Mercor’s Fast Growth Relies on Biggest AI Companies, Documents Show
The Takeaway
- Nearly all of Mercor’s gross revenue this year came from AI model makers
- Mercor’s gross revenue, or sales before paying its contractors, jumped 70% in first half from full-year 2025
- Three-year-old startup predicts it will increase revenue from non-AI customers
Mercor, a three-year-old data startup whose army of contractors teaches AI to give better answers, is growing fast but relies heavily on revenue from AI foundation model companies, financial documents seen by The Information show.
The company generated $614 million in gross revenue in the first half of the year, up 70% from all of last year, according to the documents. About 91% of the first half revenue came from makers of AI foundation models. Mercor’s top customers include OpenAI, Anthropic and Google DeepMind, the company has told potential investors, along with newer AI labs such as open-source provider Reflection AI and Thinking Machines Lab.
The concentration among the biggest AI companies points to a simmering concern among investors: Many startups owe much of their growth to a few large customers whose decision to take their business elsewhere—or handle it internally—could abruptly cool the torrid trajectory of these startups. It is a central question as Mercor discusses a new round of funding that, according to a person familiar with the matter, could value it at $20 billion, double its level last fall.
That higher valuation would follow Mercor’s March disclosure that it was among the companies targeted by hackers in a supply-chain attack. The revelation prompted Meta Platforms to pause its work with the company, according to Wired. Meta, which also owns a 49% stake in Mercor rival Scale AI, has not resumed working with Mercor, according to the person.
Mercor said last month the impact on customer information was “very limited,” and few contractors had “sensitive information” breached.
The startup predicts that the concentration of its customers in foundation models will fade over the next half decade as it sells more of its evaluation and training services to other AI startups, including app developers that need to evaluate their customized models, as well as large companies outside AI that want to make sure their agents are working as intended. The company’s fundraising material listed financial firms such as Ramp, Blue Owl and Citi as customers.
By 2028, Mercor expects gross revenue from AI startups and Fortune 2000 companies combined to eclipse gross revenue from foundation models. By 2030, it expects foundation models will represent just under one-third of its anticipated gross revenue of nearly $69 billion. Fortune 2000 companies, in contrast, will make up 57%, Mercor has told potential investors.
The documents point to other constraints on the business, which has fetched increasingly high valuations as revenue has surged.
It pays about two-thirds of its gross revenue to contractors, the legion of lawyers, doctors, writers and doctorate holders that grade AI models’ answers to questions. Those payouts kept its gross margin to 27% last year and 33% in the second quarter of this year. It anticipates gross margins rising to 46% next year and 56% in 2030.
Mercor’s past gross margins aren’t far off from comparable metrics at OpenAI and Anthropic, which both missed their own gross margin forecasts as the costs of running their AI models spiked. But Anthropic late last year expected gross margins could hit 77% in 2029, while OpenAI earlier this year predicted gross margins would close in on 70% toward the end of the decade.
Those levels would put the model makers close to best-in-class software companies such as GitLab, which reported an 86% gross margin in its most recent financial quarter—a sharp contrast with Mercor.
Rocketing Revenue
Mercor was started by college dropout Brendan Foody and two of his high school classmates and has emerged as one of the most prominent in a competitive cohort of startups. The companies, which include Scale AI, Handshake AI and Surge AI, supply specialized training data to AI companies looking to improve or fine-tune their models.
Mercor’s annualized gross revenue—monthly revenue multiplied by 12—hit $2 billion in June, double the pace from early this year, and it told potential investors it expects to hit $2.8 billion in annualized gross revenue by year end.
Similarly, rival Handshake’s gross revenue jumped 82% by April from January, to an annualized pace of nearly $1 billion.
The businesses, which rely mainly on human contractors, risk losing demand to startups that specialize in designing copies of popular apps like Salesforce or Excel to teach AI how to use those apps. Some researchers say these fake apps, otherwise known as reinforcement learning environments, work better for certain disciplines than human experts do.
Earlier this month, Mercor said it had bought Deeptune, a startup that designs RL environments, which Mercor said will help the company quickly build more realistic training environments for apps. Mercor also argues that there will continue to be a growing market for humans to train AI, particularly in high-stakes, complicated situations like legal cases.
The startup swung to a small operating profit last year, of $9 million, but sees that leaping to $226 million this year and $1.7 billion next. It forecast that its operating margin will rise from 3% last year to 12% this year and 26% next year, eventually climbing to 45% in 2030. Those forecasts would put it closer to best-in-class software stocks on this basis. Palantir had an operating margin of 46% in the first quarter.