Four Silicon Valley data-labeling companies that help train some of the world’s biggest AI models are also selling the same kind of work to China’s leading AI labs, according to a New York Post report published August 30, 2026. The companies named, Surge AI, Mercor, AfterQuery, and Turing, supply human-annotated training data to Anthropic, Google, Meta, and OpenAI. Some of them also hold contracts tied to the Pentagon. The story lands at a moment when Washington is trying to choke off China’s access to advanced AI capability through chip export controls, and it raises a question chip bans never addressed: what happens when the expertise itself, not the hardware, crosses the border.
The report says six major Chinese AI labs collectively pay US data companies roughly $500 million a year for human-generated training sets covering coding, financial modeling, and cybersecurity tasks. That figure, combined with the Pentagon-adjacent contracts held by some of the same firms, has drawn a rebuke from at least one member of Congress and revived a debate that first flared in 2024 when Scale AI walked away from its own China business.
What the Report Actually Says
The New York Post framed its story around a straightforward tension: the same data-annotation firms that help the Pentagon and America’s top AI labs train frontier models are, in parallel, filling orders from Chinese AI developers. The piece names four companies specifically. Surge AI, Mercor, AfterQuery, and Turing all provide what the industry calls RLHF data, human-written and human-graded examples used to fine-tune large language models on hard tasks like debugging code, reasoning through legal documents, or modeling financial risk.
According to the report, those four firms supply data services to Anthropic, Google, Meta, OpenAI, and the Pentagon. On the other side of the ledger, the same companies (or, in some cases, the same corporate entity under a different contract) have done business with Tencent, Ant Group, Alibaba, and ByteDance, the Beijing-based parent of TikTok. None of the arrangements described are alleged to be illegal. Data-labeling work sits outside most existing export-control regimes, which were written for physical goods like semiconductors, not for annotated text files and reasoning transcripts. Forbes reported on this same dynamic in early August, describing a wider group of American data-annotation startups quietly making China’s AI systems more capable.
The Four Companies Named in the Report
Each of the four firms named plays a different role in the pipeline, and each has a different degree of exposure to Washington. The table below lays out what the report and subsequent coverage established about each one.
| Company | Reported Chinese Client(s) | US Government / Military Tie | Reported Revenue Detail |
|---|---|---|---|
| Surge AI | Tencent | Contracts described as “small” with the US Army and US Air Force | About $1.2 billion in 2024 revenue |
| Mercor | Tencent | Holds a federal contract (specifics not detailed in the report) | Not disclosed |
| AfterQuery | Ant Group, Alibaba | Not named as a direct Pentagon contractor | At least $50 million in recurring annual revenue from Chinese AI labs |
| Turing | ByteDance | Not named as a direct Pentagon contractor | Not disclosed |
Surge AI stands out because it appears on both sides of the sensitivity spectrum at once: a nine-figure Tencent relationship sitting next to defense work for two branches of the US military. Mercor’s exposure is different in kind. It holds a federal contract while also serving Tencent, which puts it closer to a foreign ownership, control, or influence (FOCI) review than a company that only sells to commercial labs. TechCrunch has covered Mercor’s rapid growth in the data-labeling race, noting how quickly the company scaled its client roster once demand for RLHF work took off in 2025.
Why Tencent Is the Flashpoint
Tencent isn’t an ordinary customer. The Pentagon added Tencent to its list of “Chinese military companies” operating in the United States back in January 2025, a designation meant to flag firms the Department of Defense believes are linked to China’s military-civil fusion strategy. Tencent has disputed the label and sued over a similar listing in the past, but the designation has stayed in place through 2026. That means a US firm selling annotated training data to Tencent is, on paper, doing business with an entity the Pentagon itself treats as part of China’s defense-industrial base, even while some of the same US firm’s other contracts run through the Army or Air Force.
