We're here with Garry Tan, CEO of Y Combinator. Garry, really appreciate your time and thanks for having us at Demo Day.
It's great to see you. It's been a couple years. I give you credit for giving us an early heads up on vibe coding, so I'm excited to see kind of what the next thing is. But I do think, I mean, the significance of Demo Day and Y Combinator in general, you guys tend to be the tip of the spear, really leading what ends up being sort of mainstream in Silicon Valley, in tech, eventually Wall Street. What are you seeing in this batch of founders and startups that we should be paying attention to?
Yeah, the amount of hard tech companies—so literally manufacturing, industrials, defense, agriculture, I mean basically things that touch real atoms, not just bits—that's up almost 40x. So a quarter of the YC batch is actually focused on real hard tech. And I think this is the perfect reason why AI is so powerful. The way it's going to actually create abundance for everyone is when it actually solves real problems in real life, not just online. And that's what we're seeing.
I wonder if there is an AI element that makes it easier to build hard tech, physical intelligence, robots, things like that, because AI has come along so far.
Absolutely. I mean, what you're seeing with GPT-6, for instance, it literally solved the Navier-Stokes equations, right? This is something that is a 90-year unsolved equation that was solved just by zapping the rocks harder and getting back a solution. And so you can take that, and you can see that right here. This is Sable, and they create drones that can be deployed out of container ships, right? And so this company, or pretty much anyone in this hall—we're sort of in the hard tech hall of Demo Day right now where you can see physically, you know, there's a drone here, there's some skin care over there, there's things that are real atoms in this room.
And there's a robot in a military outfit. I don't know if you can see that in the shot. Who do we have over there and what are you seeing? That suggests to me that this is defense-related.
Yeah, I mean over here, OS3 is working on a robot that is actually usable for less than $10,000, which is a price point that's pretty wild. There's another one that's focused on consumer scenarios that's under $2,000. And so, when you have intelligence on tap, it means you have access to AI researchers or just even people who do solid-state modeling, they can do mechanical engineering, they can do supply chain management for you. Normally you'd have to go out and try to find the absolute best person in the world to go do that, and as a tiny startup you might not even be able to hire that person. But today, you can just open GPT-6 Astra, turn it up to max, and you have something that's not going to be exactly the same as hiring the number one person in the entire world, but you'll be in the top 10%, top 5%. You can see how that might multiply across all the engineering challenges and all the things that startups need to do to create a product that's great. That's what intelligence on tap means for society, that small teams of one or two people who find this thing—that if we could build it, people are going to want this thing—now they can build it. And so this is a revolution for hard tech.
Is it more expensive? We've talked a lot about supply chain bottlenecks with the data center build-out at least. Are some of these components getting more expensive, and what has that looked like?
Well, I mean the great thing is there will be companies that come in, and maybe we will fund them, that actually solve that problem as well. As these things develop, that's the beauty of markets, right? If there's a buyer for it, suddenly someone will want to be the supplier. And that's actually the thing that I think the world and America needs the most right now: how do we actually solve each other's problems? The most direct way to do it, especially when markets are brand new, are one or two people, instead of going off and going to work for McKinsey, they can take their time and all their energy and sit with 50 or 100 agents, and maybe they spend $100,000 or a million dollars on tokens and get the equivalent of like 100 people's work, right? But then they drive it to a supply chain, they drive it to a real-world thing like a robot that you could buy for $10,000. That could go into a factory and solve, for instance, the huge shortfall in manufacturing workers right now. If we want to re-industrialize America, we actually don't have enough skilled workforce to actually work in the factories. And so, a $10,000 robot could come in and suddenly go into those factories, and you basically produce way better whatever it is, whether it's food, whether it's clothing, whether it's something that goes into defense. Our factories can be that much more profitable, that much more effective, and that feeds into the overall economy.
And it is sort of the counterargument to the jobs discussion when we talk about robotics, especially when we talk about AI in a lot of scenarios. There does tend to be this fear from some that it's going to take jobs. You're suggesting that in robotics at least there is a lack of people to do certain jobs. I wonder how you think about the disruptive nature of AI and what it will mean for the overall economy, especially when we talk about jobs.
