Summary: The United States-China Economic and Security Review Commission published Ngor Luong’s “Two Loops: How China’s Open AI Strategy Reinforces Its Industrial Dominance” on March 23, 2026. The study argues that China’s open-source AI strategy compounds through digital and physical loops. Qwen’s more than 100,000 derivatives show ecosystem gravity: clustered tooling and learning material make the cost of switching away rise.
“Two Loops: How China’s Open AI Strategy Reinforces Its Industrial Dominance” is the title of a research report the U.S.-China Economic and Security Review Commission (USCC) put out on March 23, 2026, authored by Ngor Luong. Compress the whole argument into a single line and it reads: China has bet outright on open-source AI, and the bet pays out through two loops — a digital one and a physical one — that reinforce each other while steadily shifting the AI race’s advantage onto ground China already owns: manufacturing.
As the report tells it, most Chinese labs release their source code and model weights and sell high-end products at prices well under what international rivals charge. The consequence: with training compute squeezed, those labs still sit close to the frontier, keep closing the distance to the top Western models, and have turned out architectural and training innovations that the wider field has adopted at scale. Readers who know the story as closed models out front and open models catching up will find the picture flipped here.
The Two Loops, Explained
What the report builds its analysis on is a pair of circuits:
- The digital loop: an open-source release → cheap access → adoption spreading worldwide → quicker model iteration → adoption climbing further.
- The physical loop: AI embedded in factories, robotics, and logistics produces real-world data as it runs, and that stream loops back into better models.
Each loop stokes the other. Open models make it possible to spread AI across China’s vast manufacturing base at rock-bottom unit cost, and the data coming off those shop floors teaches the models how the physical world actually behaves. The report itself links the pattern to the “interlocking innovation flywheels” named in the Commission’s 2025 Annual Report — the difference being that the driving surface this time is an entire industrial system rather than an algorithm. And where a single breakthrough model is a one-off event, a flywheel resting on deployment scale compounds quietly: every factory that takes on an open model widens the data stream feeding the next one.
Scale Check: Qwen and Its 100,000-Plus Derivatives
Nothing conveys the digital loop’s punch like raw scale. The report singles out Alibaba’s Qwen as the largest model ecosystem on Hugging Face, its derivative models now numbering past 100,000; press coverage of the report likewise observes that Qwen together with DeepSeek effectively owns today’s open-source map. The derivative count matters less as a figure than as a gauge of ecosystem gravity: once fine-tuning tooling, inference frameworks, and learning material all cluster around one model family, what a developer pays to switch moves in one direction only — up.
There is an institutional layer worth flagging too. Beijing, the report notes, has put supporting arrangements in place: data now holds formal status as a factor of production, and companies are permitted to book data assets on their balance sheets. When a byproduct becomes capital you can put on the books, the physical loop’s data collection gets rewarded straight through the accounting system.
Export Controls Miss the Physical Loop
The sharpest reasoning in the report lives here. U.S. export controls bear mainly on the digital loop — the chips used to train frontier models are what they restrict. The physical loop falls almost entirely outside that perimeter, and open-source models keep pushing down the compute required on the deployment side, so China’s industrial-data advantage leans less and less on cutting-edge hardware. The report draws a blunt conclusion: chip controls could succeed completely and the physical loop still might not stop turning.
Restated: what the controls aim at is the ability to train the strongest models, while the open-source wager China has placed is on the ability to spread good-enough models everywhere. Because open models keep lifting where “good enough” sits, whatever bite the controls have keeps wearing away. For observers accustomed to scoring the U.S.-China AI contest by leaderboard gaps, it lands as a practical corrective.
What Developers and Policy Readers Should Take Away
If you are a developer picking a model stack: the scale of the Qwen ecosystem translates into how dense the surrounding tools, derivative models, and talent pool are, and that in practice colors fine-tuning, deployment, and hiring. Commit to the open-source route, and ecosystem gravity doubles as either your moat or your lock-in risk. If you work in policy: the report amounts to a claim that export controls by themselves cannot sustain an advantage, and that robotics, manufacturing data, and application penetration on the physical side are competitive variables just as live. For everyone else, the report supplies the geopolitical footnote to the 2026 AI race’s opening moves: half the story is the leaderboard, and the other half runs on factory floors. One note on how much weight to give it: the USCC is an advisory commission that issues recommendations rather than laws — still, its reports feed congressional hearings and the annual review that anchors China policy in Washington, so its framing has a habit of seeping into legislation.