Via nbcnews.com
A new report argues that American chip export controls miss the bigger picture: China's factories are generating the specialized data that makes AI actually useful.
The US-China Economic and Security Review Commission has published a report making an uncomfortable argument for Washington policymakers: the chip export controls designed to slow China’s AI progress may be targeting the wrong bottleneck entirely.
The March 2026 report, titled “Two Loops: How China’s Open AI Strategy Reinforces Its Industrial Dominance,” contends that China’s real AI advantage isn’t processing power. It’s data, specifically the kind of specialized, real-world data that pours out of the world’s largest manufacturing base every single day.
The two-loop problem #
The report, authored by Senior Policy Analyst Ngor Luong, introduces a framework built around two reinforcing feedback loops that compound China’s AI capabilities in ways that Washington’s current policy toolkit doesn’t address.
The first is what the report calls a “digital loop.” China’s open-source AI models, most notably Alibaba’s Qwen series, have generated more than 100,000 derivative models on Hugging Face alone. Each derivative creates new use cases, generates new data, and feeds improvements back into the base models.
The second is the “physical loop.” China’s enormous manufacturing sector, its robotics deployments, and its sprawling industrial infrastructure generate vast quantities of specialized operational data. This isn’t the kind of data you scrape from the internet. It’s sensor readings from factory floors, performance metrics from autonomous systems, quality control data from production lines.
These two loops reinforce each other. Open-source models get deployed across Chinese industry. Industry generates proprietary data. That data improves the models. Better models attract more industrial deployment.
Why chip controls aren’t enough #
US export controls on advanced AI training chips began in 2022, with the explicit goal of limiting China’s ability to develop frontier AI models. The USCC report suggests this strategy has a significant blind spot.
Training chips are one constraint. But once a model exists, even a moderately capable one, deploying it across millions of industrial endpoints generates the kind of feedback data that can close capability gaps over time. The controls target the supply side of AI development while largely ignoring the demand side, where China’s advantages are structural and growing.
The Chinese government has reinforced this dynamic by officially categorizing data as a “factor of production,” a policy designation that puts it alongside land, labor, and capital in economic planning. That’s not just symbolic language. It signals coordinated state support for industrial data generation and collection infrastructure.
The open-source wildcard #
Alibaba’s Qwen series exemplifies China’s open-source approach. By releasing capable models openly, Chinese developers have created an ecosystem where thousands of companies and researchers build on top of their work. The 100,000-plus derivatives on Hugging Face represent a distributed R&D network that no single company could replicate internally.
Washington’s response and what comes next #
The report arrives amid an already intensifying policy response. Between July and August 2026, the US heightened scrutiny on Chinese AI models and cautioned allies against engaging with competing Chinese AI initiatives. The diplomatic pressure campaign suggests Washington is beginning to recognize that chip controls alone represent an incomplete strategy.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our