When I think about human beings, the beauty about them is that they can be creative. And so, instead of thinking about jobs going away, I think about it as a transformation, right? What is your day-to-day like? There are a lot of people out there who have jobs who you might argue it's actually sort of rote, like you come in, you do this back-to-back. If a large language model could do that job, is that really stretching that person? Is that giving them a creative outlet? I might argue a lot of those jobs aren't creative. So really this is a wholesale transformation of people's jobs from rote and not creative to much more creative. You sort of have to be much more creative. And this doesn't have to happen overnight. I think that the white pill for this to me is that it will take decades for this to actually percolate into society, and that's not a bad thing. That's going to be what society needs.
It's interesting that comment about taking decades to actually percolate. I feel like there has been this disconnect. We talk to a lot of CFOs and CEOs who say that they're actually not seeing AI results, or at least improvement in margin or ROI immediately. Is that your expectation, that it actually will just take a little bit longer than some of the early expectations?
Yeah, these are organizational changes. The bad version of it is you drop a little bit of Microsoft Copilot into something, and you spend more money on tokens, but you don't really get a speed-up. The reason why it's going to take decades is that's what it took for computers to enter into business to begin with, right? And it is somewhat generational. People who are in their 40s today are sort of the people who run a lot of the modern economy at this point, but we came of age with the personal computer, whereas the generation that we took over from, even in the last 5 or 10 years, they weren't people who woke up and knew what Windows 95 was, right? And so it's going to take longer, and that's actually okay.
You mentioned tokens. I want to ask, what are you seeing in terms of cost, and how many of your founders are using open-source and looking for discount cheaper versions? Give us a sense on the ground of what the cost looks like of building on AI right now.
Yeah, absolutely. I mean, when you're trying to do brand new research—so for instance, you have to do computer vision for a drone—you're probably going to want to use the absolute best models. The economic value of being able to get a brand new algorithm that is 10 or 100 times better, it's actually worth it to spend on frontier models. But if you're automating some kind of rote task, you're going to use open-source models because you don't need 150 IQs for that. You can use an open-weight model, ideally hosted in the United States. And that's what we're seeing: you get basically 90–95% of the intelligence of the absolute frontier, but for one-tenth the cost. And so it's really an engineering challenge, and people are just sort of choosing based on whether they need the intelligence or not.
How has that changed from a year ago, and can you give us any percentages in terms of what founders are using? Was it more Anthropic and OpenAI-based a year ago, and there's been more options? How has that transformed in the past few months at least?
On our end, it's pretty wild to see even just week-to-week who has the best possible model sort of shifts. For most of this year, Claude kind of came out and captured all the mindshare, Opus before it. GPT-5, 5.6 were sort of right behind, like just there, but people sort of knew, 'Well, if I needed to be absolutely the smartest, let's do Claude.' And then even in the last couple weeks, GPT-6 Astra, I find myself using GPT-6 Astra for pretty much everything. And so this is actually a great scenario. The nightmare scenario, the doomer scenario for AI, is that there's just one company, it's the only one, it has the best access to capital, it has the best AI researchers, it runs away with it, and suddenly there's just one company that's Moloch, and that would be bad. The great thing is actually we're not in that situation. And not only that, I really like seeing OpenAI and Anthropic leapfrog each other. That gives consumers a lot of choice. And then I think that Meta is on a tear. I mean, Alexandr Wang is doing an incredible job. Muse, their new AI harness, is giving open Claude Hermes Agent and really everyone else, Grok bot, a run for their money, and it's free, which is wild.
One of the criticisms of open source, at least from the AI labs, Anthropic in particular, has accused some of the companies out of China of distilling their models, offering them at a cheaper price point. There's this debate about what you do about that, because competition, as you're saying, probably a healthy thing, but at the same time these are multi-billion dollar, potentially trillion-dollar IPOs. How do you think about distillation and what to do about it, if anything?
I would do nothing. I mean, I think that intelligence is such an incredible gift to all of mankind, and you could easily argue where was the material used to train those models. I think by all accounts you could argue that there should be an American distillation regime. I don't doubt that the big model companies are doing a great service to humanity and that they're at the frontier and driving it forward. We want that to actually be able to be fundable and to be a great business model ongoing. I think that the premium on the absolute smartest models still exists. And then as long as that's true, this is actually the ideal case. You kind of want open-weight models to give people freedom and access. And so that's sort of the equilibrium I'd be looking for. If I were a regulator, that's what I would want: I want people to have open-weight models, but I also want the American labs to still be able to maintain a technological advantage. That's a tightrope, but I think it's one that could result in the best possible outcome.
