From American Affairs Journal, Volume X, Number 2, Summer 2026:
America has no peer competitor in artificial intelligence outside of China. Several years ago, when the trajectory of AI model development was uncertain enough to allow middle powers like the United Kingdom and France to nurse hopes of remaining competitive, this may have sounded like an overconfident pronouncement. Today, it is well-earned conventional wisdom. Competing at the frontier of AI requires coordinating talent, data, energy, and compute infrastructure at unprecedented scales—scales that only the United States and China are realistically capable of delivering on, although each in their own distinctive ways. Understanding China’s approach to developing and diffusing AI is thus of existential importance to understanding America’s relative position in the world to come.
America has been surprised by China’s AI prowess before. In January 2025, the release of DeepSeek was widely described as a Sputnik moment by policymakers and business commentators. While DeepSeek’s technical advances were overstated, the media’s reaction revealed the extent to which many in the United States had become complacent about China’s lag in AI capabilities.1 Against available evidence, too many American observers believed that China was incapable of discovering AI breakthroughs on its own, whether because of constraints on their access to advanced semiconductors, or the persistent myth that Chinese companies can only copy but not innovate. Even now, many still seem to believe that Chinese AI models will remain behind American models in perpetuity, offering lesser capabilities but at a fraction of the price. Yet offering a good enough product at ultra-low prices and thereby cornering the market on less exquisite technologies and manufacturing inputs is exactly how China became a peer competitor to the United States in the first place. In AI, we are thus primed to be surprised once again.
The pervasive indifference that characterizes America’s overconfident view of its place in the AI race stems from grading the Sino-American AI race against our own preferred rubric: frontier model benchmarks, the scale of the data center buildout, and timelines to artificial general intelligence (AGI). Rarely do we measure American performance against the categories that the Chinese themselves choose to emphasize. The party-state and various Chinese companies are clearly trying to unleash AI capabilities, and Beijing’s desire for international AI leadership is beyond dispute. But their methods and benchmarks of success are different from ours, evincing a fundamentally distinct understanding of the nature of the competition.
American readers inclined to dismiss China’s focus on open source AI diffusion and applications as a case of settling for less than the frontier should consider an alternative interpretation: that the Chinese state has made a sincere and potentially well-founded judgment about where the benefits from AI development will accrue in the medium- to long-term horizon, and its leaders are organizing the many arms of the state to support their industry ecosystem accordingly. The primary questions explored by this essay are (1) how that AI-focused industrial policy is orchestrated, and (2) what the subnational dynamics between China’s provinces, municipalities, and central government reveal about their model of AI development, for which America has no equivalent.
China’s AI Division of Labor
Before describing Chinese industrial policy for AI, it is necessary to explain two paradigmatic differences in the ways that the Chinese and American governments perceive AI development and diffusion, as well as how those different perspectives influence tangible policy outcomes.
The first is the difference between how the two countries approach hardware versus software. The United States has myriad regulatory barriers to physical infrastructure buildouts that coexist with an engrained hesitation to regulate algorithms and models. China is almost the inverse. The PRC actively regulates the algorithmic layer of AI, requiring registries of proprietary data and, in some cases, imposing “ethical committees” to oversee algorithmic usage.2 At the same time, the Chinese state aggressively organizes and subsidizes physical infrastructure and deployment: data centers, compute vouchers, industry funds, procurement mandates, start-up incubators, and more.
In contrast, the U.S. federal government maintains a relatively light-touch approach to software but is mired by legal and regulatory constraints on physical infrastructure, including power generation, transmission lines, data centers, heavy industries, and until recently, chip fabrication. It is too simple to frame this as American lawyers litigating the physical buildout of AI while skilled Chinese engineers speed ahead at constructing power generation, transmission, and roboticized factories. Nevertheless, this asymmetry is a good starting place for understanding how the respective AI strategies of the United States and China diverge.
Second, high-level discourse in China concerning AI’s technological potential is conceptualized quite differently relative to the English-language AI community. In the United States, the AI race is widely seen as a sprint to AGI, an autonomous system capable of outperforming expert humans in virtually every domain. Developing AGI in a way that benefits humanity is the explicit mission of OpenAI, for instance, reflecting the influence of early AI safety thinkers from the Effective Altruist (EA) and rationalist communities, in particular. While AI development is proceeding along a continuous spectrum, AGI is considered a particularly momentous threshold, beyond which progress rapidly accelerates toward superintelligent systems with the potential to transform every aspect of our economy and society, while giving the first company or country to achieve AGI decisive economic and military advantages.
While the discourse in America regarding the correct path for AI development is uniform—AGI is brought into existence as an emergent property of scaling LLMs and related infrastructure—the same cannot be said for China. To be sure, some Chinese researchers and thinkers do share this perspective.3 Especially at model-developing start-ups, such as Moonshot and Zhipu, where each company’s CEO has explicitly stated that his mission is to achieve AGI through LLM scaling, there is a symmetry between the American and Chinese perspectives. Another perspective is the idea of Embodied AI (EAI), a viewpoint which has been mentioned in recent high-level national documents, such as the Fifteenth Five-Year Plan. Chinese proponents of the EAI perspective tend to believe that scaling LLMs is not the path to AGI.4 Instead, EAI advocates see integrating model development with physical applications, such as robotics and self-driving cars that have self-improving capacities from interactions with the tangible world, as the best path toward AGI.5
But in practice, as well as in policy, perhaps the most influential Chinese perspective is the one that treats AI as a general purpose technology, much like electricity. This view is best articulated in Taiwanese scientist and entrepreneur Kai-Fu Lee’s popular 2018 book AI Superpowers, where he describes AI in these terms: “Often, once a fundamental breakthrough has been achieved, the center of gravity quickly shifts from a handful of elite researchers to an army of tinkerers—engineers with just enough expertise to apply the technology to different problems.”6 In this case, the power of AI will accrue not to the most sophisticated model developers but instead to those who find novel applications for the technology. This is a view that explicitly rejects imminent AGI. Lee himself describes it as decades or centuries away, if it is feasible at all.7 It should not be assumed that the Chinese government agrees with this assessment, and there are new reasons to believe that party officials are taking the prospect of superintelligence seriously.8....
....MUCH MORE
We should have a very good idea of what's possible by the first quarter of 2028.
If interested see August 7's ""Inside the Race to Make AI Build Itself"" for more on the timeline.
