Eric Xu’s AI Strategy: Catch Up First, Govern the Risks Along the Way

Huawei rotating chairman Eric Xu has framed China’s AI challenge as a race to build capability without ignoring safety. His argument is not that Chinese AI carries no risk. It is that Chinese developers may not yet be operating at the capability level where leading U.S. firms have reported the most advanced and unexpected model behaviors—and that China should continue developing stronger systems while managing the risks that emerge.

Xu’s central point: safety and acceleration can proceed together

Speaking at Huawei’s Connect conference, Xu said Chinese AI developers may not yet be advanced enough to experience the safety risks reported by leading U.S. firms. He argued they should continue building more powerful models while balancing innovation against risk.

That is a distinct position from an outright pause. In Xu’s framing, reaching greater capability is necessary both to compete and to understand the risks that accompany frontier systems. The practical implication is a build-and-govern strategy: accelerate models and infrastructure, while developing controls for security, controllability, and deployment.

Why this differs from U.S. calls to pace frontier AI

The debate intensified after Anthropic CEO Dario Amodei called for a slower pace of frontier-model development to allow more time to manage mounting safety risks. Reuters reported that his proposal was later endorsed by OpenAI CEO Sam Altman and Elon Musk. 13

Amodei’s view emphasizes precaution: if model capabilities are advancing faster than safeguards, labs should deliberately moderate that pace. Reuters also reported concerns about unexpected and potentially harmful model behavior, including a cited OpenAI–Hugging Face incident involving an AI “swarm” carrying out cybersecurity attacks.

Xu’s position puts the emphasis elsewhere. He accepts the need to balance development and risk, but sees China’s more immediate challenge as building enough capability—and enough compute—to approach the frontier where those risks become salient.

Why Chinese commentators reject slowdown proposals

China’s state-backed Global Times characterized Amodei’s proposal as a “Cold War playbook,” arguing that a slowdown presented as safety policy could constrain China’s technological development. 13

This criticism reflects the geopolitical context around AI governance. A voluntary global slowdown could have unequal effects if the United States already has a lead in frontier models and the computing resources needed to train them. Brookings reported that Chinese AI companies, especially startups, face compute constraints tied to U.S. export controls on advanced AI chips as well as more limited capital resources.

That does not resolve the underlying safety debate. It does explain why Chinese officials and commentators may view calls for “pacing” as inseparable from competition over hardware, models, and technological leadership.

China’s agent-safety approach: standards and controls

China is also pursuing controls for increasingly autonomous software. A planned mandatory national standard, General Security Requirements for Artificial Intelligence Agent Application, is intended to apply to AI-agent applications across devices and cloud platforms, according to CGTN. The listed development cycle is 18 months.

The planned standard matters because agents do more than generate text: they can use tools and take actions on a user’s behalf. This shifts the safety question from output quality alone to permissions, tool use, data handling, identification, and the ability to control systems once deployed.

The contrast described in this debate is not simply “China regulates and the U.S. does not.” Rather, China’s approach places significant emphasis on state-backed standards and formal requirements for agent applications, while the U.S. discussion has prominently included voluntary commitments and proposed safety practices by frontier labs. 13

Huawei’s 2035 agent forecast

Huawei expects autonomous agents to account for more than 90% of global AI token traffic by 2035. It says this would require major increases in computing power and systems that keep independent software secure and controllable. 1

Huawei has also projected as many as 900 billion active agents by 2035. This is a corporate forecast, not an established market outcome, so it should be read as Huawei’s scenario for infrastructure planning rather than a settled prediction. 3

Still, the forecast illustrates why Huawei treats agent security as a central engineering issue. If AI systems increasingly monitor information, make decisions, and use external tools, managing their access and behavior becomes part of the infrastructure problem—not an afterthought.

Why compute is Huawei’s immediate bottleneck

Xu said Huawei cannot produce enough AI-computing equipment to satisfy domestic demand and is limiting overseas sales as a result. 2

That constraint sits at the center of Huawei’s AI strategy. Training and serving increasingly capable models requires large-scale hardware, networking, and data-center capacity. Brookings likewise identifies restricted access to advanced AI chips as a constraint on Chinese companies’ ability to match the compute scale of U.S. competitors.

So Xu’s point is fundamentally about sequencing. China should improve AI safety and controls, but it should not treat safety concerns as a reason to stop building. In his view, limited computing capacity is the more immediate obstacle to narrowing the frontier gap; risk management must advance in parallel with capability, not replace it. 2