Cursor, OpenRouter, and Where the Money Went | Martin Casado

You can put capital into these things and it tends to turn into usage. Let's say you put a dollar and you get like, let's say you put in $10 to do this. I don't know if you get $9 back on the other side of that. But what we've never been able to do in the history of this industry is put in $10 and get anything back. But now it really is $10 in and then some amount out pretty directly.

Super happy to be here. Thanks for having me.

I know.

It's been pretty wild. Yeah. 10 years in the making, but it all happened in a couple of days. Yeah.

30-ish. Yeah.

Yeah. Who's counting? Yeah. But 30-ish. Yeah. Yeah.

You know, listen, I actually took time off between undergrad and grad school and I went and worked for a national lab. I actually think it's probably a good idea to get on-the-ground experience so that it's not entirely academic. I mean, the one caveat I will say is a lot of the AI stuff is pretty deep research. I do think like, you know, on one hand, you have this whole kind of outcry against higher education and it's kind of very in vogue not to do it. On the other hand, it's kind of never been more relevant, too. And so, I think we're in a bit of a schizophrenic period, which suggests to me you should do exactly what you're feeling is right for you. I think both paths are totally legitimate.

So, okay, so one thing that we forget is like the PhD used to be the rarity for founders, right? Like it used to be like you would start a company if you dropped out of PhD or you dropped out of grad school, like very famously Sergey and Larry. This was the case for, and we'd always say like you're kind of failing if you got the PhD because clearly you didn't have a good enough idea to do it. We're actually seeing a lot of really good PhD founders now, right? Like Alexandr Wang has a PhD, we did research when I was in grad school, George Frasier, uh, C5 also has a PhD. So, I would say we're seeing more PhD founders now than ever before in the history of the industry. We've always had the young tech founder, right, Mark Zuckerberg. And so, I would actually think that even the optics on X feels like it's going towards like the younger, less experienced founder. I would say weighted, it's actually probably the trend is more in that direction.

I mean, I think that when it's all said and done and we look back on this, this wave, I think the biggest change is the fact that you can actually put a lot of money to good use. Like, you'd never be able to like, so what would have happened 10 years ago if I gave you a billion dollars, what would you do?

And then you would blow up the whole thing. The whole thing would be like a total mess, right? And so now we actually know what to do with that money. So I do think that now it's become a scale of capital game. I think the companies that do raise a lot of money, they don't necessarily blow up. They know how to use it. And I think that that actually is changing a lot of what's happening on the ground for us.

Well, yeah. So there's a kind of a law of like engineering physics, right? Where you raise a bunch of money and then you try and hire a bunch of people to build a product. And this is where the mythical man-month came from. And you just can't really speed up timelines that way, but then you end up like increasing burn and adding a lot of complexity and like the roadmap gets very diluted and then you have all the organizational complexity that happens as a result. And so I mean listen, some of these major model efforts were done with very, very small teams. Like, you know, I'm not going to say the exact model, but one of the very famous models, a multimodal model that many people use, is probably one of the most popular models, was built with a team of about 20 people. And I would say the cost of that was probably $2 billion plus, right? And so this is unheard of in the history of humanity, in the history of engineering efforts. We've never been able to have 20 people, I don't think, being able to productively use $2 billion. Like, what does that even mean, put that much money to work with that small of a team and that small of a timeline? And so I know we like to look at this wave in the context of like technical sophistication, new capabilities, but I actually think like one of the major stories is the fact that we're able to apply large amounts of money productively in short amounts of time to whatever problem that we're trying to solve.

