Mercor pays experts $5M/day as it bets big on RL environments for vertical models

ccerrato147 · x · 2026-11-11

A long thread by ccerrato147 breaks down Mercor's business model and thesis:

Expert data flywheel: Mercor claims a network of 5M+ experts; its Studio software helps them build training data, with an AI project manager coordinating and checking quality.
Spending exploding: CEO Foody says expert payouts now exceed $5M per day, vs about $1M for an entire year two years ago — expert feedback gets more valuable as models improve.
RL environments: To let AI practice inside simulated workplaces (tasks + grading rules), Mercor acquired Deeptune, whose team built hundreds of simulated apps (including Salesforce) forming reinforcement learning environments.
Vertical post-training case: Harvey (legal AI) post-trained an open model on expert data from Mercor and others; without customer data, it reports nearly 2x completed tasks vs the base model on a held-out legal benchmark.
Business logic: Enterprises view procedures and expertise as competitive advantage and don't want them training a vendor's general model; at sufficient AI usage, specialized-model training pays for itself — rising usage creates more customers for Mercor even as inference gets cheaper.

More from coding & agent