The U.S.-China Economic and Security Review Commission published a report on March 23 warning that China’s dominance in open-source artificial intelligence is creating a competitive advantage that US export restrictions cannot contain. The commission found that approximately 80 percent of American AI startups now rely on Chinese open-source AI technology, and that Chinese-origin models accounted for 41 percent of all Hugging Face downloads between February 2025 and February 2026 — surpassing US-origin models at 36.5 percent.
The report describes a self-reinforcing cycle: Chinese companies release high-quality open-source models, Western developers adopt them, community contributions improve the models further, and the resulting ecosystem attracts more users and investment back to Chinese model families. This dynamic operates independently of export controls on chips, which target hardware but cannot restrict the distribution of model weights that have already been trained and released.
The commission’s findings challenge the assumption underlying US chip export restrictions — that limiting China’s access to advanced semiconductors would slow its AI progress. While hardware restrictions may constrain training of the largest frontier models, Chinese labs have compensated through architectural efficiency, mixture-of-experts designs that achieve strong performance with fewer active parameters, and aggressive open-source distribution that builds ecosystem dominance regardless of training compute.
DeepSeek, Qwen (Alibaba), and other Chinese model families have demonstrated that competitive AI capability can be achieved at lower compute budgets than Western labs assumed necessary. The USCC report notes that Washington approved exports of Nvidia’s second-most advanced chip to China in December 2025, acknowledging that complete hardware restriction was neither practical nor effective.
The report recommends that the US government invest in domestic open-source AI development and reconsider whether export controls are the right tool for maintaining AI competitiveness. The alternative — competing on model quality and ecosystem value rather than restricting hardware access — would represent a significant policy shift from the approach adopted since 2022.
