Mercor Raises $750M at $20B Valuation | CEO Brendan Foody
Topics
AI Training Data and Evaluation Benchmarks
Unique Human Value in Everyday AI
Battle for Frontier AI Moats
Enterprise Private AI and Real Compute Demand
Mindset and Psychology of Tech Founders
Pioneers
Brendan Foody
Summary
- Mercor’s expert payouts have risen from $1 million a year to well above $5 million a day in 24 months. Foody says cumulative payouts are nearly $1 billion, but the product is a complete RL environment rather than a labor-matching fee. Experts are “one ingredient, one moat” alongside application replicas, simulated work content and scored tasks.
- Mercor’s market durability depends on humans retaining capabilities that models lack. Foody says an hour of expert work with a co-pilot might produce three to five times the model improvement, expanding demand through “Jevons paradox.” His thesis remains conditional on models needing humans to measure the frontier of their capabilities.
- DeepTune fills Mercor’s main product gap by supplying the app replicas its experts need to populate. A customer initiated the partnership because Mercor was stronger in the “worlds” and “tasks” around those apps. The acquisition joins complementary capabilities; Foody calls it “a one plus one equals three kind of thing.”
- Foody sees brand, differentiated training data and customized models as stronger moats than software or GPU resale. Harvey illustrates domain-specific model building, while Fireworks illustrates customization in inference. “The switching costs of reselling GPUs are zero,” he says; Mercor switches APIs daily according to eval results, underpinning his concern that investors sometimes overrate API revenue quality.
- Enterprise inference bills are Foody’s leading indicator for a new wave of evaluation and post-training demand. At $10 million–$100 million in inference spend, he suggests roughly 10% should fund evals; above $100 million, post-training starts to justify substantial fixed costs. O’Shea challenges the wide range, but Foody points to Mercor’s almost $1 million a day in inference as evidence that companies can scale quickly.
- His enterprise thesis assumes companies will protect their intelligence rather than surrender competitive data to frontier labs. Foody says Anthropic and OpenAI “could very well be $10 trillion-plus companies,” while predicting every Fortune 500 will resemble a “Neo lab.” Goldman Sachs is his example: letting a lab train on its people, processes and documents would resemble giving those advantages to a competitor.
- Foody and Adarsh have had opportunities to cash out, but say that is not what motivates them; they would rather work “100 hours a week” on the problem. On Scale’s Meta arrangement, Foody says, “I don’t think we would ever do something like that.” He prioritizes customer and expert outcomes over weekly revenue graphs and doesn’t expect Mercor to go public in the next 12 months.
Deep dive
1. Mercor sells complete RL environments, with experts as one input
- Foody’s trillion-dollar ambition starts with changing how people contribute to the economy: repetitive support replies, contract redlines and podcast edits become work people teach agents to perform. Human contribution shifts toward creating evals, updating skills files and training agents. Mercor wants to define that new category, rather than merely improve hiring.
- The deliverable has three parts: “apps,” cloned versions of tools such as Google Workspace; “worlds,” the messages, slides, documents and spreadsheets inside them; and “tasks,” prompts plus grading rubrics. The model attempts a task, receives a score and learns. Experts populate the worlds and tasks, making them “one ingredient, one moat” in the product Foody says Mercor delivers to labs.
- Mercor’s three platforms support that output: a talent network with over five million active job seekers; Studio, where experts use an AI co-pilot, automated quality checks and agents to build environments; and an AI project manager. Foody compares its work to human project managers setting KPIs, evaluating performance and giving feedback; Mercor’s agents now autonomously run projects, hire people, determine data quality and are accountable for customer outcomes.
- DeepTune addressed the missing ingredient: app clones, including Salesforce, Adobe and Google Workspace. A customer initiated the partnership because Mercor was stronger in worlds and tasks. DeepTune had raised a Series A from Andreessen Horowitz; it became Mercor’s second acquisition, substantially larger than its first. Foody liked founder Tim and the New York team’s cultural fit: “a one plus one equals three kind of thing.”
2. Expert supply scales through interviews and reliable payments
- Referrals supply 70% of candidates and, Foody says, tend to produce the highest quality. Mercor has conducted over seven million interviews. Its first AI interviewer, built in April 2023 using GPT-4, had about seven or eight seconds of latency. His labor-matching insight was “scalable data collection”: hiring five Stanford AI PhDs might turn up a strong candidate, while finding world-class software engineers in a remote area of India might require interviewing hundreds or even 1,000 people. Referrals and personal projects are additional context Mercor analyzes.
- Payouts now run well above $5 million a day, versus $1 million a year 24 months earlier; historical payouts total nearly $1 billion. Foody also reports growth of “more than 5X” on a “mid-nine-figure base” since the previous September. He describes that acceleration as surreal, calling human model-training work “the fastest-growing job category in history.”
- The direct relationship with labs grew partly out of a payments failure elsewhere: Mercor had supplied another player in the space with over 1,000 workers, then received tickets from unpaid people struggling to feed their families. Foody wanted control of the expert experience and to overinvest in payments reliability. Stripe Connect is the backend; he says Mercor has had no major payout issues to date.
