DSET Democratic Governance Program Director Kai-Shen Huang recently published an op-ed in The Diplomat, analyzing how China, under U.S. bans on advanced chips and restrictions on compute, is gradually developing an artificial intelligence (AI) development pathway centered on efficiency and diffusion. The op-ed cautions that China’s “Frugal Stack” strategy could reshape the global market baseline for AI models.

Huang argues that Washington has long framed the AI race as a contest of brute force—one defined by access to cutting-edge chips, scale of compute, and capital intensity. China, by contrast, is increasingly recalibrating its trajectory. Rather than single-mindedly pursuing the most advanced technologies and frontier models, it is pivoting toward a strategy of “lean diffusion,” emphasizing algorithmic efficiency, low compute requirements, and large-scale deployment of AI across industry and the real economy.

This strategic turn was clearly on display at the AGI-Next Summit held in Beijing in early January. Multiple Chinese AI startups—including members of the so-called “Six Tigers”—openly acknowledged that they are unlikely to dominate the next paradigm of artificial general intelligence (AGI) in the near term. Instead, they are pursuing what they describe as a “poor man’s way”: under conditions of constrained compute, optimizing algorithms and infrastructure to build a leaner software stack capable of narrowing the gap with the technological frontier.

Huang notes that while U.S. technology giants continue to bet heavily on large-scale model training and dense AI compute, Chinese firms are placing greater emphasis on token efficiency and reinforcement learning with verifiable rewards. Through a “Frugal Stack” approach, they aim to systematically reduce the costs of model training and deployment.

The op-ed further argues that the significance of China’s Frugal Stack lies not only in narrowing the capability gap with the United States, but also in its strong advantages for cross-border diffusion. By offering—and increasingly shaping—an ecosystem of cheap, high-performance, and computationally frugal models, China is advancing a diffusion pathway in the digital domain analogous to a “rural encirclement” strategy, making its AI technologies easier to download, fine-tune, and integrate into existing systems and industrial workflows worldwide.

Huang cautions that as Chinese open models increasingly become default choices for developers, associated applications, toolchains, and evaluation pipelines may develop strong path dependence. This, in turn, could crowd out comparatively expensive U.S. commercial APIs that are harder to scale, slower to run, and less well aligned with local linguistic and cultural contexts—particularly in regions such as Southeast Asia, the Middle East, and Africa, where Chinese AI application models may prove more attractive.

The op-ed concludes by emphasizing that the principal risk facing the United States is not a single instance of technological lag, but a strategic over-concentration on frontier breakthroughs that underestimates diffusion itself as a source of geopolitical leverage. In the next phase of AI competition, the central question may not be who builds the most capable model, but whose stack becomes the global baseline for deployment.