Mercor, an AI startup that supplies training data, evaluations and post-training work to model developers, said on Tuesday it had raised $750 million in a Series D round that values the company at $20 billion. The round included an employee tender and participation from funds managed by affiliates of Apollo, according to sourcery.vc.

The valuation rests on unusual growth. Mercor said it scaled to a $3 billion run-rate revenue in 27 months, and that more than five million experts are paid through its platform. CEO Brendan Foody framed the milestone publicly as the fastest such climb by a software company.

But the more interesting shift is in what Mercor actually sells. Analysts close to the company's positioning argue that Mercor is not a labor marketplace — the model it launched with — but a supplier of reinforcement-learning environments and evals. In that reading, its product is the scaffolding that teaches and grades frontier models: the graded tasks, the human preference data and the test harnesses that determine whether a model is good at coding, research or reasoning.

That distinction matters for how the business is priced. If Mercor sells labor, its revenue is services revenue — linear, headcount-bound, and easier to undercut. If it sells evaluation infrastructure and RL environments, the revenue looks more like software and data licensing, with better margins and more durable lock-in as labs keep needing fresh benchmarks and post-training signal.

The round's structure is also a signal. An employee tender lets early staff cash out some equity at the new price without waiting for an exit — a private-market practice that has become common as valuations balloon and IPO windows stay uncertain. Apollo-affiliated funds joining alongside traditional venture investors points to institutional capital treating late-stage AI infrastructure as a credit-like, cash-flow-backed asset rather than a pure venture bet.

Mercor's challenge now is the one facing every data-and-evals company: staying ahead of the models. When frontier labs can generate their own synthetic training data and design their own evaluations, the value of an outside supplier depends on offering something labs cannot easily replicate — human expertise, adversarial task design, or breadth of real-world environments. At a $20 billion valuation, that question is no longer academic.