20VC: Mercor: From $1M to $500M in 17 Months: The Fastest Growing Company in the World | How to Think About Margins and Revenue Sustainability in AI | Why Evaluation Benchmarks in AI are BS Today with Brendan Foody

Mercor's CEO says expert humans still power AI progress

Mercor is a fast-growing marketplace that supplies vetted human experts to train and evaluate AI models. Foody argues that as models improve, humans are still needed to define and measure the frontier, and a small share of top experts drives most of the gains. That makes expert quality and real-world evaluation the scarce inputs, and he says Mercor can pursue them profitably.

Takeaways

• Mercor hit $500M run rate in 17 months. Foody says revenue went from $1M to $500M run rate in 17 months, one month faster than Cursor. It was already at nine figures before the Scale AI acquisition, has quadrupled since, and could double overnight with enough capacity.
• Data vendors compete on expert quality, not bodies. Foody says the market moved from crowdsourced low-skill labeling to vetting experts such as bankers, consultants and doctors. Mercor's marketplace pays an average $95 an hour against about $30 at Scale AI and Search, and acts as a research partner.
• A few top contributors drive model improvement. Foody argues contributions follow a power law, with the top 10 to 20 percent of contributors driving most gains. That makes sourcing and measuring the best experts the core capability, and he says Mercor uses many models and algorithms to assess quality.
• Human experts stay essential as models improve. Foody rejects a model plateau but says synthetic data will augment, not replace, humans, who are needed to measure the frontier. He expects models to need humans in ten years because superintelligence remains distant, and RL environments may subsume much of the economy.
• Current AI benchmarks miss real enterprise outcomes. Foody calls benchmarks like humanity's last exam disconnected from enterprise results. He proposes rubric-based evaluation of real professional tasks, and says companies vibe spend on AI without product requirements, citing a 95 percent failure rate in enterprise cases.
• Customer concentration is high but expected to ease. Foody will not give his largest customer's share but says it resembles Nvidia's concentration. He expects customers to start with several vendors, then consolidate over time, as rivals like Turing, Handshake and Surge compete.
• Capital efficiency sets Mercor apart from AI peers. Mercor is profitable with positive gross and net margins, and is 25 times larger than at its Series B, which came at $20M run rate. Foody calls subsidised AI businesses with low switching costs risky and advises staying private as long as possible.