Totally. Totally. I mean, it's just, you know, again, there's always been this weird meme in venture, which is it's kind of not much of an asset class. There's only a few companies, you can't deploy a lot of money in it, which has all been very weird to me to have like this zero-sum thinking from people whose entire job shouldn't be zero-sum thinking, right? And one thing that we're learning is you actually can take large amounts of capital in private markets and deploy it and get the returns. And you know, you can just give two examples. One of them is like companies are staying private longer and accruing a lot of value that way. So that's one. And then the second one is like, you know, these models consume a lot of money and they turn that money into growth. And so this kind of almost archaic view that like, you know, venture has to be limited in capital, you know, we're overfunding it, you know, absolutely is not the case now. And listen, there's two... I think there's two conclusions you can kind of draw from that. One of that is, well, there's a new technology wave that can consume more capital, like AI can consume more capital, and so it's a result of a technology wave. I happen to think causality goes the other way, which is if you put more money privately into the private markets, it actually grows the TAM and companies don't need to go public as quickly and companies can do more privately. So, I actually think that the causality is the more money goes to private markets, the larger the markets are going to grow and then the more returns go in the private markets. But just interestingly, we're seeing both things happen at once.

When did we first... I mean listen, I think that we're one of the most tech-forward venture funds and we just kind of assume everything is going to work in the long run. And so I remember when, you know, when OpenAI was still an experiment and we had very legitimate conversations about should we be going all in on AI then, and this was like 7 years before like say GPT actually was pretty significant. And so, you know, we've always been investing in AI, we've invested in multiple waves of AI. Like when I joined the firm in, you know, say 2016, that was kind of like the drone AV wave of AI. For this generative wave, I think we're one of the earliest that were very active in it. Like we did a lot of the early foundation model companies. Right. We're in ElevenLabs, we're in OpenAI, of course, we're in Cursor, we're in Ideogram, we're in BFL, we're in Mistral. So we just did a lot of very, very active deployment. But I will say like the intent to deploy was even before that. Now again, I would say that like our conviction on technology waves tends to actually precede quite a bit the on-the-ground gains, which can sometimes work and sometimes not work. But in this case, it was actually quite prescient.

Yeah, by the way, every single way. So this is why I know the difference between like a finance-based investor that should be in a growth market and an early-stage investor. So like if you think primarily in the balance sheet and in finance, you worry about things like margins, churn, revenue quality, which are all very legitimate things to worry about if you see no strategic value in the business, right? So if you think like the entire business, forget the tech, forget the strategic value, forget the ability for somebody else to monetize it, I'm just only going to look at the balance sheet, which many, many people do and make that mistake, then sure, look at, you know, if you look at that, you'll have like difference of opinions, but that's entirely divorced from the strategic value of a company. Like some companies may lose money but they'll get to a position where they can command a tremendous transfer of value. And if you don't look at it that way, I just don't think you can be an early-stage investor. You have to look under the lens of this is a very important part of the new stack, it's a very important control point, and either somebody will want to own that because it'll accrue value or it itself will be able to accrue value even if right now when you look at it the financials don't look like something that would trade highly on a public market.

Well, I don't know if we don't believe in that. I mean, listen, we actually had an offsite recently and one of the big discussion points was like what is the future of models. And I think there's basically two paths that are meaningful to talk about. One of them is that the labs win everything and then the other one is the labs don't win everything. And you can make very strong arguments on either side of that. So, if you want, I'm very happy to detail the arguments on either side. Let me give you the argument for. So, why will the models win everything? Well, the first one is we've got three years of data and they own 95% of the market. So, if you just look at the data, you're like, well, okay, so they're doing tremendously well. They have the majority, certainly dollar-weighted, of revenue. The second one is that their ability to raise capital is just unbelievable. And then they can turn that capital pretty directly into growth, which we've very rarely seen. So, let's say OpenAI and Anthropic have raised $240 billion, I don't know, $220 billion. That's more than the entire downstream ecosystem combined, which is really unbelievable. There's another argument, and that argument is like, listen, they've been able to maintain pricing power just by being an epsilon higher. So like being on the frontier is expensive, but also being on the frontier can command a tremendous amount of pricing power. And so it's not like you have to be way better. You can just be a little bit better because these get used in competitive equilibria. We're starting to see autocatalytic effects. What do I mean by autocatalytic effects? Like if you have sophisticated enough AI, you can use that AI to create more AI and that improves your unit economics. I also think if you talk to a lot of the people deep in these labs that are working on these, they think they're actually getting further away from open source than not. And the benchmarks don't reflect that because benchmarking in AI is kind of an AI-complete problem. Like you just kind of have to use it to do the thing to know how good it is. So there's a lot of data that would suggest, listen, they have the capacity, they have the capital, they have the teams and the talent, they have the best models, and they're going to run away with it.

