The world going forward, there is going to be nothing that no one can build. Everyone is trying to commoditize the other. Value accrual is a time-dependent phenomenon. The age of the polymath is back. We will see the best companies treat teams more and more like Seal Team 6 or like professional athletes.
>> Harry: Matan, it is so good to have you in the studio. You've just insulted my continent with the suggestion that we've only come up with bottle caps while you came up with Transformers. Not wildly untrue, but this is going to be a fun show.
>> Matan: Thank you for having me, Harry. I was just doing a show yesterday with Rory and Jason, and Rory was basically saying, the fundamental question is, will we see an increase in GDP coming from AI and the coding developments that we're seeing, and will it lead to GDP increasing above the 2% average for the last 200 years?
>> Harry: Do you think we will see meaningful productivity gains from the AI tooling that we're seeing, or is Uber's concerns validated?
>> Matan: Yes, absolutely we will see tremendous growth from these tools. I think it takes time to permeate through because on an individual basis, almost on a problem-by-problem basis, we can solve problems faster. Companies generally organize around solving problems. If you're organized around solving problems and you have some set of personnel, everyone is now going to be able to solve more problems with the same number of people, or solving the same number of problems with fewer people. But it takes time for resource allocation to adjust. A lot of businesses will be grappling with: do we want to solve more problems now because of the increased leverage we get, or do we want to solve the same problem but more efficiently?
>> Harry: Do you think we will have fundamentally smaller teams, or same size teams going after a more expansive area?
>> Matan: It's really not obvious. I was watching Andrej Karpathy and he was talking about how the 10x engineer is wildly misunderstood β you won't see the 10x engineer, you'll see a smaller number of 100x engineers, and this bifurcation of engineering talent.
What is a 10x or 100x engineer? I don't agree with the language. The way I think about it is load-bearing individuals in an org. If you remove this person, things fall. In some orgs there might be people where if you remove them, nothing happens. These high-leverage people are now being handed a tool that gives them even more leverage. Those who know how to use leverage will have even more impact, and those who don't will be that much less valuable.
>> Harry: If I am a leader today, what is your biggest advice on how to think about resource allocation for tokens internally?
>> Matan: This resource allocation problem β it's not just tokens, it's dollars, tokens, and people. This is the thing every C-suite is going to be thinking about over the next 24 months. The right way is: what is the core competency for our business? What are the actual output metrics that matter? Then allocate resources accordingly. Part of the reason organizations got so bloated is everyone focused on intermediate metrics β "did you ship three features this quarter?" That doesn't necessarily matter for the business. It's finally coming back to: what are the business metrics we want to move the needle on?
>> Harry: Kirkland announced a $500 million spend to internally build their own Harvey or Lora. How did you think about that?
>> Matan: It's fun talking about core competencies. Kirkland spending half a billion to build their own AI tools β my understanding is that building AI technology is not a core competency of that firm. I was surprised. But I actually think this is good for Harvey because nothing like trying to do something yourself makes you realize, "oh, this is actually really difficult. This doesn't matter for us to have in-house. Let's go to an expert." My favorite is the amount of people that say "see, we told you how easy it was" β when they're committing half a billion dollars, that would suggest the opposite.
>> Harry: Brendan on a call the other day was saying the next 12 months would be the most value-accruing for AI infrastructure companies β the models and application layer would be most at risk. Would you agree?
>> Matan: I'd pretty strongly disagree. Sticking with Kirkland: we're used to a world where moat in software was "I know how to do this and you don't, so you'll pay me." The world going forward, there is going to be nothing that no one can build. Every single piece of software anyone will in theory be able to build. Is it worth your time and energy to go and build it, or should you go to someone else who has already built it? Just because I know how to walk out and pick up lunch for everyone doesn't mean that's an efficient use of my time.
There's a meme of the Microsoft org chart showing different segments with guns pointed at each other. That image is pretty accurate to what's happening with the models, the application companies, and the infrastructure companies β everyone is trying to commoditize the other. Everyone is trying to say "oh no, this one is irrelevant, all the value is going to be here." The reality is value accrual is a time-dependent phenomenon. It's not like there is one person whose steady state gets all of the value. Maybe for this next year, this person has pricing power; the next period, these people get it. Everyone β whether overtly or not β is trying to commoditize the people that are not them.
