Mercor CEO兼联合创始人 Brendan Foody：如何在17个月内从100万美元做到5亿美元收入年化规模

Brendan, dude, I’ve been so looking forward to this. I just had the best chat with Victor, who gave me the best intel, so you should be really quite nervous at this point. But thank you for joining me.

Thank you for having me on. I’m not sure what to expect with that, but I’m excited to jump in.

I think mothers are the most important things in the world, and Victor told me that I had to start with your ability to sell early and why your mother was nervous about it. Can we just start there?

Absolutely. I had a dozen different side hustles when I was growing up, selling things in one form or another. But one of my favorites was that, in 8th grade, I loved selling donuts. I saw that Safeway was selling donuts for $5 a dozen, so I would buy Safeway donuts, bike to my middle school, and sell them for $2 each.

I saw it was working, so I wanted to scale it up. I asked my mom to drive me to Safeway. She said that she didn’t want any giveaways, so she would charge me $20 to drive me in her minivan to Safeway, buy 10 dozen donuts, go to my middle school, and sell them for $2 each.

I had all sorts of things happen, including competition popping up selling Chuck’s Donuts, which, if people aren’t familiar, had about a $1 cost basis. They were higher-quality donuts, so I dropped my prices to $1 for 2 weeks to run them out of business because I knew that middle schoolers would care more about price as their comparative advantage.

My principal called me into their office to try to shut down my donut stand because I wasn’t allowed to sell food on school campus. I moved my donut stand 20 feet off campus so that they couldn’t police me, so to speak.

Tying back to your question, after my mom saw all of this when I was in 8th grade, she was very nervous that I would start selling drugs. It’s a small jump from donuts to drugs. She insisted that, while I’m not Catholic, I should go to Catholic high school to make sure that I stayed in touch with my values and met my co-founders there. I guess she was right all along.

Was that actually why she sent you to Catholic high school? Because she wanted you on the straight and narrow?

That was exactly why. My siblings had all gone to public school for 8th grade, and they had all gone to public school all the way through college. The primary motivation was that she didn’t want me to get into trouble.

As a principal, you’re the child who just pisses you off endlessly, aren’t you? You’re like, “The little troublemaker who moves it just outside the boundaries.”

Can I ask, Brendan, did you always know you’d be successful? What I mean by that is very specifically that, when I interview the best founders, they have a duality. They have this superiority complex: They think that they’re better than everyone. They don’t admit it because it sounds dickish, but they do. Then they have this inferiority complex where they’re not happy with their current state and want to do more and more and more. Do you have that?

I definitely had grand ambitions growing up for all the things that I wanted to do, but I don’t think it was nearly at the scale of what we’re doing today, nor did I expect how fast it would happen, because those 2 dimensions are nearly impossible to predict. I was definitely ambitious. I don’t think I had a perfect sense for what that would look like, though.

Victor told me about your not wanting to go to college. Before we dive into Mercor and the market itself, because there’s so much to unpack, I’d love to understand how you thought about college, why you didn’t want to go, and how that informs how you advise other young people on college.

I’ll start with the other story that I like to tell around my side hustle in high school, which tees up why I didn’t want to go to college.

I initially was reselling sneakers, as a lot of people my age would do in that generation. I realized that all of these sneaker resellers were eligible for AWS credits, but they weren’t claiming the AWS startup promotions. They were instead just paying big AWS bills.

I started a consulting agency where I would help the sneaker resellers create websites for their startups and help them apply to get credits. Some of those actually became venture-scale companies, and I made hundreds of thousands of dollars when I was in high school.

When I was starting to think about whether I wanted to go to college, I was thinking, “Why would I go to college to get some job at a FAANG company or in consulting, or whatever it is, where I’m making way less money? I would love to just go full-time on the things that I love doing.”

I had a big argument with my parents about whether or not I should go to college. Eventually, I appeased them and applied to colleges 10 days before the application was due.

How do you advise other young people today on the value of college, given what you’ve seen and experienced now?

I think so much of the reason that college is no longer valuable from an educational standpoint is that all of that information is available online. My parents’ preconceived notion was that they didn’t have YouTube, they didn’t have the internet, and they didn’t have all this access to information at their fingertips, so they needed to learn it from professors.

For me, I listened to almost every Stanford GSB lecture when I was in high school. I loved consuming information online and listening to all of your podcasts, Harry. I’ve been doing that since I was little. I think AI only exaggerates that by making it easier to organize, understand, and learn that information.

There’s still value to college from a social standpoint. I had a lot of fun, but I don’t think there’s too much value from an educational standpoint.

Listen, I totally agree. I went to university for about 4 weeks before I dropped out.

I didn’t know that.

Yeah, I went for 4 weeks, and then a sponsor offered me $100,000. I went to my law professor and said, “How much do you earn?” He said, “$82,000,” and I said, “Great, I’m out of here.” I hated law as well.

