Meta Chief AI Officer Alexandr Wang called on American startups to make service to the U.S. government a foundational commitment, escalating a Silicon Valley debate over companies that supply artificial-intelligence training data to both federal customers and Chinese AI laboratories.

“Serving the US government should not be a commercial convenience, it must be a bedrock principle for startups,” Wang wrote, according to a report published Aug. 8 by The Times of India. His comment responded to Aakash Sabharwal, a vice president of engineering at Scale AI, the data-infrastructure company Wang founded before joining Meta in June 2025.

Sabharwal said American companies should not sell data to Chinese AI laboratories, particularly when they also work with the U.S. government. He said Scale AI does not accept that category of business and has turned down revenue because of its position. Companies that make a different choice, he argued, risk weakening American leadership in AI and creating national-security concerns.

The exchange followed a Forbes investigation into a largely unregulated market for the specialized data used to improve advanced AI models. The report said American data businesses that serve customers including OpenAI and Anthropic have also supplied Chinese technology groups. Some of the U.S. vendors have separately worked with government entities, creating questions about whether commercially valuable methods developed within the American AI ecosystem are helping overseas competitors advance more quickly.

Forbes identified Surge AI, Mercor, AfterQuery and Turing among the U.S.-based providers connected through documents or communications to Chinese buyers. The reported customers included Tencent, Ant Group, Alibaba and ByteDance. The publication said the Chinese companies did not respond to its requests for comment, while Mercor declined to comment and AfterQuery and Surge said they do not disclose customer information.

Turing told Forbes that it supports the frontier of AI, including proprietary and open-source models, and noted that some of the strongest open-source systems originate outside the United States. That position reflects a competing view within the industry: International sales can support American businesses and the global AI research community, while sweeping restrictions could favor the largest closed-model developers and reduce competition.

The scale of the market is significant. Forbes reported that the six largest Chinese AI labs spend about $500 million a year with American data-labeling providers, citing estimates shared with two entrepreneurs in the sector. The figure has not been independently confirmed through public financial disclosures, and many of the companies involved are privately held. Nevertheless, the reported trade illustrates how valuable training data has become as developers seek incremental improvements in reasoning, coding and professional tasks.

Modern AI datasets are more than collections of text, images or software code. Vendors recruit specialists, design prompts, establish grading standards and evaluate model responses. For complex assignments, that work can involve people with expertise in fields such as finance, cybersecurity, medicine or advanced mathematics. The resulting material teaches models not only what an acceptable answer looks like but also how experts approach and assess difficult problems.

Prepackaged datasets can be especially useful because they allow multiple customers to purchase training material built with established expert networks and quality-control systems. A laboratory that buys such material may avoid months of experimentation needed to design its own tasks and scoring methods. That makes the data supply chain strategically important even though it does not resemble the physical infrastructure traditionally covered by export controls.

U.S. policy has focused primarily on restricting China’s access to the most advanced AI processors and related semiconductor technology. Those measures address computing power, one of the essential ingredients for training large models. The controversy highlighted by Wang concerns another key input: the human knowledge and feedback that help convert computing capacity into a capable system.

Forbes reported that Tencent circulated a request for datasets covering subjects including finance, cybersecurity and systems capable of improving their own performance. Tencent has disputed a previous U.S. government characterization linking it to China’s military. The reported requests nonetheless illustrate why transactions involving seemingly ordinary commercial data may attract scrutiny when the buyer is a major Chinese technology company.

No blanket U.S. prohibition identified in the reporting prevents American companies from selling AI training datasets to Chinese customers. Whether an individual transaction is permissible can depend on the buyer, the technology involved and applicable export-control or sanctions rules. Wang and Sabharwal are therefore making an argument that extends beyond minimum legal compliance: They want startups to consider national alignment when deciding whom to serve.

That approach carries practical consequences. Federal agencies increasingly rely on private companies for model evaluation, data preparation, decision-support tools and secure AI systems. A vendor seeking government work may face questions about foreign customers, access controls and the possibility that personnel or technical methods cross between projects. Even without new legislation, procurement requirements could push contractors toward stricter screening and internal separation of sensitive work.

Scale AI has direct experience with both sides of the issue. The company built its business by organizing and evaluating the data used by major AI developers, then expanded into government and defense work. Forbes reported that Scale considered a ByteDance contract before abandoning it in 2024 because of national-security concerns. Scale’s current stance gives Wang’s statement additional weight, although he now leads Meta rather than Scale.

Meta’s official biography describes Wang as the company’s first chief AI officer and leader of Meta Superintelligence Labs. His move followed Meta’s major investment in Scale AI and placed him in charge of one of Silicon Valley’s most closely watched efforts to develop highly capable AI systems. That role gives his comments relevance beyond the data-labeling sector: Meta is a major platform company, an AI developer and a potential customer for the same specialized resources at issue.

The episode also raises a difficult distinction between ordinary global commerce and support for a strategic competitor. Many American technology companies depend on international customers, researchers and open-source communities. Limits written too broadly could reduce revenue for U.S. startups or drive overseas laboratories toward domestic alternatives. Limits written too narrowly, however, may leave valuable expertise available to companies developing models that compete directly with American systems.

For startup leaders, the immediate effect is likely to be greater scrutiny of customer policies. Boards and investors may ask whether a company sells to entities in countries treated as strategic rivals, whether federal contracts impose additional obligations and how proprietary workflows are protected. Firms may also need clearer rules for distinguishing generic datasets from material that could meaningfully accelerate advanced model development.

Wang did not announce a Meta policy or call for a specific federal regulation in the reported exchange. His statement instead framed the issue as a principle for founders. That framing could prove influential in a technology sector where voluntary commitments often emerge before formal rules, particularly when national security and access to government contracts are involved.

The policy debate is likely to continue because AI development depends on an interconnected stack of chips, cloud infrastructure, algorithms, data and human expertise. Restricting one layer does not necessarily prevent capabilities from moving through another. Wang’s message is that American startups should account for that reality themselves—and that working with Washington should carry obligations extending beyond the immediate value of a government contract.