Why China Is Falling Behind in AI | by PKU Scholar Hou Hong
"Chinese entrepreneurs are left with a stark choice: bow down to platform dominance or flee overseas." – Hou Hong
This article was first made available to paid subscribers last week.
To enjoy early access to select content and full access to Sinification’s archive, consider becoming a paid subscriber. As a non-profit, Sinification depends on your support.
Today’s article is introduced by Paul Triolo, Technology Policy Lead at DGA-Albright Stonebridge Group, where he advises companies on AI strategy. He is also a Non-Resident Honorary Senior Fellow on Technology at the Asia Society Policy Institute’s Centre for China Analysis. Over the past year Paul has spoken directly with many of China’s leading AI firms, attended the World AI Conference in Shanghai in July 2025, and participated in an investor tour in October that included a wide range of major AI companies and industry players. He writes frequently on China and AI in his excellent Substack, AI Stack Decrypted. We are very grateful to him for bringing this first-hand expertise to this edition. — James and Thomas
Recently there has been much debate within Chinese industry and academia over how the country stands with respect to the United States in developing and deploying AI. A recent commentary by Hou Hong, titled Personal insights on the US-China AI gap, falls on the side of a group, including within large Chinese AI companies, who hold that major structural issues will keep Chinese companies behind their Western, mostly US counterparts. They hold that especially as competition shifts from “who has the best model” to “who builds the strongest ecosystem for deployment and compounding usage”, Chinese companies will be at a disadvantage.
The Pessimists: China’s Disadvantage in Compute and Commercialisation
For Hou, China is not portrayed as losing primarily because it cannot produce strong models. Rather, the author’s main thesis is that factor conditions, demand conditions, and competitive structure are interacting in a mutually reinforcing negative cycle—keeping China on a lower trajectory in agent deployment and especially in building an Internet-of-Agents-style ecosystem, while the US compounds advantages through faster deployment, stronger rivalry dynamics, more open ecosystem surfaces, and heavier infrastructure/ecosystem investment.
In a recent roundtable, technologists from Alibaba, Tencent and Zhipu also converged on a pragmatic view of what still holds China back from “catching up” with the US: the constraint is no longer just model ideas, but the surrounding system—compute supply chains and commercialisation.
One argued that once a technological approach is proven, China can reproduce it quickly and even optimise locally, but that there are hard bottlenecks both upstream and downstream. Upstream, China has limitations in advanced lithography/capacity and in the broader software ecosystem if compute becomes the binding constraint. Downstream, the US has a much more mature “to business” market, where firms will pay a clear premium for the best model because productivity gains map directly to economic value. By contrast, “business to business in China is hard”, and many productivity/coding-agent plays end up forced into overseas markets, while consumer AI has a lower “felt need” for ever-stronger intelligence—especially in China, where it’s often used as a search upgrade rather than a deep productivity tool.
The Optimists: The Role of Talent and a Complete Industrial System
The other camp, which includes academics at Tsinghua University, points to the advantages China has in some parts of the AI stack, such as a talent advantage in terms of both the quantity and quality of its higher education students, the data advantage with a significant lead in the number of AI applications deployed, and the advantage of a well-developed industrial and power infrastructure and a complete industrial chain. Another part of this argument is that as the pace of development of things such as cutting-edge semiconductors slow, Chinese firms will be able to catch up more quickly. The adherents of this argument typically hold that the current US model of enterprise-led, scale-up computing power is encountering difficulties as returns on increasing computing power—so-called scaling—provide diminishing gains.
A Third Perspective: Managing Agentic Development and the Closing Compute Gap