How the Money Flows
The $500 million annual figure cited in the report covers six major Chinese AI labs buying advanced instruction sets, the kind originally built for American frontier-model developers, and repurposing them to train their own systems on tasks like coding assistance, financial analysis, and cybersecurity reasoning. That’s not a huge number next to the tens of billions US labs spend on compute, but it’s meaningful because of what it buys: the same caliber of human expertise, delivered by contractors, engineers, and subject-matter specialists in the US, UK, and elsewhere, that OpenAI, Anthropic, and Google pay for their own frontier models.
Data labeling has quietly become one of the highest-margin businesses in AI. Surge AI’s reported $1.2 billion in 2024 revenue, first flagged in detail by CoinDesk’s coverage of the story, puts it in the same conversation as mid-sized defense contractors, built almost entirely on a business that didn’t exist in its current form five years ago. Nvidia’s own customer base is increasingly building rival chips to cut compute costs, and a similar dynamic is playing out in data: the labs that once relied on a handful of US annotation shops now have Chinese alternatives emerging too, which is part of why American firms serving both markets are drawing scrutiny now rather than later.
Rep. McCaul’s Warning
The New York Post report quotes Rep. Michael McCaul (R-Texas), a longtime China hawk and former chairman of the House Foreign Affairs Committee, on the record. “America is in an AI race with the Chinese Communist Party,” McCaul told the paper. He went further on the specific conflict the report describes: “If you want to do business with the US government, you should not also be helping a Pentagon-designated Chinese military company build better AI.” That argument mirrors the logic behind the Commerce Department’s chip restrictions, extended to a category, human-generated training data, that current export-control law doesn’t clearly cover.
The Scale AI Precedent
This isn’t the industry’s first brush with this exact problem. Scale AI, once the dominant name in data labeling and now roughly half-owned by Meta after a 2024 investment, walked away from its ByteDance contract in 2024 on national-security grounds after facing pressure over the relationship. That decision effectively created an opening: the Chinese AI labs still needed the same kind of human-annotated data, and a wave of newer firms, including the four named in this report, stepped in to fill the gap Scale AI left behind.
| Episode | Year | Company Involved | Outcome |
|---|---|---|---|
| Scale AI exits ByteDance contract | 2024 | Scale AI | Contract dropped after national-security pressure; Meta later takes a large stake in Scale AI |
| Pentagon designates Tencent a “Chinese military company” | January 2025 | Tencent | Designation remains in effect through 2026; Tencent disputes the label |
| Four data firms report Chinese AI lab revenue while holding US ties | 2026 | Surge AI, Mercor, AfterQuery, Turing | New York Post report triggers Congressional criticism; no regulatory action confirmed yet |
The pattern is familiar from other corners of the tech industry too. When one vendor exits a China relationship over political pressure, the business rarely disappears, it just moves to a competitor with a lower public profile. That’s effectively what happened after Scale AI stepped back, and it’s a big part of why critics argue that company-by-company voluntary restraint doesn’t work without a broader rule covering the whole category of data-labeling services.
Where This Sits Legally: FOCI, CFIUS, and the Gray Zone
Foreign ownership, control, or influence review, known as FOCI, governs whether a company holding US government contracts has financial or operational ties that could compromise sensitive work. It’s typically applied to hardware and cleared-facility contractors, not to firms selling annotated text datasets on the commercial market. The Committee on Foreign Investment in the United States (CFIUS) reviews foreign acquisitions of US companies, but it has no clear mechanism for reviewing a US company’s outbound sale of services to a foreign buyer.
That’s the gap this story sits in. Mercor holding a federal contract while also serving Tencent is the kind of dual exposure that would trigger a FOCI review if Mercor were, say, building components for a fighter jet. Because it’s providing annotated training data instead, the same relationship falls into a regulatory blind spot. Congress has held hearings on AI data supply chains before, but no statute currently forces a data-labeling firm to disclose or divest a Chinese AI lab relationship the way it would for, say, a chip fabrication contract.
