Mercor CEO & Co-Founder, Brendan Foody: How They Grew from $1M to $500M in 17 Months
Summary
- Mercor says it grew from $1 million to a $500 million revenue run rate in 17 months—one month faster than Cursor—with growth still accelerating at the endpoint. It had already reached nine figures before Scale AI was acquired and has quadrupled since; capacity is now the constraint because Mercor turns down projects daily and “could double overnight if we can meet capacity.”
- Foody argues the moat is identifying the 10–20% of experts who drive most model improvement, not merely supplying more labor. Mercor’s marketplace pays an average $95 an hour versus roughly $30 at Scale and Surge, and uses referral networks reaching Goldman, McKinsey, FAANG, medical and legal talent. Labs may initially spread work across vendors, but Foody says performance eventually forces concentration around partners finding those “10x contributors.”
- Synthetic data does not eliminate human-data demand while people can still perform tasks that models cannot. Foody expects humans to remain necessary in 10 years and calls superintelligence within three years “totally wrong”: models can win Olympiad gold medals yet still fail to draft his email, schedule a meeting or complete a multi-tool workflow.
- RL environments are a major opportunity because they convert real human workflows into learnable, verifiable tasks. Foody estimates Mercor has 50–60% of this emerging market and says lab executives believe it could “subsume the entire economy”—humans define how work should be done, then models learn to perform the repetitive execution.
- Foody considers academic benchmarks poor proxies for the capabilities enterprises actually buy. The fix is closing the “real-to-sim gap” with evals modeled on financial analysis, consulting research, software development and other real workflows: “If the model is the product, then the eval is the PRD.”
- For AI investors, retention and margins matter more than spectacular first-contract revenue, while switching costs determine whether subsidies create durable value. A company whose pilots fail 95% of the time is weak regardless of growth, while temporarily poor margins can work if distillation makes inference an order of magnitude more efficient within 12 months and sticky customers produce high LTV. Subsidizing low-switching-cost products is far more dangerous because users can leave as soon as the subsidy ends.
- Mercor’s capital strategy remains deliberately conservative despite its growth and likely near-term financing. Foody says the profitable company does not need cash and another few hundred million would not materially alter investment, but a low-dilution round could signal category leadership; he also sees the benefits of a “fortress balance sheet.” His unresolved question is whether capital efficiency is prudence—or whether Mercor should spend $100 million subsidizing supply and demand to press its advantage.
Deep dive
1. Mercor’s labor thesis began with unusually practical arbitrage
Foody’s earliest playbook was already marketplace-shaped: buy Safeway donuts for $5 a dozen, sell them at school for $2 each, pay his mother $20 for transportation, undercut a higher-quality competitor for two weeks, then move 20 feet off campus when the principal intervened.
In high school, he noticed sneaker resellers paying AWS bills despite qualifying for startup credits. He built their websites and helped them apply, earning “hundreds of thousands of dollars”—which made college look like a route toward a lower-paying FAANG or consulting job.
Foody still sees social value in college but little educational scarcity: he consumed Stanford GSB lectures online, and AI makes information easier to organize and learn. The irony is that Catholic school, chosen because his mother feared a progression “from donuts to drugs,” introduced him to his co-founders.
That background informs his rejection of the “body shop” description. Mercor’s role, he says, is mobilizing exceptional professionals who work directly with researchers—not hiding interchangeable workers behind a low-cost outsourcing layer.
2. Frontier data has shifted from crowdsourcing to scarce expertise
Early language models could use work from people writing “barely grammatically correct sentences.” Today’s problems require Goldman and McKinsey analysts, FAANG engineers, doctors and lawyers capable of producing—and helping researchers interpret—the highest-complexity data.
The causal step matters: researchers can independently diagnose an undergraduate math error, but may not understand a fifth-year Goldman associate’s work. Experts therefore provide both the training material and the judgment needed to interpret evals and hill-climb model performance.
Contribution quality is power-law distributed. On a 100-person project, Foody says the top 10–20% often produce most of the improvement; Mercor’s advantage is its referral network plus matching infrastructure that places those “10x contributors” on the work where they excel.
Foody disputed the claim that competitors lack quality algorithms: Mercor uses models to assess work, trains on supplied data to measure capability gains and operates as a research partner. His sharper distinction is cultural—Mercor pays an average $95 an hour, versus roughly $30 at Scale and Surge, because “phenomenal people that you treat incredibly well” generate quality and referrals.
3. Mercor’s $500 million run rate is now supply-constrained
Foody’s headline disclosure: Mercor moved from $1 million to a $500 million revenue run rate in 17 months, “the fastest revenue growth of all time,” one month faster than Cursor. It averaged 54% month-over-month growth for a period and is growing faster at $500 million than at any earlier point.
Scale AI’s acquisition was a real accelerant, but not the starting gun: Mercor was already at a nine-figure run rate and deeply partnered with frontier labs, then quadrupled since the acquisition. Foody’s diagnosis of Scale was nuanced—strong distribution and sales, but lost focus on product and on scaling quality. He separately emphasized that treating contributors well is central to quality.
