Brendan Foody on Mercor, AI Data and the Future of Work

Mercor co-founder and CEO Brendan Foody joins Harry Stebbings to discuss AI training data, enterprise agents and venture financing. Stebbings discloses a small investment in Mercor; Foody's assessments of the company, its customers and competitors come from its CEO.

Stebbings asks about a security incident and rumors that revenue stalled. Foody confirms an incident but disputes the other claims, saying Mercor engaged Mandiant and other security advisers, communicated with customers and added about $300 million in net new annual recurring revenue in 60 days. These are company-side statements, not an independent incident or financial audit.

Foody recalls learning about the incident on a Saturday and prioritizing containment, understanding what was accessed and communicating with customers and experts. He says Mercor subsequently made security a company value; that response does not establish that future incidents cannot occur.

He believes online discussion exaggerated the impact and says some commentary came from people with interests in competitors. He disputes an allegation about data being accessed from China. The interview does not independently establish the incident's cause, scope, attribution or the motives behind individual comments.

Foody expects demand for AI security tools to grow and says the attacker in Mercor's incident used coding agents. He argues that automated tools can let attackers review more code quickly while defenders can also use AI; the claim about this incident is his account, not a forensic finding presented in the episode.

Asked whether Mercor lost OpenAI and Meta as customers, Foody says the OpenAI relationship continues and has strengthened. He says the Meta relationship remains paused and declines to discuss specifics, adding that Meta's relationship with Scale is a factor he considers. The customers do not independently confirm these descriptions in the interview.

Foody denies that Mercor extended offers to Micro1 employees with multimillion-dollar signing packages. He says an employee contacted people with possible bonuses, including one message mentioning $500,000, but no offer letters followed. This is his response to a reported allegation, not independently verified hiring records.

He denies a rumor that Amazon tried to acquire Mercor for $13 billion and declines to discuss other acquisition matters. Asked about a hypothetical $30 billion offer, he says he would not sell because he believes independence gives the company the best chance to pursue its mission; no such offer is established.

Foody expects more jobs to exist in ten years but acknowledges substantial displacement during the transition. Stebbings questions whether historical adjustment can happen quickly enough as AI capabilities advance; both are discussing uncertain economic outcomes.

Foody describes Mercor's Apex evaluation as an effort to measure which tasks across professions models can perform. He cites progress on that evaluation, but the interview does not provide its methodology, independent validation or evidence that a score translates directly into the automation of whole jobs.

He argues that productivity can create demand for new work and says Mercor pays out more than $3 million a day to its talent network. He predicts that amount could triple or quadruple in a year; the payout is a company claim and the future figures are forecasts.

Foody expects training and managing agents to become a major category of work. He imagines employees encoding recurring customer-support or legal workflows, while acknowledging that organizations also need to capture tacit knowledge that currently resides in people's heads.

He thinks increasingly capable models may help clean and structure data, while humans contribute organizational context. These are expectations about future capability; the interview does not show that sensitive data can be processed accurately or safely without appropriate controls.

On specialized data providers, Foody says niche firms can have value but argues Mercor benefits from aggregation, reusable tooling and a talent network he says exceeds five million people. His network-size and customer-preference claims are not independently verified.

Foody anticipates consolidation if markets correct. He says Mercor has more than $500 million in cash and is very profitable, describing a history of limited cash burn; no financial statements are presented to substantiate those figures.

Asked whether reported revenue is merely gross merchandise value, Foody says Mercor reports revenue with a gross margin generally between 30% and 40%. He explains that the firm recruits experts, provides tooling and coordination, and delivers completed tasks. That is management's description, not an accounting opinion or audited presentation.

Foody says quality checks can inform which experts to recruit and claims a minority of tasks can produce much of a dataset's value. His illustration of 2,000 valuable tasks out of 10,000 is a heuristic, not an independently measured rule.

He names software engineering, finance, medicine, law and consulting among areas of demand. He expects labs to seek longer, collaborative tasks for training and evaluation; his timeline for model capabilities is a prediction, not evidence that such tasks are already reliably automated.

Foody says Mercor can mobilize experts in many fields but values people who combine domain knowledge with familiarity with frontier models. He believes labs need a broad distribution of realistic workplace context and tasks; this reflects the demand he says the company sees.

Foody recounts the financing history: he says the September 2023 seed round raised $2.3 million at a $23 million post-money valuation, when revenue run rate was near $1 million. The amounts and timing are his recollection, not verified financing documents.

He says the Series A was priced at a $250 million post-money valuation, with revenue run rate around $2.5 million when a term sheet arrived. He recalls ambitious internal projections and says the business exceeded them; these are management's descriptions of past performance.

Foody describes meeting Felicis and receiving a term sheet at a $2 billion valuation when he says revenue was around $20 million. Stebbings discloses that he invested a small amount at that stage. Foody regards the round as the most uncomfortable price relative to then-current revenue, though he calls it successful in hindsight.

He says the business later reached roughly $400 million in revenue run rate and received another term sheet at $10 billion. Asked about the next round, he says there are higher offers but the company is profitable and taking time to choose a partner. Future terms and returns are not established.

