Mercor has raised $750 million at a $10 billion valuation, according to reporting this week, making it one of the most highly valued companies in the business of sourcing human expertise for AI training. The round underscores a quiet reality of the model economy: as systems become more capable, the data required to improve them becomes more specialized, not less.
The company connects model developers with domain experts, including lawyers, doctors and engineers, who produce and evaluate training examples. That is a different business from the click-work annotation that defined the first wave of data labelling.
Why expert labour got expensive
Early training data could be collected at scale by non-specialists following simple instructions. Reinforcement learning from human feedback, and more recent techniques that require long-form reasoning traces or code review, demand people who can tell a correct answer from a plausible one.
Scarcity drives price. A qualified physician’s hour is worth more than a generic annotator’s, and the volume of expert feedback needed by frontier labs has grown faster than the pool of people willing to provide it.
The durability question
Every data-labelling company faces the same strategic concern: if models become good enough to generate their own training signal, the human market shrinks. Mercor’s counter-position is that evaluation of hard problems will remain human-led longer than generation, and that the highest-value tasks are precisely the ones that resist automation.
That argument is plausible but unproven at this valuation. The company is priced as if the expert-feedback market expands for years.
Signals in the deal
Large rounds in data services often precede either consolidation or a squeeze, as model labs bring more of the work in house. The near-term indicator is customer concentration: a business that depends on a handful of frontier labs is exposed to their budgeting cycles and their internal tooling roadmaps in ways that diversified revenue is not.