When I had Edwin on the show, he said that everyone in the space was essentially a body shop—direct quote. Is that a fair summation of the space, and how would you respond to that?

I don’t think it’s fair at all. We operate as close research partners to all of our customers, helping them mobilize some of the highest-caliber people in the world to push the frontier of model capabilities.

I think so much of our insight on the market comes from understanding how important high-caliber people are, rather than leaving them out of the narrative. I’ll give the backstory of how we really got involved in the market in the first place.

Scale AI came to us, and they used our platform to hire thousands of people. We realized that there was an enormous transition underway, moving away from the crowdsourcing paradigm that Scale and Surge pioneered: How do you get low- and medium-skilled people who write barely grammatically correct sentences for early LLMs?

The market is very quickly moving toward a sourcing and vetting paradigm: How do you find the Goldman and McKinsey analysts, the FAANG software engineers, and the top doctors and lawyers who can work directly with researchers to help them build the highest-complexity data on Earth and understand what that data is?

When we were dealing with undergraduate-level math problems, researchers could easily look at the math problem and understand why the model was making a mistake. But when we’re dealing with the kind of work that a Goldman associate would do in their 5th year, researchers can’t interpret the evaluations or all of the data they need to hill-climb and ultimately improve model capabilities.

That trend around a different engagement model and higher-caliber work caused us to take off and really catalyzed this meteoric growth.

If we extrapolate that out further and further with the advancement of models, your supply side becomes narrower and narrower. As models become smarter and smarter, the ability to do what you do requires smarter and smarter people, and there are, by nature, fewer and fewer of them.

How does that evolve to its ultimate destination, then, as we run out of really smart people?

Not exactly. The total addressable market is limited by the number of things that humans are better at than models. I’ll give an example that helps to contextualize this.

I remember when we started working on a high-complexity RL environment project, where the model would use one tool and interface with it in a task that would take a human a few hours to do. This became a famous product eventually, but we started out with 100 people, and it was easy to stump the model. It was easy to find mistakes that it was making, and over time, only 20 people could contribute to it. That is the exact dynamic that you're describing.

But then we started adding other degrees of complexity: How do we get the model to use other tools, like accessing your Google Drive, your calendar, your Gmail, your Slack, and all these different things? How do we get it to do the trajectories that a human might spend 10 hours or 100 hours on? All of a sudden, everyone else could contribute to the project again because they could stump the model.

What it goes to show is that, so long as there are things that humans are able to do that the model is not able to do, and we want those capabilities in the model—whether it's scheduling a meeting, writing emails for you, or whatever it is—we need humans to help create those verifiers and measure that frontier to ultimately improve model capabilities.

Dude, I had the founder of Cohere on the show the other day, and he said that we're absolutely seeing the reach of scaling laws being questioned, and that GPT-5 focusing on efficiency really is an embodiment of that. Do you agree that we're seeing the limits of scaling laws and entering a period of plateauing, so to speak, in terms of progression?

I don't think that models are plateauing. If we look at the last 12 months of progress in models, I've been blown away. But I do think that, to his point, we're definitely seeing a difference in the way that people improve model capabilities. It's no longer shoveling a lot of low-caliber, medium-skilled data into the model, right? It's much more these curated data sets with extremely high-caliber people that are built in a thoughtful way.

I think that transition toward RL environments and all this high-complexity data has been one of the most important things underpinning the trajectory of Mercor when we think about the supply side of that data.

When we think about the supply side of that data, you're obviously one of the providers, and a fantastic provider. There are many providers now, it would seem, including Turing, Handshake, and Surge. How do you differentiate on the supply side of data in this?

It's interesting because we saw the market shifting dramatically away from crowdsourcing toward sourcing and vetting. Once this happened, there were all these other labor marketplaces that caught on to that transition. They saw our growth and wanted to chase after it, saying the same things in podcasts and trying to position themselves in a similar way.

But I think one of the largest things we've realized is that the outcomes of data and the people who contribute to it are extremely power-law. Similar to a company, if you have 100 people on a project, oftentimes the majority of model improvement is coming from the top 10% to 20% of people, right? Just like the majority of the value in a company will often come from the top 10% to 20% of people.

What that means is that when we're able to build proprietary advantages in the way that we have—not only our supply base and the referral network to access them, but also the way that we match those experts with the opportunities where they're going to do phenomenal work—it creates so much value for customers that it's extremely difficult to compete against.

When we're able to find those people who are 10x contributors, it's very difficult to recreate.

I think a lot of people have cited a criticism of the space: They're very good at facilitation, but not great at measuring the efficiency of the data that's produced.

The challenge of being first on a show is that you say all the quotes, and then I can use them. Edwin said that none of the competitors have algorithms to measure the quality of the data that they're producing. Is that right?

That's not true at all. In fact, we use all sorts of models and algorithms to assess the quality. We train on data to see how it's improving model capabilities, and we do function as a deep-research partner to our customers.

