The AI training market just produced its most startling number yet. Mercor, a three-year-old company that pairs human experts with the labs building frontier models, has raised a $750 million Series D at a $20 billion valuation, roughly double what it was worth less than a year ago. The round, announced by the company on November 10, arrived in the middle of the sharpest stretch of AI-stock anxiety since the boom began.
A $20 Billion Vote of Confidence
That timing is the whole story. While investors spent late October and early November dumping chipmakers and neoclouds on fears that AI spending has outrun its returns, a long list of marquee names lined up to write Mercor a check. Benchmark, General Catalyst, Nvidia, Salesforce, Zoom, Felicis, Sapphire Ventures, StepStone Group and funds affiliated with Apollo all took part, alongside an employee tender that let early staff cash out. For a company whose product is human judgment rather than silicon, the vote of confidence in the AI training business is pointed.
Mercor's valuation has moved at a speed that embarrasses even this era. It was worth around $2 billion in early 2025 and roughly $10 billion by the autumn. Twenty billion is the kind of mark that normally takes a decade. Chief executive Brendan Foody says the business scaled from $1 million to $3 billion in run-rate revenue in 27 months, which he called the fastest climb to $3 billion in history. Whether or not that claim survives scrutiny, the trajectory explains why the AI training layer has become the market's new obsession, a shift Blumefield has tracked across this year's funding frenzy.
From a Job Board to the Economy's Brain
Mercor did not start life as an AI training company. Founded in 2023 by three Thiel Fellows in their early twenties, it began as a marketplace for matching specialized workers with employers. The pivot came when the frontier labs discovered they had exhausted the easy supply of internet text and needed something harder to find: expert humans who could grade a model's answers, write new problems it had never seen, and demonstrate how real professional work actually gets done.
That demand turned a modest recruiting startup into critical AI training infrastructure. Mercor now says its network spans more than five million experts getting paid to train frontier models on their specialist knowledge, covering more than 800 distinct occupations in the United States alone. A radiologist, a tax attorney and a derivatives trader can all earn by teaching a model the judgment their field demands. The company has become, in effect, a labor exchange for the thing the models still lack.
The economics are extraordinary and a little unnerving. Foody says Mercor already spends three times as much on inference as it does on employee salaries, a ratio that captures how quickly machine work is displacing human hours even inside a company built on human expertise. His conclusion is that the same shift, repeated across the economy, points toward something closer to double-digit annual growth over the next five years. It is a grand claim, and it is the thesis the $20 billion price tag is really underwriting.
Why the Labs Keep Paying
To understand the raise, you have to understand the bottleneck it addresses. The frontier labs can buy more chips and raise more capital, but they cannot easily manufacture expert human judgment at scale. As models saturate the standard benchmarks, the hard question is no longer whether a system can pass a bar exam, but whether it can do a specific job, to a specific firm's standard, without supervision. Answering that requires people who actually do those jobs, and that is the scarce resource Mercor has cornered.
This is where the AI training business has quietly split in two. The first generation of data vendors, led by Scale AI and Surge, built their fortunes on labeling images and rating chatbot responses. Mercor is chasing the higher-value tier, the expert evaluations and post-training work that determine whether a model is merely clever or genuinely useful. Meta's near-$15 billion move for a stake in Scale AI earlier this year signaled how strategically the giants now view the AI training layer, as Blumefield detailed in its reporting on the data-labeling economy.
Mercor's pitch to the labs is that frontier work and enterprise work draw on the same machinery. The expert network, the evaluation methods and the post-training infrastructure it built for cutting-edge labs are the same tools a bank or a hospital will need to make a general model its own. Foody frames the division of labor bluntly: model builders and compute providers supply the foundation, and Mercor supplies the data and evaluation that turn a raw model into something that can do economically valuable work.
The Enterprise Land Grab
The most important part of the announcement was not the number but the direction. Mercor is pushing out of the frontier labs and into the Fortune 500, betting that every large company will soon need to measure, price and improve its own fleet of AI agents. The company's stated playbook is to build public evaluations, post-train an open model to cut costs, then train bespoke models for individual customers. It is a move that turns the AI training supplier into a strategic partner.
To do it, Mercor has been buying capability. It acquired Deeptune earlier this year, folding in a team that spent two years recreating hundreds of enterprise applications where AI agents can practice professional workflows before they are let loose on real systems. The logic is that an agent needs two kinds of knowledge, how a task is generally done and how a particular company does it according to its own policies and quality bar. Without both, the company argues, an agent fails in production.
Foody has staked out an expansive vision to match the valuation. "Every job in the economy will need its own agents that can perform tasks reliably and efficiently," he wrote in announcing the round. Fresh offices in New York and London, added to the San Francisco base, show where the enterprise land grab is pointed. If he is right, custom AI training for individual companies could dwarf the data work that built the business.
The Bubble Question
None of this settles the argument raging across the market. The same week Mercor sealed its raise, Foody wrote on X that the company had reached $3 billion in run-rate revenue, yet skeptics note that run-rate is a flattering measure and that a $20 billion valuation still implies a rich multiple on a business exposed to the spending plans of a handful of labs. If the frontier labs slow their outlays, or decide to bring AI training in-house, Mercor's growth curve could bend sharply.
The deeper risk is philosophical. Mercor's entire premise is that human expertise remains the gating factor in AI progress, and that the premium on it will rise. The counter-case, argued by some of the labs themselves, is that models will increasingly generate and grade their own training data, shrinking the need for armies of human experts over time. Mercor's three-to-one ratio of inference spending to salaries is cited by the company as proof of its own momentum, but a cynic could read it as a preview of the automation that eventually comes for its own suppliers.
For now, the market has made its choice. In a month when fear of an AI bubble wiped hundreds of billions from chip and infrastructure stocks, the capital still flowing freely is the capital chasing the AI training layer. That tells you where investors believe the next scarce resource lies. Whether human intelligence stays scarce long enough to justify a $20 billion bet is the question the next year will answer.
By Blumefield | November 11, 2026