For example, we're model-agnostic. We want to give our customers the best pricing, performance, speed for whatever task they want, and we want to make sure that OpenAI, Anthropic, Google, Microsoft are all under pressure to give the best models as cheap and quick as they can. Similarly, the model companies want to make applications trivially easy to build so really the product is the model. From Mercor's perspective, it's very much in their interest that models with access to proprietary data get differentiated value.
>> Harry: What is the belief that would invalidate Factory?
>> Matan: The bear case is if one model provider gets significantly better than all the others. A key thing for us is that all the models are roughly as good as each other β one is a little better at review, one at testing, one at Python. People have a hard time keeping track of what model is number one. If one model goes way above all the others, that's a monopoly for the entire economy to worry about.
>> Harry: Is the rate of model development sustainable?
>> Matan: I was with the founder of Nebius and he said "every few weeks we see new models." I said: you're wrong. Every few days, especially Chinese open source β it's like three or four a week. Eventually we'll stop seeing them as model releases and they'll feel more continuous. Like before it was GPT-2, GPT-3, GPT-3.5, GPT-4, 4.1, 4.523 β eventually they're just going to not announce it and it'll just be "here's our model that's continuously getting better." Engineers at enterprises can't keep up with every single model, nor should they.
>> Harry: A big question is around open source β companies are spending their annual token budget by May. We see great companies use frontier models, then move to open source. Is that a threat to maintaining the market for frontier models?
>> Matan: I think it's a really important counterbalance. A lot of enterprises will realize so many of the tasks we're doing, we don't need the very frontier to do β and we can do it much faster, much cheaper with these open models. To do good resource allocation, you want to be anywhere in that cost/quality/speed tradeoff.
>> Harry: I love the memes on Twitter where it's like, "me naming a file" and it's a massive cigar with a blowtorch.
>> Matan: It's such overkill. There's a funny dynamic β an ego thing where "oh no, the work I'm doing, only a frontier model could handle." Even admittedly when I first started switching over I'd be like "I don't think an open model could handle this." And it's like, no, it probably can. It's a funny thing to mentally deal with β deciding manually or having the router do it for you.
>> Harry: Enterprises like security, reliability, ease. When you have frontier models packaged perfectly, priced clearly, secure β won't they just go for the easy option over trying to intelligently route to different open models?
>> Matan: It's easy when there's only one. But a new model comes out every week, and if you have to go through the full enterprise process for each one, it's not easy. It's also really expensive.
Probably 80 to 90%. It's typically the planning that really needs the frontier. That 10-20% could be the most important tokens β decision-making tokens. It's very similar to how we structure human orgs: leadership makes very key decisions that determine the fate of the company. They don't spend the most hours. Most human hours are spent gathering data or implementing things, but a select few hours are spent making irreversible strategy decisions β and those people are typically paid a lot.
So the frontier models are getting more expensive, you're using ultra-high reasoning for the planning. But it doesn't mean most of your tokens go there. For key steps, spend a lot. But once you have the plan, when it's time to implement β open models are typically really good.
>> Matan: There are kind of three phases happening in these enterprises. Phase one (a couple months ago): board yells at CEO, "Mr. CEO, what's your AI strategy?" CEO: "Shit, I don't know." CTO: "Hey, let's make sure we adopt AI." Phase two: AI at all costs, token-maxing, part of performance reviews β "we're going to measure how much you use AI." That was the debauchery, the long night, taking shots, using all the AI. Phase three is the hangover β you look at the bill and it's "oh my god, we are spending so much. I have no idea what the ROI is."
One of the CIOs I was speaking with realized: we've been spending hundreds of thousands of dollars per month on people asking Opus 4.8 questions like "Hey, how's it going?", "What are my macros from the food I ate today?", "What's the weather like?" We don't need the frontier of human intelligence to be doing this stuff for us β let alone it's not even work-related in some cases.
>> Harry: Will we see a contraction then?
>> Matan: We might see a short-term contraction of usage of the very frontier models. But it's healthy. Healthier to do that than to be blind and have a sudden change later.
>> Harry: Uber announced last night a $1,500 budget per individual. How do you respond?
>> Matan: I've literally seen this with dozens of our customers. Initially we'd come in and say "here are the models, go crazy" β before we had routing. Usage would go wild. They hadn't determined what parts of the codebase to dedicate tokens to versus not. We learned: we need to be having a clear conversation β "looks like you guys are spending a lot of tokens on these things. Have you thought about it consciously?" Sometimes we'll proactively set user limits.