I think the difference is that I think about our business as being at the intersection of labor marketplaces and AI research. How do we leverage our core competency in finding world-class people and pair that with the fact that we work with all of the top research labs at the frontier of model capabilities? We're not like the crowdsourcing companies, in that we try to hide all the people on the platform, pay them low rates, and so on.

One of my friends is on the board of one of your competitors, and they said that labs are incentivized to ensure that no one company dominates. They intentionally spread business around to ensure no one becomes too powerful. Is that true? Can you just help me understand that dynamic?

I think that has definitely happened in some cases, but ultimately, the thing that labs care about the most is how they improve model performance. How do they get those top 10% to 20% of people who are driving the vast majority of the model improvement?

That's how their spend allocation and investments ultimately get allocated: What are the vendors and strategic partners that are able to deliver those outcomes, and how do they work as deeply as possible with those partners?

We've definitely found that there are stories of customers who start out multi-vendoring, working with a bunch of different vendors, but ultimately get to the point where they realize that they're going to be making a trade-off in the performance of their model and the performance of the data sets if they're trying to diversify too much and lean very significantly into moving almost all of their work to us.

That's so interesting. So you expect a multi-vendor approach that then concentrates over time. Is that how you think about spend?

I do. If you look at a lot of the analogs in markets, they often start very fragmented, with many different players, but consolidate over time. So much of the reason for consolidation is that there are structural advantages in economies of scale to being the first player and having this fixed-cost investment in the best professionals in the world—the Goldman and McKinsey analysts, the networks associated with them—as well as all of the matching infrastructure for understanding exactly what tasks and jobs these people are going to do well at.

It doesn't make sense for so many different companies to be making those redundant investments. I think that the market being hot is what gives a lot of those companies more funding and more fuel, but consolidation generally happens as markets come back to earth a little bit and level up.

One thing that I worry about often is concentration of revenue. You saw it with NVIDIA, where I think it was 51% of revenue from 2 clients in one certain segment of their business. I think it was 36% in another segment of their business. What's your largest customer in terms of concentration of your revenue?

Our largest customer—I can't share the exact percentage—but the breakdown is relatively similar to NVIDIA. Part of the reason is that concentration is relevant, but the high-order bit is building a phenomenal business that's creating a lot of value for the most important customers.

Ultimately, NVIDIA is worth trillions of dollars, and some of the best empirical evidence that it's okay to have a business that leans into a handful of customers, especially when those customers are the best customers in the world—

You don't understand, Brendan. I'm the Brit who basically takes incredibly talented Americans with insanely great businesses and then critiques them.

It was when I said to Marc Benioff the other day, "Marc, single-digit growth, it's just not good enough." And Marc was like, "Dude, I have a $42 billion company. What do you have?" And I'm like, you know, that's a very fair response. I think you're right to respond with that. I absolutely love that.

Can I ask you, when Scale AI got bought, did your phone just go off the hook? Did demand just go through the roof?

It did. We were already at a 9-figure revenue run rate, and the company quadrupled since the Scale AI acquisition.

To put that in frame of reference, you were at 100, and then I saw—yes, you're at 450 now.

There are all sorts of news articles that have come out without complete information, but sorry, I didn't mean that as a spoiler.

No, but what we're sharing imminently is that we scaled the business from $1 million to $500 million in revenue run rate in the last 17 months, which is the fastest revenue growth of all time—1 month faster than Cursor's time from $1 million to $500 million. How much of that do you think was fueled by Scale AI being bought? Was that a real tipping point where you saw an acceleration?

It was definitely a tipping point where we saw meaningful acceleration. In fact, the company is growing faster now at $500 million than it's ever grown before, and the growth continues accelerating.

I do think that we had already been growing extremely quickly, and the fact that we were already such a deep partner to all of the frontier labs was one of the key things that positioned us so well when the Scale AI news happened, allowing us to expand those relationships and support customers.

When I speak to people in the space, they all say that they knew Scale was shit for a while. I'm British and very direct, which is quite anti-British, to be honest. They say that they all knew it for a while, and then it wasn't a surprise seeing that other people think they're shit too. Did everyone know that they weren't a great-quality provider?

I think people broadly knew. I think Alex was phenomenal at so many things, including distribution and sales. But in some ways, Scale lost focus on product and on scaling quality, and that was one of the largest challenges of the business.

Actually, if I had to choose the most important thing, it would be the internal link to quality: having phenomenal people that you treat incredibly well is the most important thing in this market. You need to get those people to refer all of their friends and actually help improve the frontier of models.

I think Mercor started out with this obsession with phenomenally talented people. Our average marketplace pay rate is $95 an hour, to put that in context, whereas Scale and Surge generally pay about $30 an hour. It's just a radically different approach to the way that we think about what kinds of capabilities we want models to achieve and how we want to treat the people who ultimately help to achieve those capabilities.

When we think about the hourly rate on the supply side from human-created data, one thing that challenges the model in my mind is synthetic data creation and how that supplants the need for human-created data. How do you think about a future where synthetic data creation removes the need for human data creation?

