While the United States has spent years constructing an elaborate regulatory fortress around advanced semiconductors, a parallel and largely unregulated trade has been quietly flourishing: the sale of premium AI training datasets from Silicon Valley’s top data labeling companies directly to China’s most powerful AI laboratories. A detailed investigation by Forbes reveals that the same American startups supplying training data to OpenAI, Anthropic, and even US government agencies are simultaneously selling that same data to Chinese labs building cheaper models that directly rival their American customers.
The pipeline runs to the highest levels of China’s technology sector. Tencent, which has previously been designated by the US government as associated with the Chinese military, circulated detailed requests for training data spanning finance, cybersecurity, and AI self-improvement research to prospective American vendors. Internal communications reviewed by Forbes show that Ant Group, Alibaba, and ByteDance have all established procurement relationships with major US data platforms, including Surge AI, Mercor, AfterQuery, and Turing. Collectively, the top six Chinese AI labs spend approximately $500 million per year with American data labeling companies, according to two data labeling entrepreneurs who received market size estimates from Tencent and ByteDance executives.
The strategic significance of this trade goes far beyond simple data purchases. What Chinese labs are acquiring is not just raw text or images, they are buying the proprietary “knowledge pipelines” developed during custom engagements for OpenAI and Anthropic. These pipelines include PhD expert networks, quality control algorithms, and post-training rubrics that teach models to handle complex professional tasks like financial modeling, legal reasoning, and coding. By purchasing off-the-shelf datasets built on these same frameworks, Chinese labs can bypass months of expensive trial-and-error and directly benefit from the hard-won expertise that powers Silicon Valley’s best models.
The Regulatory Gap Washington Has Not Closed
The contrast with chip export controls is stark. Washington has invested enormous political and administrative capital in restricting China’s access to advanced semiconductors, creating a complex and ever-tightening web of export controls, entity list designations, and allied coordination mechanisms. Yet the human expertise encoded in AI training data crosses borders with almost no scrutiny at all. There is currently no export control framework that covers AI training datasets, no licensing requirement for selling post-training rubrics to foreign adversaries, and no government body tasked with monitoring this trade.
This regulatory gap has made Silicon Valley’s top data platforms extraordinarily valuable. Mercor, founded in 2023, is eyeing a valuation of $20 billion, while Surge AI is reportedly worth at least $25 billion. AfterQuery grew from $100 million to hundreds of millions in annual recurring revenue in just three months. While the majority of their revenue still comes from established American clients, Chinese AI labs represent an irresistible secondary revenue stream. Mercor’s Q2 revenue from Chinese AI labs was 2% of total revenue, a small share of a $2 billion annualized business, but still tens of millions of dollars flowing to a company that also holds US government contracts.
The business logic is compelling for data vendors. Off-the-shelf datasets carry the highest profit margins because they can be resold to multiple labs. Chinese AI labs explicitly tell US data labelers that they want to buy all the data that US labs have already purchased. As AI data consultant Sean Cai told Forbes: “If you actually study the shape of the data industry in China and then the entire data supply chain in general, you’ll realize that yes, a lot of the same US data that powers US model advancement is getting sold to Chinese labs as well.”
A Third Fast-Track to AI Parity
The Forbes investigation frames the data trade as the third and most legally uncomplicated fast-track available to Chinese AI labs seeking to close the gap with American rivals. The first is poaching researchers, a practice that has accelerated dramatically, as Chinese tech giants have been aggressively recruiting AI talent from US universities and labs. The second is distillation, harvesting outputs from ChatGPT or Claude to train Chinese models, which, while not illegal, explicitly violates OpenAI’s and Anthropic’s terms of service and was flagged by the Trump administration’s science and technology advisor Michael Kratsios as a national security threat.
The third path, buying training data, is easier and comes without the regulatory headaches. It is entirely legal, commercially straightforward, and produces results that are functionally equivalent to the other two approaches. Nathan Lambert, a former research scientist at the Allen Institute for AI, told Forbes: “After the compute needed to train the models, good quality data is the most important thing. As we try to get models to move to new, more challenging domains, getting good data is the highest-leverage item that can make the model become viable.”
Some in the industry have begun raising alarms. Ali Ansari, 25-year-old CEO of Micro1, posted publicly: “Some human data companies work with foreign adversaries, and the results show today in Kimi K3. We believe it’s shameful to claim American AI dominance desires while selling millions worth of data to countries that we are in adversarial competition with.” Scale AI previously flirted with a ByteDance contract before scrapping it in 2024 over national security concerns.
The Open-Source Complication
Defenders of the practice argue that restricting data sales would cause collateral damage to the global open-source AI ecosystem. “One cannot easily restrict exports of data or AI in a way that does not materially hurt American open source,” said Cai. “It gives an unfair duopoly to OpenAI and Anthropic, and that has its own problems.” This tension, between national security imperatives and the open-source principles that have driven much of AI’s rapid progress, is likely to become a central fault line in Washington’s AI policy debates over the coming months.
The Forbes investigation arrives at a moment when the US-China AI competition is intensifying on every front. Chinese models have already captured a record share of US corporate token usage, and the performance gap between leading Chinese and American models has narrowed to the point where many enterprise users cannot reliably distinguish between them. If the $500 million data pipeline continues to operate unchecked, it will represent one of the most consequential regulatory blind spots in the history of the US-China technology competition, a gap that chips controls alone cannot close.