A three-year-old startup run by a 23-year-old is in talks to double its worth to $20bn. The numbers are dazzling. Read the footnotes.
Mercor, the AI training marketplace, has told investors it can raise at a $20bn valuation. It says it already holds at least one term sheet at that price, Bloomberg reports. The talks are early, and a deal might not close.
The figure is striking for its speed. Mercor last raised in October, at a $10bn valuation. Nine months later, investors are being asked to pay double.
The company recruits domain experts, such as engineers, lawyers and doctors, and pairs them with AI labs that need human judgment to train frontier models. Its customers include OpenAI, Anthropic and Google. Its founders are three former high-school debate teammates who dropped out of college.
They became the world’s youngest self-made billionaires at 22.
The buy that raises eyebrows
On the same day, Mercor announced it is buying Deeptune, a startup that builds the simulated software where AI agents practise real work before deployment. Think of a flight simulator, but for an agent learning to use a spreadsheet or Salesforce. Mercor confirmed the deal on its blog.
Terms were not disclosed.
There is a wrinkle. Andreessen Horowitz led Deeptune’s $43m funding round in March. One of the angel investors in that same round was Brendan Foody, Mercor’s chief executive. Three months later, his company bought it.
Foody was candid about the sequence. The angel cheque was written with a purchase already in mind, he told Fortune: “It was in a lot of ways the main motivation, actually.”
Public reporting has not established whether Mercor’s board or outside investors reviewed his personal stake before the deal closed. That is the question a $20bn price tag asks the board to answer first.
The revenue number needs an asterisk
Foody says Mercor’s annualised revenue run rate has crossed $2bn, doubling in four months. The headline is real. The footnote matters more.
That $2bn is gross billings, not what Mercor keeps. Contractors take home 60 to 70 percent of everything billed, first reported by The Information. Net of the experts who do the work, Mercor’s own revenue is closer to $600m to $800m. At a $20bn valuation, that implies a multiple of roughly 25 to 33 times net revenue. Aggressive, though not unheard of for growth this fast.
The breach it is trying to forget
The valuation talks are more remarkable given where the company was in spring. In March, a supply-chain attack on an open-source library called LiteLLM exposed up to four terabytes of Mercor data. Meta, a big customer at the time, paused all work with the startup indefinitely.
Class-action lawsuits followed. The hacking group Lapsus$ claimed responsibility.
Mercor says the impact was “very limited”. Foody speaks about it as the past, noting that OpenAI and Anthropic stayed on and that revenue doubled in the months since. That is either a display of customer loyalty, or a sign that the labs that need training data at this scale have nowhere else to go.
Why it matters
Mercor now wants to own the whole training stack: the environments where agents practise, the experts who grade them, and the benchmarks that define success. Rivals crowd in behind it, from Scale AI, worth around $29bn since Meta took a stake, to Surge AI, reportedly raising near $25bn.
The market is enormous, and the money is chasing it hard, from rival environment builders to the soaring paper valuations of the labs Mercor serves.
Whether $20bn is a fair price or a bubble marker, sceptics keep warning about froth across AI. Mercor’s answer is growth. Its investors will have to decide whether the growth, once you strip out the contractors, the breach and the founder’s own bet, is really worth twenty billion dollars.
Or whether, like OpenAI’s own long wait to list, the number is running ahead of the proof.
