The dominant frame in U.S. policy debates over AI competition has centered on frontier model capability and access to advanced chips. A March 2026 research paper from the U.S.-China Economic and Security Review Commission challenges that framing, arguing that the most consequential dynamics are running through a different mechanism entirely.
The paper, titled “Two Loops: How China’s Open AI Strategy Reinforces Its Industrial Dominance,” argues that China’s AI ecosystem operates through two reinforcing feedback loops. The first is digital: Chinese labs release model weights or code at low cost, enabling rapid adoption and derivative development that in turn generates data and iteration. The second is physical: open models enable low-cost AI deployment across factories, logistics networks, and robotics systems, generating real-world operational data that improves models and enables further deployment.
“China has opted to go all in on an open-source approach to AI,” the USCC working group wrote. “Open model proliferation creates alternative pathways to AI leadership.”
Scale in the Developer Ecosystem
The report provides concrete measures of the first loop’s reach. Alibaba’s Qwen model family had accumulated more than 100,000 derivative models on Hugging Face as of publication, more than any Western counterpart, including Meta’s Llama. The number of Chinese open models tracked by Epoch AI expanded from 32 in 2022 to 337 in 2025. During the month of November 8 to December 8, 2025, Chinese models represented seven of the top ten most downloaded large-scale models on the platform.
USCC distinguishes between open-source models, which release code, training data, and weights, and open-weight models, which release only the weights. Many Chinese models described as open are open-weight only. The distinction matters for understanding derivative ecosystems: fine-tuning and quantization are accessible with open weights, but full reproducibility requires the training code and data. The Qwen derivative count reflects the breadth of the open-weight ecosystem rather than fully open-source release.
The Physical Loop and the Limits of Chip Controls
The second loop is where the USCC report makes its most pointed argument for U.S. policymakers. China’s manufacturing base, the report argues, is not incidental to its AI strategy. It is the substrate through which deployment-driven data accumulation happens at a scale that cannot be replicated through web scraping or synthetic generation. The report cites a Guangdong intelligent-factory deployment where AI-assisted quality inspection reduced equipment repair rates by 20 percent and saved more than 1 million yuan annually in a single plant. AgiBot and Fourier, two leading Chinese robotics firms, have released open training datasets, deepening the connection between model ecosystems and physical deployment.
The policy implication is explicit: U.S. export controls are designed primarily to constrain the digital loop, by restricting access to advanced training chips. They are structurally less suited to the physical loop, where the relevant assets are operational data generated by factories, logistics networks, and robots that are already deployed. “China’s open AI model strategy and its manufacturing dominance are mutually reinforcing,” the report concluded.
An Ecosystem, Not Just a Lab Race
The two-loop framework also reframes the competitive unit of analysis. In benchmark comparisons, the relevant actor is a lab and the relevant output is a model score. In the USCC framework, the relevant actor is an ecosystem: model developers, open-weight distributors, hardware vendors, factory operators, robotics firms, and industrial data aggregators, all participating in feedback loops that none controls individually. This matters for policy because interventions targeted at individual companies or model families may disrupt one part of the ecosystem without addressing the underlying loop dynamics.
The paper distinguishes open-weight models from fully open-source models and notes that many models described as open are open-weight only. Full open-source release, including training code and data, enables much deeper derivative development than open-weight release alone. Chinese labs have generally released weights while retaining training code, which still enables broad fine-tuning and derivative ecosystems while preserving some proprietary differentiation. Understanding that distinction is necessary for accurately assessing both the reach of China’s open model ecosystem and its actual openness.
The USCC paper is an analytical assessment from a U.S. government advisory body, not a neutral market study, and its conclusions should be read accordingly. But its two-loop framework offers a more granular account of where Chinese AI competition is actually occurring than benchmark comparisons alone provide, and raises questions about whether current policy instruments are matched to the contest they are meant to address. Earlier reporting on China’s “good-enough” AI chips and models powering the Global South documented the same cost-and-deployment logic at work internationally.