America is locked in a winner-take-all artificial intelligence race with Communist China, and the winner stands to gain decisive advantages in economic competitiveness, technological leadership, and military power. For years, Washington has worked to choke off China’s access to advanced AI chips. But chips are only one piece of the puzzle.
Building cutting-edge AI systems depends on three core elements: compute (chips and data centers), algorithms (the models developed by companies like OpenAI and Anthropic), and data, the specialized training information that teaches AI systems how to perform complex tasks. A growing body of reporting focuses on that third element. So then why are American companies supplying it to both U.S. and Chinese customers?
Why Specialized Data Matters
Early AI systems trained primarily on publicly available internet content, such as books, websites, code repositories, and documents. Most of that readily accessible data has already been exhausted. To push performance further in demanding domains like coding, medicine, law, accounting, cybersecurity, military planning, and scientific research, developers now need high-quality, purpose-built datasets created by human experts.
Four American firms have emerged as the key providers of this specialized material: Surge AI, Mercor, AfterQuery, and Turing. These companies recruit subject matter experts like coders, physicians, lawyers, engineers, and scientists to convert their knowledge into structured training data. They design tasks, establish grading rubrics, define correct reasoning pathways, and implement quality controls. In effect, they refine raw American expertise into the “last-mile” data that increasingly separates the world’s leading AI models from the rest.
The Dual-Use Problem
According to reporting by Forbes journalist Anna Tong, the very same companies that supply data to OpenAI, Anthropic, and in some cases U.S. government entities are also selling to Communist China’s leading AI laboratories. Chinese labs have pursued every available path to close the gap with American systems. They’re recruiting talent, distilling (or stealing) outputs from U.S. models, and now simply purchasing the identical specialized training datasets already developed for American customers.
Forbes reported that China’s top six AI labs are collectively spending roughly $500 million per year with these American data providers. Chinese buyers have sought the exact datasets U.S. labs paid to develop, allowing the CCP to skip months of costly experimentation on task design, evaluation methods, and quality systems that Americans already paid for.
The customers are some of the largest companies in China. They include Tencent, Alibaba, Ant Group, ByteDance (the parent company of TikTok), and Moonshot. Tencent has been formally designated by the Pentagon as linked to the Chinese military and the Chinese Communist Party.
Take Surge AI. The firm has worked with Anthropic as well as the U.S. Army and Air Force, according to Forbes. The same firm has supplied data to Tencent. Forbes reviewed a message from a Tencent executive stating the company relies on Surge for data. Surge CEO Edwin Chen has personally traveled to China to meet with Chinese AI executives.
Then there is Mercor, which maintains relationships with Tencent while holding a U.S. federal contract. Chinese labs accounted for about 2% of its second-quarter revenue, its annualized revenue surpassed $2 billion in June, and the company went so far as to hire a dedicated China engagement manager. Mercor has suffered from allegations that suspected North Korean operatives used stolen identities to work as contractors, including on Anthropic-related work, from late 2024 into 2026. A March 2026 breach exposed internal and contractor data, reports of internal embezzlement on an Anthropic-related account, labor complaints, and a lawsuit from rival Scale AI alleging trade-secret theft. Roughly 91% of Mercor’s first-half 2026 revenue came from a small group of foundation-model companies.
AfterQuery has been linked through texts, audio, and purchase agreements to Ant Group and Alibaba, with at least $50 million in reported recurring revenue from Chinese AI labs. CEO Spencer Mateega has described data as “the final frontier.” His company even publicly celebrated Chinese lab Moonshot’s Kimi model for topping one of its benchmarks, with Mateega gushing that “The Moonshot team is cracked.”
Turing, per Forbes, has internal documents suggesting work with ByteDance. Contrast that with Scale AI, which declined a similar ByteDance opportunity back in 2024 over national security concerns. These companies can refuse this business, but they choose not to.
Smuggle A Chip, Get Indicted. Sell The Data, Get Rich.
The disparity in how Washington treats hardware versus data is stark. In a recent case, four men, including two American citizens and two Chinese nationals, were arrested in connection with an alleged scheme to ship controlled Nvidia GPUs, including advanced H100 processors, to China. Prosecutors described fake contracts, false paperwork, front companies, routing through third countries, and more than $3.89 million wired from China. The Justice Department called it a deliberate effort to circumvent export controls.
Transferring restricted hardware to China can land you in federal prison. Selling China the sophisticated human-derived training expertise and pipelines that make its models competitive faces a far weaker legal framework. That is a loophole big enough to drive the CCP’s entire AI catch-up strategy through, allowing them to surpass the US with AI dominance.
Some in Washington see the problem. Senators Chuck Grassley and Jim Banks have warned that protecting American AI technology from Chinese Communist Party espionage is of paramount importance. Inside the Trump administration, Treasury Secretary Scott Bessent has cautioned against open-source AI becoming “open season” on American intellectual property while Commerce has favored a softer approach. In the meantime, the commercial relationships continue uninterrupted.
What Needs To Happen
Specialized AI training data, expert rubrics, and knowledge pipelines that can strengthen Chinese military-linked entities should face export-control scrutiny comparable to advanced hardware. Congress should investigate the four firms, review any federal contracts they hold, and close the legislative loopholes that make this trade legal. And Washington should force a simple choice: American companies can do U.S. defense work, or they can do business with Chinese military-linked customers. Not both.
Communist China is coming for American chips, American researchers, and American models. And now, they’re coming for specialized training data that is America’s last-mile advantage. Whether policymakers treat that facilitation as the national-security vulnerability it is remains to be seen.
The Trump administration and Congress must act soon before Silicon Valley’s data merchants help China beat the United States in the AI race.
