Mercor CEO Brendan Foody on Expert Data, AI Evaluations and Growth

Harry Stebbings interviews Mercor co-founder and CEO Brendan Foody about expert labor for AI, model evaluation, rapid growth and company building. Foody has a direct interest in Mercor’s claims. Stebbings says he made a small investment in the company, which is relevant to his praise and questions.

Foody recalls selling doughnuts in middle school, buying them in bulk and changing prices when a competitor appeared. He says his mother worried he would get into trouble and sent him to a Catholic high school, where he met his future co-founders. The story is his account of an early interest in selling.

Asked whether he always expected success, he says he had ambitious goals but could not have predicted the scale or speed of Mercor’s growth. He recalls selling sneakers and helping resellers obtain AWS startup credits, saying he earned hundreds of thousands of dollars in high school.

Foody says online information made him question the educational value of college, although he enjoyed its social value. Stebbings says he also left university early. Their experiences are personal views and do not establish the value of college for other students or careers.

Stebbings cites a competitor’s characterization of the sector as a “body shop.” Foody rejects it, saying Mercor acts as a research partner that finds and vets experts for AI labs. That is the company’s account of its services, not an independent comparison with competitors.

He says Scale AI once used Mercor to hire thousands of people, giving the founders a view of a shift from general crowdsourced data toward more complex work by specialized professionals. He gives examples including software engineers, doctors, lawyers and finance professionals. His descriptions of other providers and their workers are competitive claims.

Stebbings asks whether increasingly capable models will make the supply of useful experts ever narrower. Foody agrees that people may stop being able to contribute to a particular task as models improve, but says adding tools and longer workflows can create new tasks where humans identify failures. His example is an account of one project, not evidence about all model tasks.

Foody does not think models are plateauing, though he sees a change in how capabilities improve: more curated, complex data and reinforcement-learning environments. That assessment and his prediction of continuing progress matter commercially to Mercor, which supplies related work.

On differentiation, he argues that a minority of contributors can produce a large share of improvement and that sourcing, referrals and matching experts to tasks can create an advantage. The suggested contribution distribution and his claim of defensibility are company views, not independently measured results.

Stebbings cites a criticism that competitors cannot measure the quality of data they produce. Foody disputes it, saying Mercor uses models and algorithms and tests data against model performance. Neither the interviewer nor Foody supplies an independent evaluation comparing vendors.

Asked whether AI labs deliberately distribute spending among vendors, Foody says some do, but claims customers may concentrate work where they see the best results. He expects fragmented markets to consolidate because matching infrastructure and expert networks have fixed costs. His account of customer behavior and future consolidation is not verified in the interview.

Stebbings asks about customer concentration. Foody declines to give the largest customer’s exact percentage and says the distribution is relatively similar to Nvidia’s. He argues serving important customers can still build a valuable business; the comparison does not disclose Mercor’s actual concentration or eliminate the associated risk.

Foody says demand accelerated after the Scale transaction and that Mercor had already reached a nine-figure revenue run rate before its business quadrupled. He reports growing from $1 million to $500 million in revenue run rate in 17 months. Run rate annualizes a current pace; it is not $500 million of realized annual revenue, profit or audited cash receipts.

He calls that the fastest revenue growth and compares it with Cursor. The interview does not provide a consistent, independently audited cross-company comparison, so the ranking remains his claim. He credits existing relationships with frontier labs as well as the Scale transaction for positioning Mercor to expand.

Stebbings relays criticism of Scale’s quality. Foody praises its founder’s distribution and sales abilities but says Scale lost focus on product quality and treatment of contributors. These are a competitor’s allegations and opinions; Scale’s detailed response and comparative quality evidence are not presented.

Foody reports average marketplace pay of $95 an hour and claims Scale and Surge generally pay about $30 an hour. The interview does not provide comparable samples, dates, job mixes or payment methodology. The figures should not be read as independently verified averages for the competing platforms.

Stebbings asks whether synthetic data could displace human data work. Foody expects synthetic augmentation but argues that people are still needed to define and measure tasks humans can perform better than a model. He believes humans will contribute to training in ten years; that is an uncertain forecast aligned with Mercor’s business.

Both criticize a gap between academic benchmarks and practical enterprise work. Foody argues that an impressive exam score does not establish that a model can build a financial model, research deck or web application reliably. His examples illustrate a measurement concern, not a comprehensive comparison of all benchmarks.

He would design evaluations that reflect real workflows and grade how models use tools, for example through a rubric for investment research. He says reducing the gap between real work and simulated tests can make model improvement more relevant to users. The effectiveness of a specific evaluation would need testing.

Stebbings asks how Foody has grown as a leader and discusses prior funding valuations. Foody says investors should consider a company’s possible future rather than only revenue multiples. He recounts run rates at fundraising and says the company later grew substantially; valuation and run rate do not themselves establish investor returns.

