Mercor

The recruiter for the machines

It is 2 a.m. somewhere, and a board-certified radiologist in Mumbai is grading an AI's read of a chest x-ray. A patent lawyer in Berlin is rewriting a model's draft of a software claim. A bond trader in Singapore is telling a chatbot why its hedge ratio is wrong. None of them work for Mercor, exactly. All of them are paid by it. This is what an AI factory looks like when the assembly line is made of brains.

Mercor is the marketplace those brains pass through. The company recruits domain experts - doctors, engineers, lawyers, bankers, scientists, journalists - and rents their judgment, by the hour, to the labs building the world's most expensive software. The labs in question are OpenAI, Anthropic, Google, Meta, Amazon, and most of the rest of the names that show up on any quarterly earnings call about artificial intelligence. The experts get paid well, often more than their day jobs. Mercor takes a margin. The model gets smarter. Repeat.

What they noticed first

In late 2022, the trio behind Mercor - Brendan Foody, Adarsh Hiremath, and Surya Midha, all 20 years old and recently dropped out of college via the Thiel Fellowship - were running a small outsourcing shop placing engineers from India into U.S. startups. It was unglamorous work, the kind that does not appear in pitch decks. But it taught them something useful. Hiring, even with software, is mostly judgment. And judgment scales badly.

Then ChatGPT happened. Within a quarter, every AI lab in San Francisco wanted the same impossible thing: very smart people, willing to do unusual short-term work, vetted in days, paid across thirty countries, gone tomorrow. The labs did not need an HR department. They needed a faucet. Mercor decided to be the faucet.

Three founders, one wager

The bet was simple and slightly heretical. Most people in 2023 thought AI would replace knowledge workers. Mercor's founders thought, correctly, that AI would first need to hire a great many of them. To learn medicine, a model has to read what a doctor flags as wrong. To learn law, it needs an attorney's red pen. The frontier of model quality, the founders argued, was no longer compute. It was access to scarce expertise, organized and paid on time.

So they built two things at once. The first was an AI-driven interviewer that could screen a candidate from a resume and a fifteen-minute conversation, ranking thousands of applicants for any niche specialty - say, "Korean-speaking oncologists comfortable with computational biology" - in under an hour. The second was a payments and compliance back-end that could move money to those people in dozens of currencies without a finance team's permission. Together, they formed something between a staffing firm and a stock exchange.

How the faucet works

From the outside, Mercor looks like a job board for very specific people. A candidate uploads a resume; the system scrapes it, runs an AI interview, scores the result, and either matches the candidate to an open contract or files them away for the next one. There are no recruiters in the loop unless something is genuinely strange.

From the inside, things get more interesting. The AI labs do not hire one expert at a time. They hire programs - hundreds of contributors evaluating model outputs against rubrics that often did not exist a month earlier. Mercor's quieter business is writing those rubrics, running those evaluations, and turning the messy output into clean training data the lab can actually use. The buzzword is "human data." The reality is closer to a global, distributed graduate seminar grading homework written by GPT.

What sits inside the platform

The Talent Marketplace matches experts to contracts. The AI Interviewer handles the screening. Frontier Data services package the result into RLHF and evaluation pipelines. A managed projects arm runs the whole engagement for customers who would rather buy outcomes than headcount. A global payroll layer pays everyone. Each piece is a defensible business on its own. Stacked, they are a moat.

Receipts, in dollars and contractors

Numbers are the only argument venture capitalists like. Mercor's numbers are loud. The company has reportedly grown from $1 million to roughly $500 million in annualized revenue inside 17 months - a pace that, if accurate, is among the fastest revenue ramps any private software company has ever reported. Most of the revenue comes from a handful of AI labs paying for what is effectively a permanent talent contract.

Organizing human intelligence

Mercor's stated mission is to "organize human intelligence to power the AI economy." Strip the press-release varnish and what is left is a real claim: that as AI absorbs the parts of work that are pattern recognition, what becomes valuable is the part that is judgment - the thing a board-certified specialist can do at 2 a.m. that a model cannot do yet, but might be able to do tomorrow if shown enough examples. Mercor's job is to find that person, pay them, and route their judgment to the place it is needed most.

There is a sharper way to put it, which Brendan Foody has put on stage more than once. The labor market, he argues, is wildly mispriced. A great cardiologist is also a fine writer. A trial lawyer can also code. Most people are stuck working on a fraction of what they are good at because the cost of discovering them is too high. AI fixes that cost. Mercor is the company that wants to collect the savings.

What could go wrong

The skeptic's case is fair and worth saying out loud. Mercor's revenue is concentrated in a small number of frontier-model customers, each of whom has the technical ability to bring this work in-house if they decide it is strategic. The category - data labeling and evaluation - has produced one breakout already (Scale AI) and a graveyard of contenders. Late 2025 brought an uncomfortable confirmation that Mercor had suffered a data breach involving contractor information, a reminder that holding thirty thousand people's bank details and government IDs is a serious job.

None of which means the bet is wrong. It means the bet has consequences. A marketplace at this scale is no longer a startup. It is infrastructure, and infrastructure breaks publicly when it breaks at all.

Back to the radiologist

It is 2 a.m. again. The radiologist in Mumbai has finished grading her batch. The model has been corrected, very slightly, on a particular kind of nodule it kept getting wrong. Somewhere in a server farm, a checkpoint is written. In the morning, a doctor in Cleveland will use a tool that is a little less stupid than it was yesterday, and will not know whose night was traded for that improvement. Mercor will know. Mercor will also have paid. That is the whole product.

If the founders are right, this kind of transaction - a graduate-level brain rented by the hour, on a global rail, to teach a model a single human thing - becomes the dominant shape of knowledge work in the next decade. If they are wrong, they have still built one of the fastest-growing companies in the history of software. Both of those outcomes are useful. Only one of them changes the labor market.

Interviews and demos