Mercor scaled from $1M to $500M ARR in 17 months — now targeting decacorn status

Summary

{
  "long_summary": "Mercor scaled from $1M to $500M ARR in 17 months — a run that CEO and co-founder Brendan Foody describes as surreal but deliberate. The company started when Foody and his high school co-founders were 19, reached a $2M run rate, and dropped out of college before locking in the AI lab contracts that drove the acceleration. The team is now 250 people across the US, India, Latin America, and the UK, and Foody says a decacorn valuation by end of year is the next target.\n\n**From Olympiads to professional domains**\n\nMercor's early work centered on sourcing extreme technical talent fast. An early signal was turning around 25 Olympiad math medalists in 24 hours to help labs train models on competition-level reasoning. Referrals drive the majority of sourcing — workers already on the platform earn $250 for successful introductions.\n\nThe bigger shift now is away from academic benchmarks like Olympiad math or GPQA toward professional domain evals. Foody says the hard question labs are now wrestling with is how to measure what it means to be a great software engineer building real products, or an investment banker doing financial analysis, or a consultant segmenting a market. These are harder to score than a competition result, so Mercor's model involves human experts defining rubrics and success criteria that models then optimize against — what Foody calls the defining trend in frontier model training over the next year or two.\n\n**The structure of the work**\n\nTraditional thumbs-up/thumbs-down RLHF is still happening across labs but Foody sees it migrating toward data flywheels built from real-world product usage. Where human experts remain indispensable is in tasks that require five or more hours of structured thinking to define how model success should even be measured — something that can't be reliably crowd-sourced from product users.\n\nFoody expects agentic trajectory lengths to extend dramatically, from single-screen tasks to workflows that would take a human 30 or 90 days, spanning 10 tools and multiple internal stakeholders. Building the environments to eval those trajectories is, in his view, the frontier problem. For browser-use tasks like booking flights, he expects simulated environments with unit tests to close the gap within two to three years. The longer tail — building a startup, running a knowledge-work process end to end — is where the harder evaluation challenges live.\n\nOn problems with very long feedback loops, like testing the 50-year health impact of a drug, Foody is skeptical that simulation can substitute for empirical analysis. His view is that models will more likely scan historical data on outcomes rather than simulate physiology — a meaningful constraint on where AI-driven scientific discovery can compound quickly.\n\n**Operations**\n\nMercor never formally mandated working hours, but the founding team defaulted to seven days a week. As the company has grown, Foody says the focus has shifted from input orientation — hours in the office — to output orientation, while still hiring for high intensity. The $40M seed round he mentioned in passing was raised when the company was at a fraction of its current scale, at which point Foody publicly called for unicorn status by year-end and hit it. The decacorn call for 2025 follows the same playbook."
}