Why the AI Labs Haven’t Weighed In
OpenAI, Anthropic, Google, and Meta are customers of the named data firms, not the sellers to China, and none of the four labs has issued a public statement addressing the specific relationships described in the report as of this writing. That silence is itself notable given how vocal these same companies have been about competitive positioning against Chinese AI developers in other contexts. A frontier lab that discovers its own data vendor is training a Chinese rival on similar material has an obvious incentive to ask questions about contract terms and data segregation, even if there’s no evidence any labeled data itself crossed between the US and Chinese sides of a vendor’s business.
Market and Industry Impact
The immediate market impact is reputational rather than financial. None of the four companies named is publicly traded, so there’s no stock to move. But data-labeling firms have increasingly positioned themselves as candidates for the same kind of government scrutiny that has hit AI infrastructure deals with China exposure elsewhere in the stack. Surge AI’s $1.2 billion revenue run rate makes it a serious business, not a scrappy startup, and serious businesses tend to draw serious regulatory attention once a story like this one breaks nationally.
There’s also a downstream effect on the AI labs themselves. If Congress moves to restrict which vendors frontier labs can use for training data, sourcing costs go up and vendor lists shrink. That would be a meaningful change for an industry that has scaled its human-feedback pipelines aggressively over the past two years, often through exactly the kind of specialized contractor network the report is now scrutinizing. It also lands against a broader backdrop of rising concern about AI-related security risk across the industry, where data provenance and vendor trust have become as important as model architecture itself.
How This Compares to Chip Export Controls
Washington’s approach to slowing China’s AI progress has focused almost entirely on hardware: restricting sales of advanced Nvidia chips, tightening rules on chip-making equipment, and pressuring allies to follow suit. That strategy assumes compute is the bottleneck. This report suggests a second bottleneck that chip controls don’t touch at all, the human expertise needed to turn raw compute into a capable model. A Chinese lab with restricted access to top-tier chips still benefits enormously from buying the same quality of instruction-tuning data that OpenAI or Anthropic uses, because it closes the capability gap without needing a single additional GPU.
That’s the structural argument critics are making, echoed in a RedState column published days before the New York Post report: hardware controls without data controls are only half a policy. It’s a dynamic that echoes other corners of the AI industry right now, where margin pressure and supply-chain scrutiny, from chipmakers facing new competition to data vendors facing new questions, keep landing in the same place. Building an equivalent enforcement regime for services, though, is much harder than restricting physical exports. A container of GPUs crosses a border and customs can stop it. A dataset delivered over the internet, or a contractor doing remote-labeling work for two clients on two continents, leaves no comparable checkpoint.
Historical Context: The Rise of the Data-Labeling Industry
Data labeling used to mean low-wage crowdworkers tagging images for a few cents a task. That changed once large language models started needing reinforcement learning from human feedback (RLHF), which requires domain experts, not just clickworkers, to write and grade responses on coding problems, legal reasoning, and multi-step math. That shift turned data labeling into a business that could plausibly generate over a billion dollars in annual revenue for a single company, something unthinkable in the crowdworking era of five years ago.
Scale AI built the first version of this business at scale, going from a computer-vision labeling shop to the default RLHF vendor for OpenAI and others before its high-profile 2024 restructuring around Meta’s investment. The vacuum that restructuring created is exactly what let Surge AI, Mercor, AfterQuery, and Turing grow as fast as they did, and it’s why all four now find themselves in a position where serving both American and Chinese frontier labs is not just possible but, commercially, close to inevitable if they want to keep growing.
Competitive Landscape: Who Else Is in This Business
Beyond the four companies named directly in the report, the broader RLHF and data-annotation market includes firms like Scale AI (now Meta-affiliated and reportedly more cautious about China exposure since 2024), Invisible Technologies, and a long tail of smaller, specialized annotation shops serving narrower verticals like legal or medical AI. Most of these companies are privately held and don’t disclose customer lists, which is part of why a story naming four specific firms and their specific Chinese clients is unusual, and part of why it’s generating this much attention now.