Customer concentration resembles NVIDIA’s, although Foody would not disclose the exact largest-customer share. His defense: concentration is secondary to creating enormous value for the most important buyers, and NVIDIA demonstrates that serving a handful of exceptional customers can still support a multi-trillion-dollar business.
Labs have, in some cases, spread spend to prevent one supplier becoming dominant. Foody says this can reverse when diversification degrades data and model performance: fragmented markets consolidate because of the fixed investments in elite networks, quality systems and matching infrastructure.
4. Human demand expands whenever workflows become harder
Foody defines the addressable market as everything humans can do better than models. Synthetic reviews and augmentation may make expert interaction more efficient, but advancing the frontier still requires a “human reference point” against which an absent capability can be measured.
His best example began with 100 people stumping a model on a single-tool, multi-hour task. As the model improved, only 20 could still contribute; adding Drive, Calendar, Gmail and Slack, then extending trajectories toward 10- or 100-hour workflows, let the full group find failures again.
That is why Foody expects human trainers to remain necessary in 10 years. Models may hold Olympiad gold medals and exceed PhDs on reasoning while remaining unable to draft his email or schedule a meeting; he therefore calls superintelligence better than humans at everything within three years “totally wrong.”
Mercor does less traditional RLHF, but Foody estimates it holds 50–60% of RL environments. Executives and CEOs at leading labs believe these environments could “subsume the entire economy”: humans encode the framework for recurring research or operational work, then models learn to execute it.
5. Useful evals must resemble the work buyers actually need
Foody agreed that Humanity’s Last Exam, PhD reasoning and Olympiad math are poor measures of economic utility. Enterprises care whether a model can build a financial model like Goldman, create a consulting research deck or produce a web app like a software engineer.
The required shift is closing the “real-to-sim gap.” For Harry’s investment research, an eval might grade how a model searches online, cross-references PitchBook information, tests a product and applies tools—using a rubric much like a professor grading an essay.
Separately, Foody criticized companies for “vibe-spending on AI” without defining success. Evals establish the ground truth for each deployment: “If the model is the product, then the eval is the PRD.”
6. Retention, margins and switching costs separate durable AI revenue
Foody’s first test for application companies is retention, supported by customer conversations. If 95% of pilots fail, initial contract velocity means little; unparalleled retention and customers who genuinely love the product indicate real market fit despite today’s low-friction pilot budgets.
Margins remain fundamental, though context matters. Mercor has positive gross and net margins, but Foody accepts aggressive model-serving economics when distillation could make inference an order of magnitude more efficient within 12 months and current subsidies purchase sticky, high-LTV relationships.
His red flag is a competitive category with low switching costs: hundreds of millions—or billions—of subsidies create no durable value if customers migrate when discounts disappear. Mercor sees this tension in coding, where Cursor leads internal usage, Claude Code follows, and switching remains surprisingly easy despite emerging codebase-specific models and data flywheels.
Foody expects more engineers, not fewer, in five years. If AI makes engineers 10 times more efficient, he believes companies would build substantially more software and ship more features and iterations, making engineering an amplified and more valuable role.
Foody is less alarmed by broader AI capex on a 10-year horizon, while conceding pockets of exuberance. Code and foundation models attract substantial hype, but he also sees real value: Mercor’s engineers already receive “incredible” utility from Cursor, Claude Code and Cognition.
7. Valuation follows possibility, but Foody still chooses durability
Mercor was at $1.5 million in run-rate revenue when it met investor Victor; by the term sheet it had just passed $2 million against a $250 million valuation—over 100 times revenue. A later term sheet arrived around $20 million at another roughly 100-times multiple; Mercor is now 25 times larger than it was at that Series B.
Harry observed that $10 billion would equal only 20 times today’s $500 million run rate. Foody said financing is likely, mainly for low-dilution signaling, but profitability means the company neither needs capital nor would invest materially differently with several hundred million more. He separately cited the benefits of a “fortress balance sheet.”
He also leans toward remaining private. Jack Dorsey’s advice was to stay private as long as possible because quarterly reporting can erode long-term orientation; a three-year horizon may look frothy, while extraordinary businesses could look cheap over 10 years.
Foody’s live debate is whether to abandon capital efficiency and burn $100 million subsidizing marketplace supply or customer projects. Harry would do it only under credible competitive pressure and with the ability to undercut rivals; Foody’s default remains fundamentals, despite demand sufficient to “double overnight.”
His operating philosophy has evolved similarly: “996” described an intensely committed early team, not mandated hours. Mercor now optimizes more for output than face time, while using purpose and fast-appreciating equity to recruit “missionaries, not mercenaries” in a market where Zuck can offer $100 million in cash.
On models, Foody has moved from specialized-only toward “a lot of both” after o3’s generalization and GPT-5’s capabilities changed his view. He thinks today’s largest model builders probably already exist, though he is not certain and allows for startup breakthroughs. He also calls Gemini Flash’s smaller models extraordinary and underappreciated on evals.
Customization remains the opportunity because APIs have “low switching costs, not much pricing power” and, in his blunt formulation, “is not a good business.” He still expects foundational models to be huge businesses while enterprises increasingly customize them around their own tools, knowledge bases and processes.