It ties to what I was saying earlier about how the total addressable market is bound by the amount of things that humans are better at than models. Of course, there's going to be synthetic reviews and synthetic augmentation to make it more efficient to engage with humans. But ultimately, if you want to push the frontier and get the model to do something that a human knows how to do but the model doesn't know how to do, then you need some human reference point to measure that, right?

Every time there have been questions about whether we're going to have superintelligence that's able to teach itself and do everything, that has turned out not to be true. We've continued scaling up the number of experts contributing to improving these models, especially in all of the professional domains that are most economically valuable.

In 10 years, do the models still need humans to help train them?

I very much believe so. The question comes down to when we'll have superintelligence. Once we have superintelligence and models are better than humans at everything, then of course humans won't be able to contribute to models or measure the frontier that models aren't able to do. But I still think it's a very long road.

These models have gold medals in Olympiad math and are better than the best PhD at reasoning, but they can't draft an email for me. They can't schedule a meeting. They can't do so many of the basic things involved in using a handful of tools to do a task that takes me a few hours. That entire road to automating the economy and building agents for everything is paved with humans creating eval workflows.

Do you think the current method of evals is bullshit?

How so?

We train or assess the effectiveness and efficiency of models based on humanity's last exam and all this other crap, which doesn't actually determine practical usage in society.

Absolutely. We're releasing a lot of announcements on this soon. One of the largest inefficiencies in all of AI research is that the evals people have been relying on—Humanity's Last Exam, PhD-level reasoning, or Olympiad math—are wholly disconnected from the outcomes that consumers and enterprises actually care about.

They want a model that is able to build a financial model like Goldman, build consulting research decks like a consultant would, or build a web app in the way that you'd expect a software engineer to be able to do. I think that transition is going to be very meaningful and one of the most exciting shifts in AI actually being useful in the economy.

Okay, I get you. So then, when we think about evaluations, what is the right way for them to be done? If I gave you a magic wand on evals, what would you change to make assessment more effective?

The number one thing is bridging the divide in the real-to-sim gap. How do we make sure that the tasks over which we're building evals and hill-climbing as closely as possible reflect the distribution of capabilities that people care about?

Think about the things that you do in your day-to-day job and how they could be evaluated for a model. Say you do research on investment opportunities that you're considering, where there's online research, cross-referencing their PitchBook data, using their product, and all these different things.

Imagine you could create a rubric that, similar to how a professor would grade an essay, grades how well the model is doing all of the online research and using the tools associated with doing that. I think that will be one of the most important trends as we move away from the era of academic evals toward measuring the real capabilities that users care about.

As you see the scaling of the company from where we were at the beginning to $500 million in revenue today, you need to change a lot as a leader. How have you seen your leadership change, and what have been the most difficult elements to grapple with?

I'm relatively young. How old are you?

I am 22. I turned 22 in April.

Wow. Okay. When you raised at $2 billion, people thought it was particularly crazy, if I'm being honest. In investors' eyes, that was a pretty punchy price. Now it looks ridiculously cheap. How did you think about valuation when raising?

Too many people think about valuation through the lens of market comps and revenue multiples, and not enough through the lens of what's possible with this company. What extraordinary thing can this company achieve, especially when you have such meteoric growth?

I'll give you a couple of fun revenue numbers at each of our valuations. When we met Victor, we were at $1.5 million in revenue run rate. He gave us a term sheet when we were at a little over $2 million in revenue run rate, so it was over a 100x multiple on revenue.

Because he paid $200 million.

He paid a $250 million valuation at our Series B.

I'm sure the Benchmark partnership thought that was insane at the time. At the Series B, when Felicis gave us the term sheet, we were at $20 million in revenue run rate, so it was a 100x multiple on revenue. But what they saw in talking to customers was the phenomenal experiences that we were creating and that our growth was going to continue.

Now we're 25 times larger in revenue scale than we were at the Series B, but we're in a spot where the business is so profitable that we don't really need to go out for financing or spend too much time thinking about financing, even though we often get a lot of offers and interest.

Dude, that is hilarious because I obviously met Adarsh, and he very kindly let me put in a small check. I had no idea you were at $20 million. I thought you were way bigger.

I think we let you put in a small check a little bit later, because the company—keep in mind—was growing over 50% month over month. We averaged 54% month-over-month growth for a while during that time period, so it wouldn't shock me if it was a couple of months after the round and we were at a meaningfully higher revenue scale.

Dude, I'm thrilled. Otherwise, I was massively off to my partnership, and I was like, "Yeah, yeah, yeah, yeah. They're way past where they'll be," with them looking at us and going, "What?"

I love that. Do you need to raise more money? Again, I am direct to a fault. The rumors of a $10 billion valuation—if you're at $500 million, that's only 20x, and given your growth rate, that would be cheap.

Definitely. That's what I've been thinking about, too. We honestly haven't given it much thought. We've gotten a bunch of offers from existing investors. We haven't really shared any materials on the business. There's just been outside diligence and offers based on that.