3. Expert demand lasts as long as human capabilities stay ahead
- Foody says an AI co-pilot might make an hour of expert effort yield three to five times the model improvement. Instead of shrinking labor demand, he sees “Jevons paradox”: greater productivity unlocks more teaching because “99% of knowledge still lives in people’s heads.”
- O’Shea’s durability challenge — is this category temporary? — gets a conditional answer: demand is bounded by what humans can do that models cannot, digitally or physically. That might mean running a wet-lab experiment, alongside purely digital work. Foody likens self-evaluation to an English student told to grade every essay alone: measuring beyond one’s own capabilities is difficult. Math offers checkable answers; most domains need external feedback.
- Foody says investors have predicted superintelligence within three years for the past three years, including seed-stage objections that Mercor’s market would disappear. His counterclaim depends on that frontier remaining open; he says many top researchers at the labs expect human feedback to matter for decades. The skepticism has also reduced competition while Mercor retained access to capital, although he thinks the thesis is “on the verge of becoming consensus.”
4. Long-term operating discipline supports the decision to stay independent
- High standards, “time travelers” and customer obsession are Foody’s three operating values. The research frontier leads AI-native applications, which lead the broader economy; planning backward from that future matters more than copying competitors. He pushed back when someone on the executive team put up a weekly revenue graph, preferring customer happiness, expert happiness and expert utilization. “We have no interest in being second place in a market.”
- His account of setbacks includes the security incident, losing an early team member and roughly 18 months initially serving seed-stage customers; he says 80% of such companies go out of business within 12 months. “Chaos is the norm.” O’Shea raises Alfred Lin’s “shock absorbers” analogy; Foody agrees leaders should scrutinize problems during euphoria and become cheerleaders during rough patches.
- Foody says he and Adarsh have had chances to cash out and never work again, but would rather work “100 hours a week” on the problem. Discussing Scale’s arrangement with Meta, he describes the core team moving over and the remaining business focusing mostly on Meta, then says, “I don’t think we would ever do something like that.” M&A remains attractive but can magnify both mistakes and positive impact.
5. AI moats depend on brand, differentiated data and switching costs
- O’Shea suggests Foody said AI app-layer companies would die; he corrects her: “Not exactly.” Harvey’s moat combines brand with learning which customer problems matter, then creating benchmarks and training sets around them — “not training on customer data.” Survivors need frontier capabilities in their domain because “the model is the product.” Mercor’s public legal benchmarks, APEX and APEX Agents, helped interest Harvey in its work.
- Frontier labs benefit from economies of scale: distribution and revenue fund more compute and RL environments, improving subsequent models. Foody adds brand and recursive self-improvement, but says the latter “won’t be a vertical takeoff”; it will make progress slightly faster over time.
- For inference providers, “the switching costs of reselling GPUs are zero.” Foody argues that buying GPUs which then appreciate 50% can produce temporary margins, without establishing durable economics. He singles out Fireworks for executing on the more defensible approach: helping customers customize models that carry high switching costs and cannot simply be moved to competing infrastructure.
- API revenue quality is consequently easy to overrate, in his view. “We switch between APIs on a daily cadence,” using whichever model performs best on Mercor’s evals. He expects mature companies with substantial inference bills to behave similarly; large revenue alone does not resolve the switching-cost problem.
6. Enterprises become model builders as inference spending rises
- Foody predicts “every Fortune 500 is going to look like a Neo lab,” curating its own evaluation and training tasks. Anthropic and OpenAI “could very well be $10 trillion-plus companies,” but he does not want them to subsume the economy and become the only place to access frontier intelligence. Goldman Sachs illustrates why: its people, processes and documents embody competitive advantages; letting a frontier lab train on them would resemble handing them to a competitor.
- His adoption heuristic is inference spend. Below $10 million, companies generally do not care much yet about evals or optimization and are still figuring things out; at $10 million–$100 million, he suggests roughly 10% of the inference budget should fund eval datasets and routing decisions. O’Shea pushes back: “That is such a wide gap.” Foody concedes it is, then notes Mercor spends almost $1 million a day on inference: some businesses scale through that range quickly.
- Above $100 million in inference spend, companies such as Harvey and Cursor need to seriously consider post-training. Foody says companies can spend tens of millions of dollars a year — sometimes hundreds of millions or billions, as with Cursor — because they can amortize the fixed training cost across their inference. He expects enterprise inference bills to rise materially and adoption to follow that sequence. He is most excited about traditional Fortune 50 companies recognizing the opportunity and threat; he doesn’t expect Mercor to go public in the next 12 months.
- The practical instruction is to “start with the eval.” Foody says most frontier training compute now goes into RL, where models repeatedly attempt problems and receive rubric scores, rather than simply reading documents or learning input-output pairs through SFT. Subject-matter experts must create tasks and rubrics; establish trustworthy evaluation first, then expand the task distribution to improve against it. Dumping company documents into a model misses how he says learning works.