Mercor co-founder and co-CEO Brendan Foody say di biggest thing wey dey hold artificial intelligence back no be say models weak, na say nobody don sabi how to measure wetin dem no fit do yet. Foody tok for MTS podcast say lab RL dey become powerful well well, and once dem get evaluation for something, models fit saturate am. Mercor just raise $750 million in Series D, and na company wey dem describe as $20 billion.
Mercor business na to supply data and evaluations wey frontier labs and top application companies dey use find weak spots for their models. Foody tok say di company move from near zero to about $3 billion in annual revenue in 27 months.
Foody frame di new capital as money to expand customer base, no be to build new capability. Mercor existing business dey sell data to frontier labs and top application-layer companies. Di new money go towards bringing di same technology to mainstream enterprises wey want build their own evaluations and post-train their own models.
He describe adoption as two-step. First na to build evaluations to establish ground truth for how model performs on specific work. Second na post-training, tuning di model so e do that work better. Top AI startups don reach phase two; he name Harvey, Sierra and Cognition as companies wey get post-trained models in production wey meaningfully affect gross margins. Most classic enterprise software companies, by contrast, don complete or nearly complete phase one, and dem still dey experiment with post-trained models wey carry large share of production workloads.
Foody tok say money spent on inference na wetin dey push company to move. For pretty much every major agent deployment wey dey run above $10 million a year in inference, companies dey start build evaluations to decide which model to use, whether e work in production, and how to optimize cost. Evaluation no be academic exercise; na procurement tool.
Di pattern dey show for benchmarks Mercor co-build with companies wey dey deploy agents. Apex benchmark launched February 2026 when frontier models score about 23%. Dem now score roughly 80%.