Do you like that? Don't laugh. Is it a nice feeling, or is it, "Hey, let me just focus and do my work"?

I think there's a bit of both. Parts of it feel validating, but parts of it feel distracting. We just want to focus on creating phenomenal experiences for our customers and for the experts in our marketplace.

I think it's likely we'll do a financing soon, with low dilution, largely because there are a lot of benefits to signaling ourselves as the market leader in RL environments and in all of the high-complexity data that we produce. We'll keep you updated, Harry, and stay tuned.

Do you think a big financing will do it? If you think about it, I'm just intrigued. As you said, Surge AI has big revenue numbers. They're at over $1 billion in revenue now.

Like, is it the financing that'll do it?

Well, obviously, the financing won't so-called do it, but I think it can definitely play a part from a signaling standpoint. Making a little bit more noise about that, what we do, and how we see the market developing over time could be interesting.

If you had truly unlimited resources, what would you do differently?

This is a tricky question because I feel like we're at a point where we're trying to invest as aggressively as possible, but the business is still profitable and we're not trying to be profitable. And so I don't think that having another few hundred million in cash would meaningfully change the way that we're investing, but I do think that having a fortress balance sheet has its benefits—having sort of the new mark of the company, et cetera.

And so I don't think it would change the way we're investing too dramatically.

You're going to go, “Harry, what are you talking about?” But at $500 million and growing at the rate you are, you'll soon be at a scale where an IPO is very possible. Given public pricing now being better than private pricing in a lot of markets, do you want to go public sooner rather than later?

It's not something I've given too much thought to because it's sort of surreal, considering we started the company in January 2023. And all my college classmates just graduated in May.

But I think there's a lot of benefits to staying private. It's funny, I remember when I was talking with Jack Dorsey before he invested, one piece of advice he gave me was that we should stay private as long as possible.

What was his reasoning for that? Super interesting.

Well, I think it's just that it allows you to stay very long-term oriented. Public companies get so caught up—even though founder-led companies tend to be more resistant to it—I think public companies still get more caught up in the quarterly numbers and aren't as focused as they maybe should be on all of the long-term drivers of value and moats.

And so I think that that is one of the core reasons: allowing us to stay very long-term oriented, especially when there's also so much access to capital in the private markets.

Do you think there's too much cash in the private markets today?

I mean, I don't know, because it's sort of like a supply-and-demand question. If I were an investor, I would definitely think that there's too much cash in the markets, right? It's sort of spurring higher competition. You could say there's a load of shit competitors who are getting funded to the tune of hundreds of millions that shouldn't be getting funded.

I think that's definitely the case. My heuristic for this is the age-old saying—or the idea, at least—that it's probably overestimating in the short term and underestimating the long term. If we're evaluating things on a 3-year time horizon, it wouldn't shock me if we feel like things are frothy and it's a crazy time. But if we're evaluating things on a 10-year time horizon, all of these extraordinary businesses that are being built will look like a discount. And the challenge right now is just saying, “Are we in 1996 or 1997, or some other time?”

Dude, you weren't even born then, so you can't talk about that. That's when I was born. I have a FIFA game from when you were born, and that really makes me feel old.

Did you see the MIT study or release?

Yeah.

What did you make of that?

I think it ties to the exact point you were making earlier about how evals are bullshit, right? When we start showing that we have Olympiad gold medals or PhD-level reasoning, that doesn't mean that it's going to be useful to enterprises. In fact, in 95% of cases, we're seeing these failure cases.

And the answer is that we'll need evals for every one of those implementations and examples, because evals are the way that we measure the truth and have a static point of understanding what the models are capable of. If we think about the model as the product, then the eval is the PRD. So many people have been vibe-spending on AI without actually writing the PRD of what they want to implement and how they measure that it's going to be successful.

Dude, I need your help. Okay, you said “vibe-spending on AI.” The revenue numbers that we see from some players in the application layer are just awe-inspiring, scaling in a way that we've never seen before in my history, anyway. How do you think about the sustainability of revenue for the majority of AI companies, and how would you advise me, a friend, and investor?

I think the most important thing is looking at the numbers and anecdotes around retention to see the revenue health and whether there's real value. If you meet an application-layer company where 95% of their pilots are failing, it's probably not going to be a good investment. But if you meet a business that has extraordinary, unparalleled retention numbers and you talk to those customers and hear about how much they love the product, then of course it's a really exciting opportunity.

And so I think that those signs of true market fit are the most important when there's sort of a lower friction to accessing initial pilots or contracts.

Totally get you there. The other element is margin. And the margins are pretty terrible in a lot of cases, especially when you take into account free-user giveaways, which there's a lot of.

Should we give a shit about margin structures given how early we are in the cycle, or, yes, we should? It's always fundamental.

I think that the answer is yes. Both of those matter, and it's very contextual. On one hand, I am a huge believer in capital efficiency. We have very positive gross and net margins, unlike most AI companies. But on the other hand, I also see the case that if you're able to distill models and make them an order of magnitude more efficient in 12 months, then it could make sense to run really aggressive margins on serving models.

It really comes down to the stickiness and whether those subsidies today are driving large LTVs that make sense long term. But I think the case where I would be hesitant is when there are very competitive markets with low switching costs, so that people are pumping hundreds of millions in subsidies, maybe billions in subsidies, and then all of a sudden the customers are switching over to a competitor if those subsidies dry up.

How do you feel about the often-banker concern that the level of capex is concerning because of the revenue generation required to make up that capex? Do you share that concern, or do you think this is a supercycle? Of course, the investment is required and the revenue will show itself, like Masa believes it will.

I'm less concerned about the broader capex because I think that if you have a 10-year investment horizon, all of these things—or the market generally—will look like it's at a discount. But I think that there are definitely cases of exuberance, right? And people need to just be thoughtful about which investments are going to have those positive 10-year-horizon ROIs and which don't make as much sense.

What segment do you think is most overhyped, overexuberant? Not company, just segment.

Nothing jumps out to me on that because obviously I think the things with the most hype are code and foundation models and maybe starting use cases in finance. And I feel like the value being created is also very real. The amount of utility that our engineers get from Cursor, Claude Code, and Cognition is incredible.

And same thing: do you use all 3 internally?

Yeah, we let people choose, and so various people use different products.

What is the distribution?

I think it's a lot of Cursor usage, closely followed by Claude Code. But it's hard because it's very dynamic. The market is changing so fast and the products are improving so quickly that I think some of that distribution will change over time.

Do you think there's switching costs between those?

There are surprisingly low switching costs. And so it makes me think that definitely some of these products are moving in the direction of adding more switching costs, right, with an understanding of how you interact with the platform and having data flywheels around that, or custom models for your codebase.

But I think a lot of those sources of defensibility are taking more time to develop, and right now the market is very competitive, which has driven a lot of the negative gross margins that we've seen companies have in the coding space.

In 5 years' time, will you have more or fewer engineers?

I think more, and the reason is that engineering is such an elastic role, right? If we could build 100 times more software, or say we make engineers 10 times more efficient, we would probably build 100 times more software, right? Insofar as maybe not unique platforms, but the amount of features those people would ship and the iterations on every ranking algorithm, et cetera.

And so I'm a huge believer in the fact that AI, especially in domains like software engineering, will be an amplifier in making people more productive and making people more valuable rather than diminishing their value.

You mentioned code there being one, and you mentioned models being another. Do you think the biggest model providers have been created already, or do you think some of the biggest in the future are yet to be created?

I think the largest model creators already exist, but I'm not 100% sure about that.

I definitely have some caveats about it. My expectation for why the largest model builders exist is just the extraordinary capex in terms of both data and compute investments that go into that, as well as building out all the teams of researchers, which has quickly become phenomenally expensive. At the same time, I think there may be other breakthroughs that help enable more model progress, and those could play a role coming from startups.

I love the kind of dual-sided mindset there. You mentioned the expense of talent. Is the expense and the economics around talent today in AI, in SF, just nuts?

It definitely is. I mean, certainly also beyond my wildest imagination a couple of years ago. But I think what it's really amplifying is the importance of having a really strong purpose, more so than just paying people well, because lots of companies can pay people well.

I really like you, Brendan. You're awesome, but come on, dude. When Zuck puts $100 million down, you're like, “Okay, yeah, I'm out of here.”

Look, I agree. You still need to obviously reach parity with respect to the economics of things and, of course, give people a lot of upside in the business. But part of purpose isn't only the mission of the company, but also the economic upside associated with that mission. I'm not sure startups can pay someone $100 million in liquid cash, but we can give people equity grants that are appreciating extraordinarily quickly as part of the vision of the company to help people capture upside in this purpose.

I do think that is increasingly important: having an employee base of missionaries, not mercenaries, and people who are in it for the long haul.

Will Zuck's spend work, do you think? He's got all the mercenaries together, who are very talented, brilliant people, but does that work?

I think so. I think there's an extraordinary team there, and so it'll be fun to see what they build. But these things are always hard to say.

Which team do you think is underappreciated and doesn't get the love that it deserves? It's interesting because I feel like OpenAI gets a lot of the love, with ChatGPT being the brand that everyone talks about. I feel like Anthropic gets a lot of the love around code and Claude Code. xAI definitely gets much more on the consumer side as well.

I definitely feel like a lot of the Gemini Flash models are also extraordinary and underappreciated on evals, especially their small models. I'm always amazed by them. So if I had to choose, maybe not a company but especially a set of models, I think the DeepMind team did a phenomenal job on a lot of those smaller models.

Do you think we live in a world of many unbundled, specialized models or fewer monolithic, generalized models, like the providers you mentioned?

I used to be very much in the camp of a lot of specialized models. Now I think it'll be a lot of both. I'm much more split in that.

What changed to cause that change of mindset?

The amount of generalization that we're seeing, especially with o3, blew my mind. It was just a phenomenal model and generalized so well. I think GPT-5 as well is a phenomenal model.

When there's still so much headroom in these foundational capabilities, it feels structurally more efficient to have those as individual investments to improve model capabilities. But I still think we're just in the first inning of model customization, with every enterprise wanting models to know how to use its own set of tools, all of its own knowledge bases, and the processes that it has codified. That'll be another huge area of investment over the coming decade.

Do you buy sovereignty as a reason why a model provider wins? We've got Mistral in Europe, and you have Cohere in Canada. Is sovereignty a reason why a model provider wins, maybe in a scoped part of the market?

I could see why, for example, there would be a lot of benefits to having Mistral be an expert in European law, which might have nuances from other kinds of law, and they've just invested far more in having the best model there, where it doesn't make sense to use other models.

But I don't think the largest companies, per se, are going to be those that invest in a specific geography. I think it's going to be a broader set of capabilities and the general-purpose models that people use every day to code, build products, or do their day-to-day work.

I don't know if you know this, but I'm particularly disliked in Europe because of my affiliation with, or affection for, the 996 work culture. Truly, my DMs are basically a war zone nowadays. 996 is a model that you very much espouse, too. Can you talk to me about why you're 996-bullish first?

Well, not exactly. I need to offer a key clarification, which is that we've actually never mandated hours. When we were talking about 996, it was more so a description of how the early team worked. In fact, the reason we talked about 996 was because people were working so much more than that, and we wanted people to go home a little bit early so that they could be well rested.

I think that intensity is, of course, extremely important in building a generational business. At the same time, I think we've become less focused on the in-person elements of that intensity and recognized that it can be expressed through outputs. When the market for talent is so competitive, it especially makes sense to optimize for working with the best people, less so than optimizing for face time.

Fascinating. So now you're at the stage where you need to bring in execs and make the language with which you speak more conservative.

Well, I don't know exactly.

I love it. I work with so many companies where they're 996, 996, 996, and then suddenly it's, “We need to bring in that CPO,” and he's never going to be 996 because he's a stellar CPO from a big company. You're like, “What? No, no, it's all about impact. It's about impact.” The language changes to be a lot more neutral. I think my lesson is that you need to do that.

Look, I think the thing is, when there were about 20 of us in a room, working with our India team as well, everyone just loved what they did. If people left to go home for dinner or had something else going on, we wouldn't bat an eye.

You just fire them, give them their box, and say, “Come in tomorrow.”

I think the truth is that all along, it's been much more about hiring people who give a damn, love what they do, and are obsessed with it in the way that we are, rather than specific hours. Early on, those were highly correlated, but I think that as the company expands, they're not always as perfectly correlated, and there are definitely exceptions.

Totally get that. I just want to ask one final one before we do a quick fire. I got asked this brilliant question the other day that's really stuck in my head: What would you do if you weren't scared?

An example for me, just so you have a framing, would be that I'd move to Silicon Valley. I'd compete in the coliseum of technology rather than sitting in London, being happy being a big fish in a small pond. What would you do if you weren't scared?

It's an interesting question because I feel like I live in a very risk-on way, always trying to make big bets. Maybe one ties to capital efficiency, and that part of it—maybe it's for better or for worse, right? Part of the reason that we've run the business in a very capital-efficient way is that I've always been very thoughtful about how markets will develop over time and how we ensure that we're building a super-durable, sustainable business that will be around in 10 years.

But I often wonder if maybe we should just start burning hundreds of millions of dollars a year. So that's one thing in my head.

Could you? It's not—

Possible. I think we could find a way.

How do you spend it on talent? Like—

I think on subsidizing either the supply or demand side of the marketplace. How do we get great people on the supply side, or how do we subsidize customer projects?

The business certainly doesn't need to do these things. We have the demand to double overnight if we can meet capacity, and we have a supply base that loves us and is growing phenomenally quickly. But at the same time, I do think that if I were trying to burn $100 million, I could figure out a way to do that.

We can go away for a weekend. I'll show you how to burn $100 million.

Harry, what do you think as an investor? How would you handle that? Would you be scared and capital-efficient, or would you be maximally aggressive about burning money?

I don't live the competitive landscape that you do. If I'm feeling continuous pressure from competitors that I feel are good, and I have an ability to undercut them in a way that they don't undercut me, I would absolutely leverage cash reserves to subsidize it. I'd be a loss leader until I could bluntly strangle them out of the market.

Interesting. Yeah, but it depends. If you don't feel that competitive pressure—which is clearly not showing in your numbers—cash can actually be a bit of a problem at certain stages.

You know, when you look at large companies, you need to make your cash work for you, and you really have to buy the growth in a lot of cases. You just don't want to get to that stage.

I totally agree, and I think that's why we've always erred on the side of capital efficiency and fundamentals.