The biggest question I ask is: Marc Benioff says he spends $300 million on Anthropic for his devs. That is 3.8% of salaries. What will that number be in three years? If it's still 3.8% β fucked. If it's 20% β fucked again. And if it's Brendan at McKay, who says he's spending more on tokens than headcount β fucked again, but more positive.
>> Harry: What do you think that percent of dev salary is in 3 years?
>> Matan: It's more nuanced than we might think. It could be as low as 0% for some individuals and as high as thousands or tens of thousands of percent for others. It depends on the unique skills of those individuals. I'm saying individuals, not devs β because the way we organize roles is going to be very different. Some people get more leverage by using more tokens; others don't really need tokens at all. If your org has a standard number where every engineer is at "this percent of salary in token use," you're probably painting with way too wide a brush. If I had to give a median order-of-magnitude: comparable to salary. Same order of magnitude, within a three-year timeline.
>> Harry: What is the strongest opinion you have that most people disagree with?
There's a very common Silicon Valley fallacy: research is the pinnacle, then engineers who implement the research, then sales and marketing and all that dirty stuff. "Oh, if only we could build a better product and it would sell itself." It's completely delusional. The product at Factory is the entire journey from the very first time they hear our name till their 10th renewal after a decade of being a happy customer. When salespeople close a deal, engineers say "we closed a deal." When engineers ship a feature, salespeople say "we shipped a feature." Name a legendary company that has a shit sales or marketing team. You can't.
It's like astronauts in space β no gravity, your muscles atrophy. Gravity will come back. If you don't have a good sales and marketing team because you don't give it respect, the second gravity returns, all of your muscles will be atrophied and you won't be able to compete.
>> Harry: Does what it takes to be a great engineer change when you become prompter and manager of agents?
>> Matan: Yes. The best engineers are the ones that don't see sales and marketing as dirty work but as an important part of the product. You're no longer just shipping features β you own full end-to-end outcomes. The parts of engineering that become less important are funny enough the things Silicon Valley brags about: competition winning, Olympiad type, memorizing language nuances. If you memorized syntax that someone else didn't, it doesn't matter.
>> Harry: I think people misunderstand the VC mindset. In a world where we desperately seek certainty, we look for validators β math Olympiad or whatever. That serves as a good crutch.
>> Matan: Yes, but it's a crutch. We have people on our team that won Olympiads and they're great. But there's something β some high schools where you "must do the Math Olympiad, must go through this funnel." That's actually anti-signal because you're not owning your fate. Then there are people from the middle of nowhere where no one else in their high school ever did this, who took the agency: "I think this is fun, I want to compete." Those are the positive signals.
>> Harry: Does she remind you of Matt Damon in Good Will Hunting?
>> Matan: I'll take that to the bank.
Growing up I was obsessed with math and physics and so jealous that hundreds of years ago, Da Vinci, Euler, Newton could be polymaths β their fields were relatively shallow. In the 2000s, pre-AI, fields were so deep β theoretical physics, string theory β you could spend 50 years catching up before contributing anything new. With AI, we're completely the opposite. These tools get you up to the frontier way faster than ever before. If you're good at thinking around constraints, holding uncertainty, knowing there are unknowns β you can be a polymath. Pushing forward developer marketing while at the same time pushing forward token caching for software development agents, while also being an incredible solution engineer.
>> Harry: What do we do today that we'll look back on and say "I can't believe we did that"?
>> Matan: For an engineering team β writing release notes. Five years from now: "people that get paid so much money spent hours of their time doing this?" Same with documentation. Stripe was the pinnacle β incredible documentation. That's going to be commoditized.
>> Harry: When agents are the buyers and you're selling to agents, how does the world change?
>> Matan: Agent to agent doesn't give a shit about UI or design, but it fundamentally cares about data structures, integrations, documentation. The best organizations who are the most agent-native put in a lot of guidance on UI, aggressively pruning anything unnecessary. The future is engineers who build the factories that build their software β like Tesla's robotic arms. Humans designed that assembly line to optimize throughput. In software, human engineers won't write as much actual code, but they build the scaffolding around the factory.
>> Harry: This is the first job you've ever had, which I think is always a funny thing to say.
