Mercor Shows Move to Expert Labor

Mercor’s jump to a $3B gross revenue run rate shows that AI training has moved beyond cheap clickwork into a market for scarce expert labor. Labs are now paying for doctors, lawyers, PhDs, and engineers to grade model outputs, write hard prompts, and build evals for reasoning tasks, which makes the bottleneck less GPU supply alone and more access to vetted humans who can teach models expert judgment at scale.1 2 3 4

• The work has shifted from low cost overseas annotation toward high skill review. Earlier RLHF often used roughly $1 per hour contractors to rank outputs and flag unsafe content. More recent reasoning models need experts who can judge math, legal, medical, and science answers, often at far higher hourly rates.2 3
• Mercor’s model is concrete. It vets experts with AI video interviews and tests, then routes hourly contracts from labs and tech companies to those workers. Gross revenue grows fast because it counts total customer spend before contractor payouts, similar to other labor marketplaces where 60 to 70% can pass through to workers.1 2 4
• This demand is broadening, not narrowing. The same expert network and task routing systems used for LLM post training are spreading into model evals, enterprise agent testing, and even robotics data collection, where companies like micro1, Turing, Handshake, and Mercor compete for the same AI training budgets.1 4

The next phase is a fight to own the expert supply graph. Companies that can verify credentials fastest, keep quality high, and expand from text labeling into evals, agents, and robotics will capture a larger share of AI development spend, because frontier models will keep needing specialized human judgment even as the models themselves improve.1 3 4