So, what does Peter say?

I think Peter is more on the side of capital efficiency. You know, he's seen how these things play out and the ups and downs of markets, and he's been more in that camp.

And don't get me wrong: I'm still incredibly bullish on the market and AI. I think we're much more at, like, 96 or 97. But I think that focusing on fundamentals at least does buy you a lot of durability and long-term value, just having the right values and culture that can be easy to lose sight of in this one-way door of not being efficient.

Final one, I promised. Brendan, you said there was double the demand than the supply. Is it that much of a supply constraint that, if you had the resources or reserves on the supply side of data, you could double the business?

Definitely. We turn down projects every day. The reason is we're very focused and disciplined about working with the best customers in the world and doing phenomenal work for them. So the capacity question is: how do we scale up our ability to do that? That's my biggest focus right now.

Brendan, what does your mom say?

It's evolved over time. When I dropped out, she was very upset. Now I think she's come around.

Yeah. Have you done secondaries?

A very small amount.

Do you advise founders to take them or not take them?

I think the most important thing is making sure it's not distracting, right? Because ultimately, the vision that we're selling the company—selling everyone on—is that we are fully committed, and I want to demonstrate that in every aspect of the word: this is our life's work and the thing that we plan to spend the next decades on. Showing that on every dimension is important.

I want to do a quick-fire round. I'm going to say a short statement, and you're going to give me your immediate thoughts. Does that sound okay?

Sounds great.

What's one widely held belief about AI that you're like, "God, that's so wrong. Just please stop"?

That we'll have superintelligence in 3 years that's better than humans at everything. I think it's totally wrong.

Wow. Okay, good. I agree with you totally. You can be the CEO of OpenAI for a day. What would you do that they're not doing?

I think model customization is a really exciting opportunity because APIs will have low switching costs and not much pricing power. It's not a good business. Focusing more on model customization is a really exciting opportunity.

Do you think OpenAI will win the consumer, with ChatGPT as their Trojan horse, and Anthropic will win business and enterprise, with Claude Code winning that segment? It certainly seems like that. What question should every AI company be asking themselves that they aren't?

I really like the thing Sam Altman says: will models being dramatically better in 1 to 2 years improve your business or worsen it? I think that, in so many ways, is the most important question to see if you're building a business that's durable and well positioned for the future.

He said it first on our show, and that was the show—yeah. And that was the show where he took a 20VC jumper and put it on. Before the show, I was like, "Brendan, this is unbelievable branding."

Well, Sam wore one of our Bour jackets the other week, which I was over the moon about.

Yeah, I was too. Okay. And then he gets on camera, Brendan, and do you know what he says?

What did he say?

"Startups, we're going to steamroll you." And I am a startup investor. My job is to inspire entrepreneurs. I'm like, "Oh no. Oh no." But yes, dude. What have you changed your mind on in the last 12 months?

You know how I talked about how I thought there would be a lot of model customization? This is a little contradictory. I think I've indexed more on a lot of generalization and just that foundation models will be huge, huge businesses, while I still think they should invest more in customization as well.

What investor do you not have that you would most like to have? It doesn't need to be a fund. It could be a person. It could be anyone.

I think Jeff Bezos. I've admired Amazon so much, and just the early clarity of thought in the business and long-term focus. I think there's a lot of analogies, so I would love to learn from him.

Why do you not have him? With the cap table you have, getting him would not be impossible at all.

I haven't met him. I'll have to—I haven't put too much time into it. I've been meaning to.

You can give yourself one piece of advice going back to January 2023, starting Mercor. What do you know now that you wish you had told yourself back then?

I would say, focus on foundation model labs. I didn't understand the scale of the opportunity with foundation model labs in January 2023. I think being the first company to realize that, especially in how our marketplace fit into it, was one of the most impactful things. If I'd realized that 9 or 12 months sooner, that would have been even more exciting.

How penetrated into their spend are we? You know, when you look at them, they are absolutely destroying a lot of their economics to win this race. They can only do that for so long. How penetrated are we into their spend?

There are different buckets within their human data spend. There's the RLHF buckets, which we don't do as much of, but then there's the new data types that everyone's moving towards, called RL environments, where our rough estimate is that we're at 50% to 60% of the market. So we're doing quite well on that and expanding market share quickly.

Do you think market share is still continuously expanding? How much room does the market itself have left to expand?

I've talked to multiple executives and CEOs at leading labs who believe that RL environments will subsume the entire economy, because it doesn't make sense that humans would be doing monotonous, redundant work of redundantly researching different companies each week or guests for your podcast. It makes way more sense for humans to build the framework of how to do that, so that models can then learn how to do it and do it for us. I think that is going to be a ridiculously exciting transition.

Dude, I've so enjoyed having you on the show. This is why I don't send questions. We have all this ahead of time, and none of it has been covered because this was way more interesting. Thank you so much for being so flexible with my questions. You've been fantastic, dude.

No, I love it. Thanks for having me on, Harry.