What differentiates the four named companies from the rest of the field isn’t that they work with China, multiple firms in this space likely do to some degree, it’s that they also carry visible US government or defense-adjacent business at the same time. Industry tracker AI Weekly flagged the same pattern shortly after the report broke, noting that the overlap, not the China business alone, is what turns a routine commercial relationship into a story a national newspaper puts on its front page and a member of Congress puts on the record about.
What Happens Next: Five Predictions
- Expect at least one Congressional hearing or formal letter of inquiry directed at the Pentagon and one or more of the four named companies within the next few months, given the tone of the reported criticism.
- Frontier AI labs will likely add contractual language requiring data vendors to disclose or restrict work with Chinese AI labs, mirroring how cloud providers already restrict certain customers under export-control policy.
- Watch for at least one of the four companies to publicly clarify or scale back its Chinese AI lab relationships in the coming weeks, following the same pattern Scale AI set in 2024.
- Lawmakers will likely push to extend some form of export-control or FOCI-style review to data-labeling and RLHF services, closing the gap that currently exists between hardware rules and service-sector activity.
- Expect renewed scrutiny of other categories of “invisible” AI supply chain vendors, including compute brokers and fine-tuning contractors, as journalists and regulators realize hardware controls are only one part of the picture.
The Bigger Picture for AI Supply Chains
This story is really about a blind spot in how the US thinks about protecting its AI edge. Export controls were built for a world where the thing worth restricting was a physical object, a chip, a piece of fabrication equipment, a server. Frontier AI training increasingly depends on something much harder to track: skilled human judgment, delivered as a service, by contractors who can serve any client willing to pay. Closing that gap means either extending existing frameworks like FOCI and CFIUS to cover services, which is legally awkward, or writing new rules specifically for the data-annotation industry, which barely existed in its current form three years ago.
Until that happens, the four companies named in this report, and the many similar firms that weren’t, will keep operating in a market where selling to both the Pentagon and Chinese AI labs isn’t against any specific rule. It’s simply a business decision that a national newspaper and a member of Congress happened to notice this week.
Frequently Asked Questions
Which companies were named in the New York Post report?
Surge AI, Mercor, AfterQuery, and Turing. All four provide human-annotated training data used to fine-tune AI models, and the report says all four have done business with Chinese AI labs while also serving American AI companies or US government-linked contracts.
How much do Chinese AI labs reportedly pay US data companies?
The report puts the figure at roughly $500 million a year, paid collectively by six major Chinese AI labs to US-based data-annotation firms for advanced instruction sets covering tasks like coding, financial modeling, and cybersecurity.
Is it illegal for a US company to sell training data to a Chinese AI lab?
Not under current law. Data-labeling and RLHF services fall outside most existing export-control frameworks, which were written to cover physical goods like semiconductors rather than annotated datasets or contracted human labor.
Why does Tencent’s involvement matter specifically?
The Pentagon designated Tencent a “Chinese military company” in January 2025, a label the Department of Defense uses to flag firms it believes are tied to China’s military-civil fusion efforts. Tencent disputes the designation, but it remains in effect, which means US firms selling to Tencent are doing business with an entity the Pentagon itself has flagged.
Did Scale AI face a similar situation?
Yes. Scale AI, once the dominant data-labeling vendor for US AI labs, exited its ByteDance contract in 2024 following national-security pressure. Meta later took a large investment stake in Scale AI. The business that Scale AI walked away from didn’t disappear, it moved to newer competitors, including some of the firms named in this latest report.
Have OpenAI, Anthropic, Google, or Meta commented on the report?
None of the four labs had issued a public statement addressing the specific vendor relationships described in the report as of this writing. They are customers of the named data firms rather than parties to the China-facing contracts described.
What could Congress actually do about this?
Options include extending FOCI-style review to data-annotation contractors that hold federal work, requiring disclosure of foreign AI lab clients for any vendor serving US government contracts, or creating a new export-control category specifically for AI training-data services. None of these currently exist in statute.
Does this affect the chips the US restricts from sale to China?
No, chip export controls are a separate policy track and remain unaffected. This story highlights a different bottleneck, human-generated training data and expertise, that current chip-focused export rules don’t cover at all.
