The Twenty Minute VCMercor CEO & Co-Founder, Brendan Foody: How They Grew from $1M to $500M in 17 Months
At a glance
WHAT IT’S REALLY ABOUT
Mercor’s 22-Year-Old CEO On Building A $500M AI Data Rocket
- Brendan Foody, 22-year-old CEO and co-founder of Mercor (called “Marqor/Macaw” in the convo), explains how the company scaled from $1M to a $500M revenue run rate in just 17 months by supplying elite human experts to train and evaluate frontier AI models.
- He argues the data-labeling market has shifted from low-cost crowdsourcing to highly vetted “Goldman/McKinsey/FAANG-level” talent building complex reinforcement learning (RL) environments that mirror real professional workflows.
- Foody pushes back on claims that synthetic data or plateauing scaling laws will eliminate the need for human experts, insisting that as long as humans can do things models can’t, there will be a massive market for expert-driven data and evals.
- The conversation also covers Mercor’s economics, valuation philosophy, competitive dynamics with players like Scale and Surge, broader AI investment froth, founder secondaries, work culture, and how evals must evolve to reflect real-world capabilities instead of academic benchmarks.
IDEAS WORTH REMEMBERING
5 ideasElite human expertise is now central to frontier AI training.
Mercor’s core thesis is that the market has moved from low-cost crowd work to sourcing and vetting top professionals (e.g., Goldman bankers, McKinsey analysts, FAANG engineers) who can design and execute complex RL environments that actually move the needle on model performance.
The most valuable data contributors follow a power-law distribution.
Foody notes that in a 100-person project, 10–20% of contributors drive most of the model improvement—similar to company value creation—so Mercor differentiates by identifying, attracting, and matching these “10x” experts to high-impact tasks.
Current AI evaluation benchmarks are largely misaligned with real utility.
He argues that focusing on Olympiad math or ‘humanity’s last exams’ is “wholly disconnected” from what enterprises care about; future evals must measure performance on realistic workflows (e.g., building a real financial model, writing research decks, or operating tools like email, calendar, and Slack).
Synthetic data will augment, not replace, expert human input for years.
While synthetic data can scale and cheapen some training, any capability where humans can still outperform models requires human-generated ground truth and verifiers; until superintelligence exists, humans remain essential to pushing the frontier of model capabilities.
Vendor consolidation will favor those with the deepest talent networks and matching intelligence.
Labs often start with multiple data vendors but tend to concentrate spend with partners that deliver the biggest model performance gains; Mercor bets that its referral-based supply, high pay rates, and sophisticated matching algorithms will make it a winner as the market consolidates.
WORDS WORTH SAVING
5 quotesWe scaled the business from one to 500 million in revenue run rate in the last 17 months, which is the fastest revenue growth of all time.
— Brendan Foody
The total addressable market is limited by the amount of things that humans are better at than models.
— Brendan Foody
Eval’s are bullshit when they’re about Olympiad gold medals and humanity’s last exam. They’re wholly disconnected from the outcomes that consumers and enterprises actually care about.
— Brendan Foody (paraphrasing and agreeing with Harry Stebbings)
Having phenomenal people that you treat incredibly well is the most important thing in this market.
— Brendan Foody
If we think about the model as the product, then the eval is the PRD.
— Brendan Foody
High quality AI-generated summary created from speaker-labeled transcript.
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome