Adarsh Hiremath @ Mercor: The Fastest Growing Startup in Silicon Valley | E1261

Summary

• Mercor raised $100M at a $2B valuation led by Felicis, with Benchmark and General Catalyst participating — roughly 5% dilution; Harry cited $50M ARR in November while saying he may have quoted it wrongly, and reported hearing growth of “50% month on month continuously for quite a while.” The founders weren’t focused on fundraising: “we didn’t intend on doing the fundraising… it sort of just came to us,” and the money is a balance-sheet play for a long-term labor-aggregation goal, not a spending plan.
• The core thesis: human data and talent assessment have become the same thing. Data labeling has shifted from crowdsourcing (drawing boxes around stop signs for Waymo) to finding the expert who can make a model better in a specific domain — “figuring out who that expert should be is 100% a talent assessment problem.” AI labs hiring post-training experts is a forcing function on Mercor’s endgame of a unified global labor market.
• Against the synthetic-data bulls (Harry cites Jonathan at Groq), Hiremath holds that data is the bottleneck more than compute or algorithms, and it isn’t zero-sum: “evals definitionally have to be outside of model capability,” so humans must build them, and SFT/RLHF/RL environments all need expert humans — “for a very very long time.” Low-quality human data pushes nothing; high-quality human data is, again, a talent-assessment problem.
• When software costs approach zero, the businesses that succeed will be built on network effects: “the businesses that succeed… will be built on network effects. The companies that don’t could even give away their entire code base and still be alive.” SaaS changes shape — the next era replaces entire services end to end, which is what Mercor is to a recruiting agency.
• The business runs with no sales team and takes that can exceed 30% on a case-by-case basis, fully automated from resume pull to AI interview (spins up in under 10 seconds for any role) to payment. The pricing logic: unlike Uber’s 4.8-vs-4.9-star driver, “there’s a huge difference between the top 0.1% and the 80th-percentile person” — so “it’s not a question of price, it’s a question of quality” and the take is “often a second thought.”
• The infamous 996 (9am–9pm, 6 days) is framed as a side effect, not a policy: “the only reason we actually just floated those numbers out is because we didn’t want our team working on Sundays.” Hiring indexes on the one thing you can’t teach — caring — and the hardest scaling lesson was that “scaling culture is harder than scaling software.”

Deep dive

1. Debate partners, a dev shop, and “dude, how hard could this be”

• Adarsh met co-founder Surya at age 10 — “the only elementary schoolers who wanted to compete in high school debate” — and frames the debate partnership as his first startup: “we had 50/50 equity in each other’s success,” constant win/loss feedback, and the same lesson that picking the right partner is the most important decision you make.
• Mercor started with no business ambition: a dev shop recruiting “really really exceptional folks from India,” until the founders realized the people mattered more than the software — so they automated the candidate side, then the company side, and the marketplace was born.
• The dropout decision was made with no seed round, no likely Thiel Fellowship, and “a little bit of revenue” — an emotional decision. In a three-desk Palo Alto office, Surya said: “dude, how hard could this be.” The surreal moment wasn’t the $3M+ General Catalyst seed wiring in — it was changing their Gusto salaries to $500 a month: “I felt like we made it.”

2. The round: $100M at $2B, raised by accident

• The Benchmark round happened via helicopter: Victor asked Brendan if he’d ever been on one; “before you knew it Brendan was on a helicopter with Peter Fenton,” and the firm was chosen “very very quickly.” About six months later — at eight figures of revenue, heads down, not fundraising — came the next round because they wanted to be in business with a likely Sundeep Peechu and Felicis.
• The new raise: $100M at $2B led by Felicis, with GC, Benchmark and others participating. Harry’s framing — “you dilute 5%, you get $100M on the balance sheet, phenomenal round” — met Adarsh’s discipline on deployment: “it gets really dangerous when people raise the money and then think that they just have to spend it immediately.” Building a unified labor market “is going to take a long time”; the balance sheet should be commensurate.
• Governance is lean: the board is the three founders plus Benchmark, that’s it — and no, he doesn’t enjoy fundraising. “The thing that we really really enjoy is moving the business forward.”

3. The thesis: human data and talent assessment are now the same problem

• Five years ago human data meant crowdsourcing — “likely Waymo wants a bunch of their images labeled, you get a bunch of people across the world to draw boxes around stop signs.” Today it’s “GPT-4o or whatever model is not good in a particular domain, so we actually need an expert to make the model better — and figuring out who that expert should be is 100% a talent assessment problem.”
• Lab work isn’t a niche detour but a forcing function: the unified labor market needs “tons of smart people on the platform and the ability to predict job performance” — “which happens to be the exact set of problems that a lot of the AI labs are having.” Labs hire experts through the platform “to essentially help with post-training models.”
• The one metric he watches: customers keep expanding — net retention is over 100% “by a large margin.”