The host discloses a small investment and asks about the timing. Foody recalls a period averaging roughly 54% month-over-month growth and says revenue could have changed significantly between conversations. This is a company-reported historical growth rate, not a forecast that such growth will continue.

Asked about a rumored $10 billion valuation, Foody says the company had received offers and could raise soon, partly to signal market leadership. He does not confirm a completed round at that valuation. A prospective valuation, financing and eventual investor return are distinct.

Foody says Mercor is investing aggressively yet remains profitable and that additional cash would not necessarily change its investment pace dramatically. He sees value in a strong balance sheet. The interview does not provide audited statements or a reconciliation for the profitability claim.

On a potential IPO, he says he has not thought about it much and recalls Jack Dorsey advising the company to stay private as long as possible. Foody values the ability to pursue long-term objectives without quarterly public-market pressure. This is a preference, not a listing timetable.

Asked whether private markets have too much cash, Foody says investors may see excess competition and that some weak competitors receive substantial funding. He believes short-term exuberance can coexist with long-term opportunity, invoking the late 1990s. His view of future returns is speculative.

Responding to a study mentioned by Stebbings, he says academic scores and enterprise outcomes can diverge, and that teams should define and evaluate the work they expect AI to perform. The interview does not establish a general failure rate for all enterprise AI projects.

For application revenue, Foody suggests examining retention and speaking with customers rather than relying on early pilots alone. He contrasts failed pilots with strong retention and customer enthusiasm as possible signs of product fit. These are investment heuristics, not a guarantee of returns.

On margins, he says Mercor has strongly positive gross and net margins, while arguing that low margins may be defensible if future efficiency improves and customers stay. He warns that subsidies are riskier in markets with low switching costs. His company’s margin claim is not accompanied by financial statements, accounting details or exact figures.

He is less concerned about broad AI capital expenditure over a ten-year horizon, while acknowledging exuberance and saying individual projects need positive returns. This is a long-term investment judgment, not evidence that all infrastructure spending will pay off.

Foody says Mercor engineers use tools including Cursor and Claude Code, with usage changing as products improve. He sees low switching costs and intense competition in coding tools, while noting vendors are trying to build defensibility through customization and user data. His descriptions of other companies’ margins are not independently verified.

Asked whether his company will have more or fewer engineers in five years, he predicts more: greater productivity could make it practical to build much more software. This is a conditional forecast about his company and demand, not evidence that AI will raise engineering employment everywhere.

Foody thinks the largest model builders may already exist because data, compute and research teams are expensive, but explicitly leaves room for new breakthroughs and startups. He praises smaller Gemini models and says his thinking has shifted toward a mix of general and specialized systems.

On sovereignty, he sees possible opportunities for models specialized in a region’s law or other local needs, but expects the biggest companies to serve broader use cases. His assessment of which model providers will win is a forecast, not a settled outcome.

Stebbings describes Mercor as supporting a 996 culture. Foody clarifies that the company never mandated hours: he says 996 described the early team’s work and that some people worked even longer. He still values intensity, while saying the company increasingly focuses on output and talent rather than time physically present. This is management’s account of its culture.

Asked what he would do if less afraid, Foody wonders whether Mercor should burn much more cash to subsidize marketplace supply or customer projects. He says the company has prioritized capital efficiency and claims it has enough demand to double if it could meet capacity. That demand claim is not a demonstrated revenue forecast.

Stebbings, an investor in Mercor, says he might use cash to subsidize a business under intense competitive pressure, but not necessarily otherwise. Foody says he and investor Peter Fenton favor capital efficiency and fundamentals. Their discussion concerns strategic trade-offs, not an instruction that subsidies will work.

Foody says Mercor turns down projects and is focused on expanding its ability to serve customers. He says he has sold a small amount of stock in secondary transactions and wants any such activity not to distract from the long-term mission. His statements are not independently verified transaction data.

In the quick-fire round, he rejects the idea that superintelligence better than humans at everything will arrive in three years. He recommends asking whether much better models in one or two years would improve or weaken a startup’s business. Both comments concern uncertain future capabilities.

Foody considers customization attractive and says general foundation models may also become large businesses. He praises OpenAI’s consumer position and Anthropic’s coding and enterprise position as he sees them. These are competitive observations, not established future market shares.

He says he wishes he had focused on foundation-model labs earlier when Mercor began in 2023. Asked about market share, he offers a rough estimate of 50% to 60% for reinforcement-learning environments, while distinguishing them from other human-data categories. The interview supplies neither a market definition nor an independent measurement.

Foody reports that some lab executives expect reinforcement-learning environments to spread across much of the economy, with people designing workflows that models can learn. He shares that optimism, but the pace, scope and employment effects are speculative. Mercor has a direct commercial interest in this market expanding.