Chinese AI providers need to move faster to encounter the frontier challenges US firms already face, while unveiling plans for 900 billion autonomous agents by 2035
Huawei’s rotating chairman Eric Xu delivered a blunt message to China’s AI industry at the company’s Connect conference on September 17: you’re not moving fast enough. In a speech that doubled as a reality check, Xu argued that Chinese AI model providers need to accelerate development to the point where they actually encounter the frontier safety risks that leading US firms are already grappling with.
Xu’s framing was notably different from the usual nationalist tech triumphalism. Rather than claiming China is neck-and-neck with US AI leaders, he essentially acknowledged a capability gap by pointing to the absence of frontier safety challenges in Chinese computing environments. Current Chinese computing capabilities, he argued, haven’t exposed domestic developers to the kinds of risks and edge cases that companies like OpenAI, Anthropic, and Google DeepMind routinely encounter.
The prescription, in Xu’s view, is to balance aggressive development with proper risk management. Move fast enough to find the problems, but build the guardrails before they become catastrophic. What makes the statement particularly noteworthy is its source. Huawei isn’t some scrappy startup lobbying for fewer restrictions. It’s China’s designated backbone for domestic AI infrastructure, the company that the entire ecosystem depends on for the computing power that makes frontier models possible in the first place.
900 billion agents and the chips to run them #
Alongside the urgency about development speed, Xu painted a picture of what Huawei thinks the AI future looks like. The company forecasts that autonomous AI agents will handle over 90% of global AI processing traffic by 2035. The scale Huawei envisions is staggering: up to 900 billion active AI agents operating globally within a decade. For context, that’s roughly 100 agents for every human being on Earth.
To serve that future, Huawei is expanding its Ascend chip ecosystem with two new processors. The Ascend 960DT is slated for the first quarter of 2027, followed by the Ascend 960PR in the third quarter of that year. These chips are designed specifically to handle surging domestic demand for AI compute, a demand that exists in large part because US export controls have choked off access to Nvidia’s most advanced GPUs.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Huawei’s domestic AI computing market share now reportedly exceeds Nvidia’s within China. Even with that advantage, Xu conceded that Huawei cannot currently satisfy all of China’s domestic AI computing needs, and the company has signaled only minimal plans for overseas expansion.
The export control backdrop #
US restrictions on advanced semiconductor exports to China began tightening in 2019 and escalated significantly with the October 2023 export controls that specifically targeted AI-capable chips. Those rules effectively barred Nvidia from selling its most powerful processors to Chinese customers, creating the supply vacuum that Huawei has been racing to fill.
The result is a bifurcating global AI infrastructure. US-allied markets run on Nvidia, AMD, and increasingly custom silicon from hyperscalers. China runs on Huawei’s Ascend stack and whatever domestic alternatives can reach production scale.
What this means for the AI race #
For the global semiconductor landscape, Huawei’s planned 2027 chip releases represent a meaningful escalation. The autonomous agent forecast also deserves scrutiny. If Huawei is right that 90% of AI processing will serve autonomous agents by 2035, the infrastructure requirements look radically different from today’s chatbot-heavy workload. Agents need persistent compute, real-time inference, and robust safety layers — a different engineering challenge than training large language models, and one that favors companies that control the full stack from chips to cloud services. Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our