Invest1 distinct publisher3 min readPublished
The money is small measured against these vendors' own revenue and large measured against anything else in the training-data supply chain, which is why a control that does not yet exist would be cheap to impose and hard to enforce.
The Investor · Invest desk

Compiled by The InvestorSomething wrong?How this is made
Start with the only percentage in the file. Mercor's Chinese contracts came to 2% of its second-quarter 2026 revenue [4], and its annualized run rate crossed $2bn around June of that year [3], so the China line runs on the order of $40m a year [1]. AfterQuery reports at least $50m of recurring annual revenue from Chinese labs alone [6]. Those two add to roughly $90m, about 18% of the $500m that American labelling firms collectively take from China's six leading AI operations, which leaves something near $410m attributable to Surge, Turing and vendors the report does not name [2][7].
Across six buyers, $500m averages about $83m a lab [3]. What $83m buys is less a dataset than access to English-fluent domain experts, often holding advanced degrees, which the report identifies as the input hardest to reproduce inside China [8]. That is a price paid for a labour pool, not for a commodity, and it is why the pool cannot be stood up locally on a purchase order. Set the whole China flow against Surge's $1.2bn of 2024 revenue and it comes to roughly 42% of one vendor's single year [4][2], a sum that matters far more to the six buyers than to the sellers who supply it.
That asymmetry is the interesting deal term. Mercor's $40m is two cents on its own revenue dollar and about 8% of the entire China pool [5], so a prohibition would cost the American vendor a rounding error while removing from the Chinese buyer the one thing it cannot manufacture. Export controls normally run the other way, with the domestic seller doing the lobbying against them. And the reason there is no rule is the category rather than the politics: the Bureau of Industry and Security has built restrictions around chip exports, chip-making equipment and cloud computing access, and nothing that governs the sale of annotated training data [10]. BIS has simply put its enforcement capacity into those three categories and never built a fourth for training data.
This is probably wrong, but the dollars look like the less consequential exposure. Anthropic buys from the same annotation services [9], the US military work sits inside the same firms [2], and one named buyer, Tencent, is designated by the Pentagon as a Chinese military company [5][11]. Concentration in that layer means the task specifications that reveal what each side is trying to teach its models pass through a small number of commercial counterparties. That is closer to a counterintelligence question than a trade one.
How this plays out could go several ways. Commerce writes something covering services and the trade reroutes through offshore intermediaries that no $500m estimate would capture; the vendors self-sanction first, because 2% of revenue is cheaper than a problem with defence-adjacent contracts; or the category stays unowned and the flow keeps growing. What would break the thesis: the aggregate is single-sourced to one report [12], and it stacks a 2024 annual revenue against a mid-2026 run rate, so $500m cannot be read as a same-period share of anything [2][3]. If annotated work of comparable quality is available from non-US suppliers, a control is a press release. On the numbers as given, the whole China trade is about 16% of just two vendors' disclosed scale [6], which is roughly why nobody has yet had to choose.
Ranked by verification strength, evidence, and original report placement.
The Commerce Department's Bureau of Industry and Security has built increasingly sophisticated restrictions around chip exports, chip-making equipment and cloud computing access, but there is no equivalent framework governing the sale of annotated training data.
The Pentagon has designated Tencent as a Chinese military company, meaning a US defense-adjacent data firm is simultaneously servicing an entity the Department of Defense considers part of China's military apparatus.
All of the revenue and client figures in this story come from a single report published by cryptobriefing.com.
US data labeling firms collectively earn roughly $500 million per year from China's six leading AI operations.
Surge AI reported $1.2 billion in revenue in 2024, built on providing expert-annotated datasets to both US military branches and Chinese AI laboratories.
Mercor crossed an annualized revenue run rate of $2 billion by June 2026.
Distinct publishers with included, body-backed reporting in this cluster.
cryptobriefing.com
1 article · August 30, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
invest
The chips never move: Washington's fix for the Southeast Asia compute loophole1 distinct publisher
product
Nvidia's H200 finally lands in China at about 1% of the order book2 distinct publishers
invest
A BIS draft would reprice the offshore racks Tencent rented at $80,000 a chip1 distinct publisher
build
China's accelerator swap makes Cambricon supply, not export policy, your ship-date risk1 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One unsourced report carrying every number
The $500m pool, Surge's $1.2bn, Mercor's $2bn run rate and AfterQuery's $50m all appear in exactly one place, and Crypto Briefing names no filing, study or person behind any of them. Five vendors are described; none is quoted. Four Chinese buyers are named; none is asked. What survives outside this report is narrow but real: the Pentagon's designation of Tencent, and the plain fact that the export rulebook covers chips, tooling and cloud while saying nothing about annotated data.
Real trade, unmeasurable size
What can actually be watched is small: a 2% revenue-mix line at Mercor for one quarter and a $50m floor at AfterQuery. Both are relayed, not disclosed — no contract, no customer confirmation, no procurement record. That is enough to accept that Chinese labs are buying US annotation at commercially meaningful volume, and nowhere near enough to size it, which is why our own arithmetic leaves roughly $410m of the stated pool sitting behind no named line item.
Scandal framing, rounding-line arithmetic
The headline promises a national-security breach; the numbers inside behave like a footnote. Mercor's Chinese business is 2% of its own book, and the whole $500m pool is smaller than half of Surge's 2024 revenue. Meanwhile the part that genuinely is underplayed — that no rule has been written for this category at all, so there is nothing to violate and nothing to license — gets three sentences at the end.
Motive not visible
We cannot see who gains from these figures being in print. Crypto Briefing credits nobody, no vendor or buyer speaks, and no agency is positioning itself in the text. Numbers this specific about private companies usually reach a reporter from someone with a reason, but naming that reason would be invention, so we leave it open.
Plausible frame, unverified digits
Low, and low in a particular direction. The shape of the story — American annotators selling to Chinese labs while no export rule covers the product — is plausible and would surprise nobody if confirmed. The digits carrying it are unattributed private-company revenue from a single crypto-trade outlet, and our own ratios inherit that weakness whole. Treat the frame as a lead worth chasing and the figures as claims awaiting a second name on them.