4. Data is the bottleneck — and the synthetic-data pushback

• Asked whether data constrains model improvement more than compute or algorithms, the answer is categorical: “data is the bottleneck — that would be an accurate statement.” Harry pushes back with a view he attributes to Jonathan at likely Groq that synthetic data is higher quality, without “the dregs of the internet like Reddit.”
• The rebuttal: it’s not zero-sum — synthetic data will matter — but “evals definitionally have to be outside of model capability”: to judge a model you need a human-created eval set better than the model at that task. Same for SFT, RLHF, and RL environments. And the phrase “low-quality human data” concedes the point: “high-quality human data will [push models] — and again, that’s a talent assessment problem.”
• How long do expert humans matter? “For a very very long time.” His model of the human-AI relationship isn’t unidirectional handoff: AI might get you 60–80% of the way, a human takes you the rest — “and finding that human will become harder and more valuable.” Harry’s counter — don’t you need fewer humans as you approach perfection? — gets the labor-market answer: work moves “towards specialty and sophistication.”

5. The machine: no sales team, 30%+ takes, quality economics

• Mercor has not a single salesperson outside the founders — growth is inbound from customers who hired through the platform, and the constraint “is more of a bandwidth thing than any tactical or coordinated sales motion.” The wow moment isn’t the demo or the price: “it’s usually when the first couple candidates start working with them.”
• The entire candidate journey is automated — resume and salary-expectation ingestion, a personalized AI interview spun up “in under 10 seconds” for any role from engineers to lawyers to doctors, through getting paid. The take is case-by-case, “for some customers over 30%,” and tolerable because of his Uber contrast: a 4.8 vs 4.9-star driver barely differs, but “there’s a huge difference between the top 0.1% and the 80th-percentile person… it’s not a question of price, it’s a question of quality.”
• The India origin is personal — his and Surya’s parents immigrated from India, so recruiting started at their schools — and the founding anecdote is a hiring inefficiency in miniature: the best engineer he’s worked with came via a Facebook ad, failed the manual interview, then sent “a really really long message about what exactly he got wrong” and was hired anyway. Today the number-one source of workers on the platform is the United States, with clients mostly US too.

6. 996 as side effect: culture is the hard part

• On 996 culture: “the only reason we actually just floated those numbers out is because we didn’t want our team working on Sundays.” He calls 996 “a side effect rather than an objective” of selecting for mission-focused people who “don’t want to wait until Monday to move the company forward” — and he doesn’t think it can be done effectively remotely: “that motivation, that intensity, you feel in the same room.”
• Hiring index: “you can teach people a lot of things… the one thing that you can’t quite teach people is to care.” The unspoken scaling lesson: “scaling culture is harder than scaling software” — the culture of the first 20 people “is in some ways the strongest the culture is ever going to be,” and keeping it is “the most important part of building a legendary company.”
• Harry said he had heard 50% MoM growth; Adarsh called that level a “perpetual stress test on the business” — things constantly break, and “everyone in the company needs to keep outgrowing themselves.”

7. Zero-cost software, programming in English, and 100 billion jobs

• Cursor changed how he builds — “you can basically snap your fingers and it’ll get done” — and the implication is that software gets commoditized very quickly. “The businesses that succeed in a world where software costs approach zero will be built on network effects. The companies that don’t could even give away their entire code base and still be alive.” Mercor’s two: marketplace liquidity, plus a job-performance data flywheel that surfaces the best person for a role “even if they themselves don’t know it.”
• On the don’t-study-CS advice: programming becomes more important at a new abstraction — “the leap from assembly to Python was maybe even a bigger leap than the leap from Python to natural language.” The future coder may be someone with average skills by today’s standards… orchestrating thousands of superhuman coding agents. His change of mind in 12 months: the next era of SaaS “will replace entire services end to end.” His worst product call: betting everything on chat UI — “we may have mistimed a little bit.”
• On models: the market is shifting to reinforcement learning (“you’re already seeing this with o1, o3, the DeepSeek models”), yielding many specialized models at the application layer but only a couple of foundation-model companies — the cloud analogy “roughly holds.” Mercor uses a variety of models and has been particularly thrilled with OpenAI, and “the whole product gets better as the models get better.”
• The 2035 math, worked backwards: a couple billion job seekers, a couple dozen jobs each (“factoring out all the jobs Mercor creates for AI agents”) → 100 billion jobs created and a unified labor marketplace solving matching “across every role, across every company.” His contrarian close: “being a recruiter is the highest prestige position in any company… you can gather all you need to know about a company from seeing the talent inflows and outflows.”