Foody predicts the next year will favor infrastructure businesses serving model labs over application companies built on those models. He argues frontier providers can add applications, while data networks and compute can develop advantages. Mercor itself sells infrastructure to labs, so this is a commercially interested thesis.

He says Mercor could serve more demand if it had the capacity and speculates it might raise prices without a large demand impact. He also says the company considers long-term customer and expert relationships, not just short-term pricing. Neither elasticity estimate is tested in the interview.

Stebbings discloses investments in application-layer companies and argues that legal software can be defensible through workflow depth, sales, adoption and customer service. Foody counters that models are becoming more capable and application software may be easier to recreate; their disagreement is a competitive forecast.

Foody predicts agents will be able to build complete SaaS applications within a year. He says network effects, such as integrations or connections across organizations, may nevertheless protect some incumbents. These are hypotheses about technology and competition, not evidence that all SaaS products can currently be replicated.

He distinguishes pre-sale distribution from post-sale deployment. In his view, helping a customer capture tacit knowledge and adapt agents can be more defensible than selling software alone; this aligns with Mercor's own services and data business.

Foody sees opportunity in AI-enabled services but warns that a company must actually gain efficiency from AI. He says Mercor has about 100 to 150 people in delivery and describes a first project managed end to end by an AI project manager. The reported experience and automation are company claims, not independently evaluated results.

Foody argues that improved price-performance can increase total demand for models. He says Mercor currently spends more on tokens for internal agents than on employee salaries, a striking company-specific claim for which the interview supplies no expense breakdown.

He describes agents for interviews, candidate ranking, accounting and fraud detection, each with an evaluation to choose models and manage price-performance. He says Mercor has conducted more than five million interviews; that figure and the quality of the evaluations are not independently verified.

Foody predicts enterprises will increasingly evaluate models against their own workflows and switch among providers. He expects API-level competition to intensify while user routines and application workflows can remain sticky; switching costs and model comparability can differ by deployment.

He guesses that in five years an average enterprise might spend more on compute than headcount. This is a long-term forecast, not a measured industry trend, and depends on uncertain prices, staffing, adoption and what counts as compute or labor expense.

Stebbings questions whether evaluations capture real work. Foody agrees that academic benchmarks may not reflect business outcomes and argues for testing complete workflows involving colleagues, documents and deliverables. An evaluation still requires careful design and validation to support such claims.

The two discuss startups using frontier models as references and adapting open models to lower costs. Foody predicts both greater overall demand and stronger open-model competition, and expects much inference eventually to use open or customized models. He also considers leading labs attractive investments; these are forecasts, not advice.

Asked for valuations in five years, Foody speculates that at least one leading lab could exceed $10 trillion. He sees continued importance for Nvidia but expects more chip competition. These highly uncertain valuation and market-share guesses are not facts or recommendations.

Foody worries about unequal distribution of AI's gains, even while arguing that allocating compute to highly productive firms can be efficient. He says the societal implications of concentration still need to be managed; he does not provide a policy model proving a particular distribution.

He proposes eliminating income tax for the bottom half of Americans and considering other revenue sources, including capital gains or carbon. Stebbings objects that higher capital-gains taxes could discourage investment or encourage relocation. Foody agrees migration effects require analysis; their discussion is a policy debate, not tax advice or a verified revenue estimate.

Foody describes meeting successful founders and realizing many are approachable peers. He says seeing friends build companies expanded his sense of what he could attempt; it is a personal lesson rather than a claim that outcomes are equally accessible.

On Europe's model ecosystem, he expects talent and capital concentration at existing labs to be difficult to reverse, but sees possible value in post-training, applications and localization. These are his competitive and geopolitical predictions, not a conclusion about any region's inherent capability.

He says localized data can make models more useful for different legal jurisdictions, while predicting large labs will also invest in that work. His example is about potential model adaptation, not evidence that a model is qualified to provide jurisdiction-specific legal advice.

Foody says competition for researchers is intense and recounts a candidate he understood had an approximately $20 million annual offer from Meta, later clarifying it could be liquid stock. He also says top researchers may receive tens of millions in stock per year. The anecdote and compensation figures are unverified and do not describe typical salaries.

He expects unusually high pay to persist for a small group while a broader supply of experienced people may develop. The magnitude and timing of that adjustment are uncertain.

Foody describes Mercor growing from about 40 people to several hundred and says finance and legal support have helped management keep up. These headcount and scale comparisons are his estimates, not independently audited company data.

Contrary to the host's criticism of HR, Foody says people functions and clear feedback become important as a young team scales. He says Mercor has not mandated a 996 schedule, while describing his own and other leaders' long working hours; this is his account of company policy and culture.

In the quick-fire round, Foody says he would like Mercor to go public in the next few years but is not rushing this year or next. That is an aspiration rather than an announced offering or timetable.

He says revenue growth at frontier labs has increased his conviction in their long-term importance, but declines to choose between OpenAI and Anthropic as an investment. He also praises a competitor's research focus; the rankings and market shares he offers are his personal judgments.

Foody credits people in the Prod community for advice, early business and working-capital support when Mercor was young. He says they did not take equity, and closes by acknowledging how many circumstances could have led the company to a different outcome.