On March 23, 2026, the U.S.-China Economic and Security Review Commission (USCC) published a research report, “Two Loops: How China’s Open AI Strategy Reinforces Its Industrial Dominance,” written by Ngor Luong. Its core argument fits in one sentence: China has gone all in on open-source AI, and the strategy runs through two mutually reinforcing loops — digital and physical — that steadily move the advantage of the AI race onto the terrain China already dominates: manufacturing.
The report observes that most Chinese labs publish source code and model weights, and price high-end products well below global competitors. The result: even under compute constraints, Chinese labs operate close to the frontier, keep narrowing the gap with top Western models, and contribute widely adopted architectural and training innovations. For readers used to the narrative of closed models leading while open models chase, this report offers the inverse framing.
What the Two Loops Are
The report’s analytical skeleton is a pair of circuits:
- The digital loop: open-source release → low-cost access → widespread global adoption → faster model iteration → even more adoption.
- The physical loop: embodied AI deployed in factories, robotics, and logistics generates specialized real-world data, which feeds back into model improvement.
The two loops fuel each other. Open models let AI roll out across China’s enormous manufacturing base at very low unit cost, while data from the shop floor makes the models better at the physical world. The report explicitly ties this to the “interlocking innovation flywheels” cited in the Commission’s 2025 Annual Report — except this time the driving surface is not an algorithm but an entire industrial system. And unlike a single breakthrough model, a flywheel built on deployment scale compounds quietly: every factory that adopts an open model widens the data stream feeding the next one.
The Numbers: Qwen and 100,000 Derivatives
Scale best explains the digital loop’s force. The report names Alibaba’s Qwen the largest model ecosystem on Hugging Face, with over 100,000 derivative models; press coverage of the report likewise notes that Qwen and DeepSeek effectively dominate today’s open-source landscape. The derivative count matters less as a number than as a measure of ecosystem gravity: when fine-tuning tooling, inference frameworks, and teaching materials all accrete around one model family, developers’ switching costs only move in one direction — up.
The institutional layer deserves attention too. The report notes Beijing has built supporting structures: data is formally designated a factor of production, and enterprises are allowed to carry data assets on their balance sheets. Turning data from a byproduct into bookable capital rewards the physical loop’s data collection at the accounting level.
Why Export Controls Can’t Stop the Physical Loop
The report’s sharpest argument is here. U.S. export controls act mainly on the digital loop — restricting the chips used to train frontier models. The physical loop sits almost entirely outside their reach, and open-source models keep lowering the compute threshold needed at the deployment end. That means China’s industrial-data advantage depends less and less on cutting-edge hardware. The report’s inference is direct: even fully successful chip controls may not stop the physical loop from turning.
Put differently: the controls target the capacity to train the strongest models, while China’s open-source strategy bets on the capacity to distribute good-enough models everywhere. As open models keep raising the definition of “good enough,” the controls’ effect gets diluted. For anyone used to measuring the U.S.-China AI competition by leaderboard gaps, this is a practical corrective.
Takeaways for Developers and Policy People
For developers choosing a model stack: the sheer size of the Qwen ecosystem means density of surrounding tools, derivative models, and talent — in practice that shapes fine-tuning, deployment, and hiring. Pick an open-source route, and ecosystem gravity is itself either your moat or your lock-in risk. For policy people: the report effectively argues that export controls alone cannot sustain an advantage, and that the physical side — robotics, manufacturing data, application penetration — is an equally live competitive variable. For everyone else, the report adds a geopolitical footnote to the opening leg of the 2026 AI race: the leaderboard is only half the story, and the other half is running inside factories. One caveat on weight: the USCC is an advisory commission that writes recommendations, not law — but its reports shape congressional hearings and the annual review that anchors China policy in Washington, so the framing tends to migrate into legislation.
Sources
- Two Loops: How China’s Open AI Strategy Reinforces Its Industrial Dominance — USCC
- USCC Warns China AI Data Advantage — Crypto Briefing
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.