China’s open-source AI strategy is creating a self-reinforcing competitive advantage that U.S. export controls were not designed to counter, according to a warning from the U.S.-China Economic and Security Review Commission.
The document, titled Two Loops: How China’s Open AI StrategyReinforces Its Industrial Dominance, argues that Chinese laboratories have created a global diffusion engine through permissive licensing and pricing that undermines U.S. rivals, and that its widespread adoption is generating advantages in industrial data that no current regulatory framework addresses.
More and more companies around the world are using China’s AI
The magnitude of that adoption is reflected in the figures. The Qwen model family from Alibaba is now the largest ecosystem on Hugging Face, with over 100,000 derived models, according to the article. Between November and December 2025, seven of the ten most downloaded models on the platform came from Chinese labs, including derivatives from DeepSeek, whose rapid rise has already altered companies’ AI acquisition forecasts.
The article also adds that commercial penetration extends deeply into American companies. A partner at the venture capital firm Andreessen Horowitz estimated that approximately 80% of American startups use Chinese base models to create derivative applications, the article notes.
Regarding prices, the Kimi K2.5 from Moonshot AI costs four times less than OpenAI’s GPT-5.2, while matching it in capability scores, according to the article, which cites benchmark data from Artificial Analysis.
“Companies are no longer making a clear and deliberate decision about which AI model to adopt,” said Sanchit Vir Gogia, chief analyst at Greyhound Research.
The traditional management of third-party risks was never designed for this, adds Gogia. “It does not track the lineage of models, fine-tuning chains, training data inheritance, or runtime routing behavior.”
Fear for data security and malicious actions
The security risks associated with that invisible exposure are documented. The U.S. National Institute of Standards and Technology evaluated the DeepSeek models in September 2025 and found that agents based on the safest DeepSeek model were, on average, 12 times more likely than cutting-edge U.S. models to follow malicious instructions, according to the article.
In simulated tests, the kidnapped agents sent phishing emails, downloaded malware, and stole users’ login credentials.
The article also noted that Chinese models are subject to restrictions on political content, and that companies that route their workloads through providers linked to China face jurisdictional risks regarding data, a concern that has raised regulatory scrutiny across Europe and South Korea.
Gogia pointed out that companies should treat AI as a software supply chain: “A bill of materials for the model should reflect the origin of the base model, the history of derivatives, the datasets used, and the geographical location of hosting, and it should be continuously updated, as these systems change in production without triggering traditional acquisition processes.”
The deployment in factories widens the gap
The main concern of the commission is that China’s advantage operates through two distinct cycles.
The first is digital: open models drive adoption, adoption drives iteration, and iteration produces more capable models. The second traverses the physical economy: AI deployed in factories, logistics networks, and the robotic sector of China generates real-world data that feeds back into the improvement of the models.
Beijing formalized this in 2020, designating data as the fifth factor of production, and became the first country to allow companies to include data assets in their balance sheets, the commission noted in the document.
The document stated that the most important models for business use are not the large advanced language models, but rather the small models specific to concrete tasks. In this type of models, China’s open ecosystem is already ahead.
A research document from Nvidia cited by the commission revealed that small models handle most of the operational subtasks in agentive AI systems at a cost between 10 and 30 times lower than that of cutting-edge alternatives.
“The real question is no longer which flagship model a provider uses,” Gogia stated. “The real question is which models are actually performing tasks in production: where they come from, how often they are updated or replaced, and how they are managed.”
And in the U.S. they don’t know how to respond
The commission acknowledged the initial steps of American companies but warned that the overall response remains uncoordinated. OpenAI launched its first open-weight models since GPT-2 in August 2025, followed by Nvidia’s Nemotron 3 in March 2026.
However, according to reports, Meta is preparing to shift its next-generation model to a closed approach, only through API.
“If it continues,” the document states, “Meta’s departure from openness would leave the United States without a developer of cutting-edge models that strengthens its open AI ecosystem, precisely at a time when open development backed by the Chinese state is accelerating.”
