Amidst fierce U.S.-China tech competition, the idiom “four ounces deflecting a thousand pounds” (四兩撥千斤), a concept rooted in Chinese taichi, has become a recurring fixture in Chinese official documents, state media commentary, and public statements by AI industry leaders. The phrase has become a shorthand Beijing increasingly deploys to frame its strategic response to U.S. export controls.
DSET has released its latest policy report, “Four Ounces Deflecting a Thousand Pounds: Analyzing Chinese Elite Perspectives on Winning the U.S.-China AI Competition,” drawing on open-source intelligence (OSINT) methodology to systematically analyze Chinese public documents and AI elite discourse. Taking China’s own vantage point, the report examines how, under the “thousand-pound” pressure of U.S. restrictions on advanced chips and equipment, Chinese actors are deploying a “four-ounce” strategy across six layers of the AI stack—market applications, AI models, compute, infrastructure, talent, and financing—leveraging existing advantages in energy capacity, domestic market scale, industrial supply chains, open-source ecosystems, and state-backed capital to mount a sustained, full-spectrum challenge to the U.S.-led AI ecosystem.
The report was presented by co-authors, Nathanael Cheng and Emory Tsai-yi Wang, and featured commentary from David Lin, Senior Advisor for Tech Leadership at the Special Competitive Studies Project (SCSP), and Wendy Chang, Senior Analyst at MERICS, who shared expert insights and perspectives on the findings.
The primary research question was how China’s AI elites understand and enact the strategy of “four ounces deflecting a thousand pounds”—leveraging advantages across the AI stack to respond to U.S. export controls. DSET focused primarily on leaders in China’s AI industry, including tech giant and startup executives and scientists, business leaders, academics, and infrastructure operators.
Cheng and Wang argued that high-end training compute is the weakest link in advancing China’s frugal AI stack strategy. To address this constraint, China is simultaneously accelerating market application diffusion and driving down model costs, while investing in talent, financing, and infrastructure to compensate for the shortfall in advanced chips. While Chinese AI leaders widely acknowledge challenges across all six layers of the stack—market applications, models, compute, infrastructure, talent, and financing—this does not imply that China’s competitiveness can be underestimated. China’s four-ounce strategy has absorbed external pressure across multiple layers, converting constraints into domestic innovation imperatives. For the United States and its allies, competing effectively requires an economic security strategy that addresses the full AI ecosystem—not only leading-edge chips and frontier models, but the broader economic and industrial foundations that determine how AI capabilities are developed, deployed, and scaled.
The presenters thanked co-authors Jin-Chian Seer and Jeremy Chih-Cheng Chang for their invaluable contributions.
Wendy Chang of MERICS noted that China is charting a distinct AI development path. Unlike the Silicon Valley model—characterized by high-cost technological breakthroughs followed by enterprise adoption—China prioritizes rapid diffusion through low-cost solutions. This approach is exemplified by the rapid adoption of the open-source tool OpenClaw, which achieved record uptake within weeks through coordinated support from both industry and government.
She further argued that this trajectory has structural roots: China’s system excels at scaling and replicating successful models rather than generating original innovation; its strong manufacturing base provides advantages in physical AI production; and constraints on compute force alternative technological pathways. By aggressively promoting applications and accelerating diffusion, China is seeking to carve out its own path in the global AI competition.
David Lin of SCSP offered three additional observations. First, China’s continued dependence on the U.S. software ecosystem—particularly Nvidia’s CUDA parallel computing framework—poses significant structural constraints. Second, Beijing’s “anti-involution” policy has a dual character: domestically, it encourages coordination and resource consolidation among firms; externally, it pursues multiple avenues for technological advancement, including chip smuggling, model distillation, and early access to frontier models. Third, while China continues to expand investment in basic research, foundational scientific capabilities remain a key weakness.
In the Q&A session, audience members asked about Taiwan’s role in U.S.-China competition. Cheng and Wang highlighted the potential of “Physical AI” and the robotics sector. Taiwan’s export-oriented manufacturing base positions it as a critical partner, with firms such as Foxconn already collaborating closely with Nvidia to optimize production through digital twins and software integration. The authors suggested that the United States and Taiwan should jointly develop a “non-red supply chain” in robotics and physical AI, leveraging Taiwan’s manufacturing strengths and U.S. software capabilities.
Responding to questions on China’s semiconductor self-sufficiency, the authors noted that while China is advancing domestic ecosystems such as Huawei’s Ascend series, limitations in training large-scale models persist. As a result, many Chinese AI firms continue to rely heavily on Nvidia chips, often accessing them through transshipment or remote use of data centers in third countries such as Malaysia—highlighting China’s ongoing shortage of advanced compute resources.
