China’s use of open‑source AI threatens the US lead in AI development, US Commission warns

A new analysis from the US‑China Economic and Security Review Commission argues that Beijing’s open‑source AI push is creating a self‑reinforcing edge—one that current US export controls aren’t built to blunt.

Chinese AI models are not just catching up—they’re proliferating. Systems like Qwen and DeepSeek now dominate download charts on Hugging Face, signaling a rapid, global diffusion of Chinese open‑source AI. But the headline trend isn’t just about model popularity. According to a new Commission paper, China’s playbook is converting open models into factory‑floor gains, spinning up proprietary industrial data at a pace the United States is ill‑equipped to counter.

A two‑loop strategy the US didn’t plan for

The paper—titled Two Loops: How China’s Open AI Strategy Reinforces Its Industrial Dominance—frames Beijing’s approach as a twin engine. The first loop is digital: models, code, and compute. The second is physical: deployment of AI across manufacturing, logistics, and other industrial systems. US export controls, the authors argue, aim squarely at the first loop by restricting access to advanced chips used for training frontier models. But they largely miss the second loop, where real‑world deployment generates exclusive data that compounds over time.

US export controls primarily target the digital loop—cutting off high‑end training hardware—yet are not well suited to addressing the physical loop of deployment‑driven data creation across China’s manufacturing base, the Commission warns.

As open‑source models improve and become more efficient, they require less compute to deploy effectively. That shift lowers the hardware barrier for integrating AI into factories and supply chains. The result: Chinese firms can scale AI‑powered workflows, capture operational data at industrial scale, and feed that data back into better models—without needing the latest, restricted chips.

Open source as a diffusion engine

Chinese labs and companies have leaned into permissive licensing and aggressive pricing, creating what the Commission calls a “global diffusion engine.” Accessible models accelerate adoption among developers and enterprises worldwide, often undercutting proprietary US offerings on cost and flexibility. Every new deployment, especially in manufacturing contexts, becomes a sensor—collecting task‑specific, high‑value data that is difficult to replicate or purchase on the open market.

This creates a structural advantage: while anyone can download a base model, only the operator running that model across thousands of machines, assembly lines, and workflows accumulates the fine‑grained data necessary to optimize it for real‑world performance. The paper argues that China is systematically capturing this advantage through scale, policy alignment, and ecosystem coordination.

Why chips aren’t the chokepoint anymore

Export controls made it harder to train the very largest models, but the Commission suggests that’s no longer the only—or even the primary—bottleneck for competitive AI. Today’s open models are increasingly capable out of the box, and can be fine‑tuned on modest hardware for targeted tasks. In that environment, compute scarcity matters less than access to proprietary, high‑quality data.

As open models reduce the compute required for effective deployment, China’s ability to generate proprietary industrial data at pace and scale becomes increasingly independent of access to cutting‑edge hardware, the paper says.

In other words, the locus of competition is shifting from training the biggest models to deploying the most models in the most places—and learning from the resulting data. That’s a contest where industrial capacity, integration speed, and local data feedback loops may outweigh frontier parameter counts.

Policy blind spots and a moving target

The Commission’s core warning is that current US policy frameworks don’t address the physical loop. Restrictions on advanced accelerators target the training phase but leave the deployment ecosystem largely untouched. Meanwhile, permissively licensed Chinese models spread quickly, lowering barriers for global adoption and intensifying the data flywheel for those with the most machines, sensors, and processes to instrument.

The paper also underscores a pricing dynamic: open and low‑cost models from Chinese labs pressure US incumbents who rely on closed systems and enterprise subscriptions. As developers default to what’s cheapest and easiest to integrate, market share and influence shift—further entrenching the data advantage of the most widely deployed stacks.

What’s at stake for US competitiveness

If the Commission’s analysis holds, US leadership in AI won’t be decided solely in research labs or data centers. It will hinge on whether American industry can embed AI across production, logistics, and maintenance at comparable scale—and whether policy can evolve to recognize that deployment data is the new moat.

The implications cut across:

• Industrial policy: Incentivizing AI deployment and data generation in domestic manufacturing, not just subsidizing chips or cloud capacity.
• Open‑source strategy: Supporting competitive, permissively licensed US models that can match the diffusion speed of Chinese alternatives.
• Standards and ecosystems: Building interoperable stacks that make it simple for factories and suppliers to adopt AI—and own or securely share the resulting data.
• International coordination: Working with allies to align on data governance and avoid being outcompeted in the physical loop by subsidized, rapidly spreading open models.

The Commission’s message is blunt: the US focused on the chip bottleneck while China built a deployment flywheel. With Qwen, DeepSeek, and other open models sprinting up the download charts, the question isn’t whether the technology will spread—it’s who will capture the compounding returns from the data it generates on the factory floor.