>> Matan: Prior to this I was a theoretical physicist. Literally never had a job aside from physics. I was obsessed with physics since I was 12 because I was a bad student and my geometry teacher told me I had to retake geometry in high school. I never tried in school but always prided myself on being good at math. I was like "are you kidding me? I'll show her." My first Amazon order ever was textbooks: algebra 2, trig, pre-calc, calc 1/2/3, differential equations, linear algebra. I studied all of them between middle school and high school, did all the problems, then took exams to place out of all those classes.
I asked my dad what the hardest math was. He said string theory β which is technically physics, not math. So I was like, okay, I'm going to be a string theorist. That was all I cared about for the next 12 years. Went to Princeton because they had a great physics professor β Juan Maldacena. I was the first undergrad to write a paper with him. Came to Berkeley for my PhD. Only at Berkeley did it all come crashing down: "Holy shit, I've just been doing this because it's hard and because someone said I couldn't. What the hell do I do with the rest of my life?"
I stumbled on a Stanford VC podcast with a guy whose name I recognized β I had cited his paper. He was a theoretical physicist who sold a company for a billion dollars and was a partner at Sequoia. He could maintain eye contact, which is rare for theoretical physicists. I emailed him: "Hey, I'm Matan, I also used to be a physicist, wrote a paper with Juan." He responded that day, invited me to Sand Hill. A 30-minute meeting became a three-hour walk. At the end he said: "You absolutely need to drop out of your PhD. Either join Twitter β Elon just took over, it's hardcore β or start a company."
The next day I went to a hackathon in San Francisco and across the room saw this guy who also went to Princeton β Eno. We joked it was intellectual love at first sight. He's a thousand-x better engineer than I ever will be. For the next 72 hours we put together a better demo. I called the investor: "Hey, I have something cool." He said "eh, it's okay." I was like "are you fucking kidding me, this is going to change the world!" He said "would you work on it full-time?" "Yeah." "Drop out of your PhD and send me a screenshot."
My parents immigrated from the Soviet Union with basically nothing. Me doing a PhD was their pride and joy. But there was so much momentum. I dropped out, sent the screenshot. He said "you have a meeting with the Sequoia Partnership tomorrow morning."
>> Harry: You'd never presented to a venture partnership before. So what happened?
>> Matan: I made a shitty deck. We went to Sequoia HQ. I didn't even know who the hell they were. Retrospectively, I know Alfred and Pat and Roelof were all in there. I was probably dismissing some of their questions β "oh yeah, we'd solve that easily." This was April 2023, way before anyone was thinking about agents, before people were even using Copilot. We were talking about fully autonomous software development agents. The next day Sean called: "Hey, we want to give you a check." $1 million at $5M post.
>> Harry: 20% post β on the last round that'd be a $300 million position. People said you should shop it around. Why didn't you?
>> Matan: When you have a connection like that β no one else would have believed in me except him. I literally had never had a job before. No other partner would have done it. Trust, loyalty, and belief matter so much more than the price tag. We're building a legendary company β it's not just 10 years, it's a lifetime.
>> Harry: Would you tell founders to take a discount for Sequoia?
>> Matan: Generally yes. They're the best firm β especially if there's a special connection with the partner. What really matters: people that are there for you when the days are tough and when it's not obvious. When you're a hot company raising a hot round, everyone's your best friend. There's this one investor β old guard β and people told me "he's really good at making you feel good about yourself." I left the meeting being like "I'm the fucking man, this is my destiny." Then 30 minutes later: "Oh my god, he got me. He did the thing."
>> Harry: How did you get Ivanka Trump as an ambassador? Does she actually provide value?
>> Matan: Through one of my best hires β Francesca, who came from the Chainsmokers' fund (people are surprised, but Alex Pall and his partner are incredibly good investors). Francesca was relentless wanting more allocation. I joked "if you want more, you could just join us." We both went "oh, interesting" β turned out to be a strong fit. She was close with affinity at Ivanka's firm. Ivanka is genuinely one of the kindest and smartest people I've met. There are famous people who let you down β she doesn't. Incredible network, generous with time, does the "dirty work" investor help that some name-brand investors don't.
>> Harry: How does the cognition / Claude Code / Codex / Cursor market mature? Is this an AWS / GCP, or an Uber/Lyft?