Well, so yeah, this is where I'm just like an annoying pedant. So to me, recursive means something like, I'm a, you know, this, my PhD is in computer science. I've done programming languages. Recursive is when you take something and you make another copy of that thing wholesale. Like if I have a compiler, code could, a compiler that would be recursive. Where autocatalytic is, is use the thing to, like, you know, help you make that thing faster as a tool. Like if I use a computer to design a chip because I'm using like whatever, like some chip design software. Would you call that recursive self-improvement?

No. I mean, well, not in any sense because we've never used that word. Like, listen, we all use software to build software and we never, we don't call it RSI. And we never have. And so I feel like what we're seeing today is much more autocatalytic effects, which we've had forever, but for whatever reason, I think this is a legacy of Bostrom, we call it RSI, recursive self-improvement. That said, there's definitely focus on recursive self-improvement and that's great, but it's a subset of the broader phenomenon, which is very important for all of us to understand, which is autocatalytic effects. So just give me an example. If I use AI to create a really good GPU kernel that's faster than a human being can do and that improves my ability to like run or build AI, I would say that's autocatalytic. I would think that's very important for us on the economics, the convergence properties of the industry, but it's hard to call it RSI.

Oh, another one for it, by the way, is they own all the supply.

In a supply-limited market. Okay. So the arguments against, the arguments against is the surface area is really expanding. And I mean, really what's working are primarily code and language reasoning. And so there's a bunch more domains where actually probably you need a services arm, probably you need kind of more connection with the customers, and this would be very hard for a single organization to do. Second one is, you know, like there is a wealth of open source models that are doing very well and we've seen some recent examples of that. And there's a maturing ecosystem around them to serve those. I think a lot of the reason the labs are so far ahead is they have very cheap access to capital, which will almost certainly rationalize. Like listen, if I give you free money and I say you're worth like whatever trillions of dollars, like of course you can use that to grow, but at some point in time I may be like, you know, you're probably worth, you know, something different because we've rationalized the company and then it'll be harder for you to grow. I think that right now they're taking advantage of the supply constraints because they can actually buy GPUs in bulk. When you have more supply, they won't be able to do that. So I think actually the landscape right now is very much in favor of the large labs, primarily from funding and supply, and once those rationalize, the growing surface area is likely to fragment, which almost always happens in the history of this industry for sure, and I think that will play in favor of everybody else. If I were to guess, this is Martin totally guessing, no idea. If I were to guess, I'd say supply constraints will ease in 2028-ish. I think that the big labs will probably, dollar-weighted, get 80% of the market going forward because that's historically what we've seen for large incumbents. But I think token-weighted, 60% will be long tail and open source. And one more thing about all of this, I'm sorry to kind of rant on it. I just think it's such an important point. We're seeing increasing value go to the actual apps. The apps are doing incredibly well and so I think they will start to actually erode on some of that margin share, and you don't view them as necessarily competitive with the labs. But I do think that we're going to capture more and more of the value going forward, but it'll be a couple years.

Yeah. So listen, so OpenRouter, OpenRouter is many things. I mean, it's an API router, but one thing you could think of it as is, you know, maybe you have your Anthropic key, maybe you have your OpenAI key, maybe you have your Grok key, but for all the other models you use OpenRouter, right? It's basically a two-sided marketplace to give you access to a bunch of models. It's a single place for visibility, it's a single place for analytics, it's a single place for access to these models. It's on the token path. It is the leader in kind of a brand monopoly in the space from a brand standpoint. It's a very well-known brand in the space. And I think with all of these phenomenons where you have a couple of incumbents, somebody has to aggregate kind of the longer tail marketplace, which I think will impact the dynamics of the model market going forward for sure.

