Roughly 80% of US AI startups now build on Chinese open-source models. That number, from a report published today by the U.S.-China Economic and Security Review Commission (USCC), frames a problem that chip export bans were supposed to prevent.
The report, titled "Two Loops: How China's Open AI Strategy Reinforces Its Industrial Dominance" and authored by Ngor Luong, argues that China has built a flywheel the US isn't equipped to slow down. The core idea is simple: Chinese labs like Alibaba, DeepSeek, and others release powerful models for free. Developers worldwide adopt them. That adoption generates feedback, which improves the next model, which drives more adoption.
The Numbers Are Hard to Dismiss
Alibaba's Qwen model family now has over 100,000 derivative models on Hugging Face, the largest AI model-sharing platform. Qwen has surpassed Meta's Llama in cumulative global downloads. DeepSeek's R1 reasoning model briefly overtook ChatGPT as the most-downloaded app on the US App Store after launch. Chinese models dominate usage rankings on both Hugging Face and OpenRouter.
These aren't niche research projects. They're production models that startups and businesses are shipping products on top of, largely because they're free or dramatically cheaper than Western alternatives.
The "Two Loops" Problem
The report's central framework splits China's advantage into two reinforcing cycles.
The first is the digital loop: open-source model releases drive global adoption, which drives iteration, which drives more adoption. This is the one US policymakers have tried to counter with chip export restrictions since 2022. The logic was straightforward - cut off access to the best GPUs, and you cut off the ability to train competitive models.
That logic has not held up. As the report puts it, "This open ecosystem enables China to innovate close to the frontier despite significant compute constraints."
The second is the physical loop, and this is where the report gets genuinely concerning. China is deploying AI across manufacturing, robotics, logistics, and autonomous driving at a pace that generates enormous amounts of real-world industrial data. That data feeds back into model training, creating advantages that no amount of chip restrictions can touch.
USCC Vice Chair Michael Kuiken told Reuters there is "a deployment gap in the embodied AI space between the U.S. and China" that "over time compounds itself." As AI shifts from chatbots toward agents that interact with the physical world - humanoid robots, factory automation, self-driving vehicles - the country with more deployment data wins.
What This Means for People Using These Tools
The practical tension here is real. If you're a startup founder picking a base model, Qwen and DeepSeek are genuinely good and genuinely free. The security concerns the report raises - potential political bias baked into Chinese-trained models, data handling questions, dependency on a geopolitical rival's infrastructure - are abstract compared to the very concrete reality of your runway.
Western research organizations have flagged risks of over-reliance on Chinese open-source AI, but adoption keeps climbing because the economics are overwhelming. A free model that's 95% as good as a paid one wins the spreadsheet argument every time.
The USCC is an advisory body, not a policymaking one. This report doesn't change anything by itself. But it signals that the conversation in Washington is shifting from "restrict chips" to "we need a broader strategy," which could eventually mean new rules around which open-source models US companies can deploy in sensitive applications. For now, the open-source genie is thoroughly out of the bottle.