>> Matan: What's necessary for the best consumer outcome is models that are separate from applications. As a consumer, you don't want to use applications provided by the same people giving you the model β incentives are misaligned. If I'm a model provider giving you a coding tool, I want you to use as many tokens as possible because I'm an API business. I don't have a strong incentive to be token-efficient. Versus an application layer that lets the enterprise decide between providers β model providers had better be the best, cheapest, or fastest.
This is different from cloud. Cloud providers said "sign a three-year deal, big discount." Then they jacked up prices, and once you standardize on one, it takes 2 years to switch. Everyone has scars from that. So every CIO I speak to is keenly aware: we cannot throw our lot in with just one model provider.
>> Harry: Help me understand Replit running the same prompt on three models at the same time?
>> Matan: That's fine for consumer use cases where you're not as cost-sensitive. For enterprise β if you're a non-technical person building an internal dashboard, you probably don't need 10 different models generating different iterations of it.
>> Harry: Are we entering a danger zone for security? A huge amount of net new code that may not be as secure?
>> Matan: Yes, it's going to be crazy. Code generated is growing exponentially; security efforts aren't growing in kind. There's a lag. Probably big incidents in the next couple years. We haven't even seen the most adversarial behavior yet β people can use these tools to be quite adversarial. The security market is really important.
>> Harry: Should US startups be allowed to operate so extensively on Chinese open-source models?
Yes. Using an open model is fine. The concern about "trigger words" β like spy movies where someone is in robot mode when you say the right word β theoretically yes, but if you were trying to make a model with that, you'd do it as late as possible. If you did it in an early model and someone discovered it, no one would ever use your models again. Not a big concern. If deploying correctly in enterprise (not as consumer), data exfiltration you can fight against. From a patriotic standpoint though, I think it's pretty embarrassing that we don't have frontier open models in the United States. I do hope we reclaim superiority there.
>> Harry: Quickfire β Nebius vs CoreWeave. Who has a larger market cap in 5 years?
>> Matan: I'm strongly biased as an application person. I hope for a world where users don't even know which one is under the hood. For me, I'd want CoreWeave to be β because Nebius has more ambitious plans to be full stack, which will eat into our plans. Businesses need to think about core competencies. If people are trying to expand beyond their core competencies β Kirkland and Ellis, have fun β I don't think it makes sense.
>> Harry: Can you sell to enterprises today without an FTE model?
>> Matan: Yes. Have a good product. The FTE thing blows my mind. For us, FTE is about acceleration β if a customer would scale to $1M in 6 months with just product, FTEs accelerate that to 3 months. If I'm sending FTEs as services, I'm not Accenture. We're not a services company. If you need FTEs to make the product work β sorry my friend, you have a shit product.
>> Harry: Anthropic or OpenAI on IPO day?
>> Matan: They're approximately equivalent business-wise. The biggest difference is volatility β there's just been more random chaotic events at OpenAI. So Anthropic.
Yes, this really upsets me. It's been not only disingenuous and wrong, but really hurt the psychology of a lot of developers and people in the world. It's for selfish reasons. If you're trying to raise unprecedented amounts of money β hundreds of billions β the best way to convince people is to say "all of capitalism is gone, the only company left will be me, so you better give us your dollars." Then when it comes to IPO, when humans now have money you want them to put in your IPO, suddenly it's "whoa whoa whoa, humans are pretty important, there will be jobs again, we like you guys." That pisses me off.
>> Harry: What's ironic is the ones who've never said it are the ones who've never needed the money. Zuck or Demis have always had a very different stance.
>> Matan: For all the philosophizing about AI and intelligence, incentive is driving the outcome, and the incentive is "I want to raise a lot of money."
>> Harry: Which legacy company has most embraced AI?
>> Matan: EY β the accounting firm β is one of our largest customers. They are so agent-native. They saw what happened with cloud, the scars of being late, and have great engineering leaders who said "this is going to be scary, but we are going to make our org agent-native if it's the last thing we do." That's brave new world.
>> Harry: Final one β what have you changed your mind on most in the last 12 months?
>> Matan: There was a brief period I thought it might be just one or two companies running away with frontier. What seems clear is it's probably going to be at least four, approximately as good. That is a win for humanity. The bad case is when there's one that's really really good. That's a hot take because right now people are enamored with maybe one or two.
>> Harry: Matt Damon, it's been so wonderful. I'll let you go back to Robin Williams.