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.