You know, it's a good question. I go back and forth on this. Like, we've actually learned that these models are a lot stickier than people assumed. Like, everybody talks about just swapping them out, but it actually doesn't happen very often. And I think a lot of this could be actually like procurement dynamics, like I bought a bunch of credits from OpenAI, like why would I swap them out? So you can think about like smart routing in one of two contexts. One of them you can be like, you know, almost from like a quality standpoint, like you'll answer the question better on the right router. I think that's a very, very hard problem and maybe an AI-complete problem, and right now I think that doesn't happen very often. You can think about it in a second way, which is basically cost performance, which is like I will minimize the cost. And because tokens are so expensive, that's actually kind of a high priority for many. And that we're seeing great results, right? If you've seen like for example, Cursor Auto and Cursor Router does a great job, OpenRouter does a great job. And so I would encourage, you know, if you think about smart routing, I think it in the context of, you know, how do I for the given task choose the right model on the Pareto frontier relative to cost as opposed to like I'm somehow going to choose the right model that's going to answer the question in the right way.

I think it's AI-complete.

By AI-complete. Okay. So, let's imagine you're trying to answer the question, what question does the smartest thing in the universe need to answer? Like I think you need the smartest thing in the universe to answer that question. You see what I'm saying? The only thing that can answer that question is actually like the smartest model, and then in which case you just give it to the smartest. It just becomes a cost optimization type of question. Well, listen. So, okay. So, there are many opinions. I don't think we know the answer to this. Like clearly, and you hear this often, clearly the frontier models are getting very jagged. Meaning, and this isn't a core technical capability. It's like the decisions of the people that created it decided to make it very good in say front end or very good in 3D or very good in language. And so it's not like this one's smarter than this one. It's like this one's very good at this type of thing versus, you know, very good at this type of thing. That for sure I think we have a lot of specialist models, a lot of, you know, and then a model router will just decide this is better in code, this is better in language, this is better in front end, this is, you know, whatever. So like that will happen, but I think today if you look at like model routing, the gains, you know, as deployed by the application companies, you know, I mentioned Cursor, but we've got a huge portfolio of application companies, I think the primary gains actually tend to be keeping the quality bar high while decreasing the cost.

Listen, I think that you need a two-sided marketplace and in many ways you can view one of the products of OpenRouter is the demand, right? So I can go to the next model provider and say, hey, listen, we've got, you know, millions of users all the time and we can bring that to you, and this is a very common thing. And then you can go to all the developers and say, hey, listen, you know, the next time a model comes out, it'll be on here and you'll have access to it on your existing system and so forth. And so I think there's a tremendous amount of value in the two-sided marketplace, independent of all of that. That said, clearly over time there's going to be more and more ability to do this sort of arbitrage. I think for whatever...

And it's a hard question to answer just because we don't even know how to benchmark these models to begin with.

The pricing is dynamic, it's often subsidized, and you know, and then, and then honestly, like the next big model comes out and it tends to be Pareto-efficient everything anyways, and so like then they would route to like one model like Opus or GPT-5 anyway. So I think right now that like the actual routing piece is, you know, the gains are uncertain, and yet this is a very, very popular, very successful, you know, internet property and brand, and I think it's more the two-sided marketplace.

I mean, right now these guys get free money. I would use it too, right? Like, you know, if I could raise, you know, three times more money than the entire downstream ecosystem combined. I mean, you know, like this is a very classic thing. Of course, you'd subsidize in order to get kind of single users. And by the way, all the margin tends to be made in the enterprise anyways, right, for this business. And so why wouldn't you do that? And by the way, this is...

By having run businesses, marketing is hard, man. Like, you know, like it's really hard. Field marketing, content marketing. You never know if I put a dollar in what it's gonna get me. Like events, like it'll maybe get leads. Will those leads qualify? I have no idea. I'm gonna do this content. Like it's got this atmospheric value, but like it's very, it's always been hard to like spend a marketing dollar and be like, 'This is going to be this many sales-qualified leads,' except for in very mature organizations. A remarkable thing about AI is you can spend a dollar and get users because like there's unlimited demand for tokens. And so in many ways it's kind of disrupting the entire marketing. And so this is what we're seeing. We're seeing like if I'm a company and I want to attract users for the top of funnel, why wouldn't I subsidize?