Mercor raised US$750m at US$20bn selling expert judgment to AI labs

01What Mercor sells: expert hours, turned into training data

A model learns language from text gathered at scale. To do a lawyer’s or an analyst’s job, it needs examples and grades written by people who do that job. Mercor’s own guide lists the stages: supervised fine-tuning on pairs of prompts and expert answers; reinforcement learning from human feedback, in which people rank a model’s answers; and evals, sets of tasks with known good answers that test whether training worked. It says a lab will likely need human experts, not synthetic data, to push past what current models can do.²¹ The newest layer is the “environment”: recreated software where an agent practises a workflow, with tasks and “verifiers” that score it. Mercor says its experts write the tasks and verifiers.¹²

For the expert, the job description is short: review model outputs in your field, create examples that “teach models how experts think”, and grade quality against set standards. Most projects require at least undergraduate-level expertise, many need years of experience or an advanced degree, and applicants first sit an AI interview that adapts its questions to their field. Projects run for weeks to months and “can be extended, shortened, or concluded early” depending on the customer.⁴

The pricing is cost-plus. In Mercor’s words, “the vendor and client together decide what the right level of compensation is for the talent, and the vendor takes a fixed percentage of that.” It recommends paying by the hour rather than by the task, saying per-task pay pushes people to rush and skip hard tasks, and that many of the top labs it works with have adopted hourly pay. It recommends hiring annotators as contractors, who, it notes, do not get employees’ benefits.⁵ Its expert page gives an average contracted rate of US$90 an hour and 734,000 roles created.⁴ The network and the workforce are different sizes: over five million people are in the network, while in May Mercor counted more than 30,000 weekly active contractors.¹¹⁰

02From a hiring app to a data supplier in two years

Mercor was founded in 2023.² Its founders, chief executive Brendan Foody, chief technology officer Adarsh Hiremath and chief operating officer Surya Midha, were three 21-year-old Thiel Fellows when it raised its Series B; its mission page now signs Foody and Hiremath as co-CEOs.¹⁵³ It started as a hiring company that judged candidates from interview transcripts, résumés and portfolio sites, then, in CNBC’s account, found it had built a network of specialists and turned to hiring them to train AI models.¹⁶ When it raised its Series A in September 2024, Reuters still called it an AI hiring platform, with more than 300,000 people in its candidate pool.¹⁴ At its Series B in February 2025 it said its “current focus is on contracting experts to help advance the frontier of AI”, and that it was growing more than 51% a month.⁷

Foody set out the logic in March 2025. Mercor began 2024 “in our apartment with no US employees, under $1M in annual revenue, and only seed companies as customers”, and grew 6,400% that year. Labs were the wedge: they report how a contractor performed within days, against about three months at a typical employer, which trains Mercor’s hiring models faster, and they make “unreasonable asks” such as 300 people in two days, which force automation. The plan he published had four steps: hire people to train AI models, use those contracts to learn to predict job performance, expand to short contract roles, then “hire people for all jobs”.⁶

A competitor’s deal opened the market wider. In June 2025 Meta paid US$14.3bn for 49% of Scale AI, and some large labs, including Google and OpenAI, reportedly cut ties with Scale. “It just doesn’t happen too often in startups where your biggest competitor gets torpedoed overnight,” Hiremath told Forbes.¹⁶ Four months later Mercor’s Series C valued it at five times its Series B.⁸ The field is still crowded: Wired lists Surge, Handshake, Turing, Labelbox and Scale AI as rivals.¹⁷

03US$20bn, and what a US$2bn run rate means

The valuation rose eight-fold, then five-fold, then two-fold. General Catalyst led a US$3.6m seed in 2023; Benchmark led the Series A at US$250m; Felicis led the Series B at US$2bn and the Series C at US$10bn; the Series D is at US$20bn.¹⁴¹⁵⁸¹ That is 80 times the Series A valuation in 26 months. The US$20bn was on the table by July, when Bloomberg reported that Mercor had told investors it held a term sheet at that price.¹⁹

The revenue claim behind it is a run rate. A run rate takes a recent period’s revenue and scales it to a year; TechCrunch described Mercor’s as the latest month times 12.¹⁵ Foody wrote on 6 July that Mercor “crossed $2B in ARR in June, just 4 months after hitting $1B”.¹³ In May its engineers put the path at US$0 to a US$500m run rate in 17 months.¹⁰ A run rate is a month’s pace, not a year’s revenue already earned, and Mercor’s projects can be shortened or ended early by the customer.⁴

It is also gross. Mercor bills the buyer for the expert’s hours plus its percentage and counts the whole bill. The Information, which first reported the figure as “gross annualized revenue”, reported that contractors receive 60% to 70% of billings, as The Next Web relayed; on that range, Mercor keeps about US$600m to US$800m a year of the US$2bn.²⁰ Mercor called itself “a profitable Series C company” in May; it publishes no accounts.¹⁰

04Who pays, and what the breach showed

Mercor names no customers on its own pages. It says it spent two years building for “the frontier labs”, and that big tech, open-model developers, application companies and the Fortune 500 are now adopting the same work.¹ In May it described itself as “trusted by 6 of 7 Mag 7 companies”, the seven largest US technology stocks.¹⁰ In February 2025, most of a US$75m run rate came from AI labs, and Mercor said it worked with the five largest, OpenAI among them.¹⁵ Wired named OpenAI and Anthropic among the labs that rely on it.¹⁷

In late March 2026 a malicious release of LiteLLM, an open-source tool for routing requests to AI models, harvested credentials from systems that installed it, and Mercor was one of them. It confirmed the incident on 31 March.¹¹¹⁸ A group claimed to hold 4TB of its data; Mercor did not comment on that claim, and five contractors sued over their personal data, Business Insider reported.¹⁸

Mercor’s account, published on 25 June after a forensic investigation with Google’s Mandiant, is that “only a very limited subset” of nearly five million experts had sensitive information affected, that the impact on customers was limited because many work on their own platforms, and that “all frontier labs have increased their work with us”. It added round-the-clock managed detection, dependency audits and regular credential rotation.¹¹ Neither company has said publicly whether Meta resumed.

The chart shows the mechanism of concentration. Payouts track hours, hours track projects, and projects come from a few buyers who can end them early. When one large buyer stops, its experts stop billing on the same day. Payouts doubled between May and November, after the pause; Mercor has not said which buyers account for the rise.⁴¹⁷

05What the US$750m is for

Mercor says the money goes into infrastructure and research for evals and post-training, and into more environments where agents learn complex workflows. It wants enterprises to buy what labs buy: evals built from a company’s own workflows, and the training that follows.¹ It launched an enterprise platform in March that maps how staff work, through screen capture, internal wikis and AI-led interviews.²⁴

Deeptune is the environment half. Announced on 9 July, it had recreated “hundreds of enterprise applications”, from spreadsheets to Salesforce, for labs to train agents in; Andreessen Horowitz led its US$43m Series A earlier in the year; Mercor was already a customer, and Foody an early investor. No price was disclosed.¹² Foody told Fortune that a purchase was “in a lot of ways the main motivation” for his investment, The Next Web reported.²⁰

Research is the third. Mercor says it raised the score of the open Qwen3.5-397B model on its APEX-Agents benchmark from 16% to 27% of tasks fully passed, and in October it post-trained NVIDIA’s Nemotron 3.5 Super for legal, software and accounting work, with Harvey, Cognition and Ramp as partners.²²²³ NVIDIA is now also an investor. Mercor has opened offices in New York and London, has more than 400 staff, and says it spends three times as much on AI inference as on employee salaries.¹²