Will Open-Source Threaten Anthropic's Business & Do Margins Matter in a World of AI | Matt Murphy

Like, it's great to be here. It's taken me seven years to earn my way back onto the show now that you've become so famous.

Yeah. Well, he'll be mad if I don't give him a call out, but Anjan, you know, Mida, was the one who introduced me. So Anjan worked for me, with me at Kleiner Perkins when I was there as a young kind of associate, but he was so spiky at the time. So he's always kind of just been in the flow. We were talking about AI and he said, 'Hey, Matt, you got to meet Dario and Tom. This is the one company.' You know, I said, 'Let's do it.' Got on the phone with Dario and Tom the next day and I, you know, I personally was like, 'All right, I'm in.' And I'll give you the kind of like the broader story, but there was part of it that was really easy and part of it that was hard, as you can imagine. So, you know, at the time you have like a $600 million venture fund. You kind of try to average $15 million into a company, and along comes a company that's like pre-revenue and, you know, wants a $4 billion plus valuation. Too early for our growth vehicle. Where does it kind of fit? But you know, the easy part was, okay, OpenAI is absolutely ripping, ChatGPT taking off, but Dario was the creator of that within OpenAI. As you know, the reason why he left is because basically he's like, OpenAI is doing too many things. This is the one, this is the one big opportunity. So you had that kind of like unique insight, knowledge, conviction around this opportunity. You meet him and he's just like this amazing technical thinker, researcher. A lot of the best researchers want to work for someone like that because it kind of mirrors them. It's like that's the leader they gravitate to. And then, you know, another easy part of it was they had, you know, they had basically, it was pre-revenue, pre-launch of the model, but all the benchmarks you could see that they were kind of better or at the same level of performance as ChatGPT at the time, and they'd spent like, I don't know, a fiftieth of the capital. So these compute multipliers, you're like, all right, there's something special under the hood technically. And my partner Tim Tully, who was the CTO of Splunk, great, thankfully, you know, part of the team we've built out here, I had Tim to kind of dive in with Tom. Right, so that's all kind of like, look, this is a massive market. These markets are never dominated by one player. There's going to be an alternative. Who's better positioned to be the number two player than Anthropic? The hard part was what I mentioned, like, you know, wait, why are we doing this? It's like a $4 billion valuation of a venture fund. That's not what we're going to, what we should be doing, what our LP is going to say. But fortunately and great,

I don't remember exactly that part of it. But basically, you know, if there was a mistake, and it's hard to look at this through the lens of having made any mistake, it's basically like, hey, look, the opportunity is there for you to lead. But I'm like, well, you know, we can't really do this out of the growth vehicle in the venture fund. We can only do so much. So we said, hey, we want, we're all in. We want to be part of the round. And I'm very grateful that I have a set of partners who were just like, look, let's just do this. Let's just get into this. This is one of the biggest waves. We've pivoted the firm to be all in AI. Let's jump on this thing and see what happens. And that led to everything from there. But if I had a partnership that was more rigid around, hey, that doesn't fit, then this never would have happened, and we would have never gotten to the point where we led the next round and all of that. But anyway, so that's kind of the quick story of it. Very fertile.

The first check was a little over 10. And then, you know, so that was kind of the starter check because the average, like I said, you try to kind of in a venture fund kind of have this kind of narrow range of what you invest. But then the next round is when we did the 500 plus SPV.

Yeah. Yeah.

Well, first of all, I'm glad you weren't in my partner's meeting. But seriously, I mean, there was that in the room. And at the same time, you know, I had a couple other partners, and this is what you want. You want to have partners that debate things you listen to. But it's like, look, there's never going to be a perfect entry point into this market. If we wanted to be in this market, this was the way in. If we said, 'Hey, look, we're just kind of priced out. We can't be in foundation models or labs of any kind,' then, okay, you sit on the sidelines. But we were like, we have to be in this market. We're building the firm around AI, and this is absolutely the best company. So just don't overthink it and get in. And honestly, I think that's been really a hallmark of how we've operated. I think other firms can be, and not to throw any shade at anybody because I have such great respect, but you know, you kind of get into these situations where we have to own 15 or 20% ownership or we don't do this and don't do that. And I think the new Menlo that I'm part of has shown extreme flexibility to just do what makes sense. Let's get in this great company because once you're in, hey, if it takes off, there's plenty of opportunity to put more capital in.

By far. I mean, look, if you can get ownership, it's magical because, you know, just if you own a lot and the company's worth a lot, that's going to be great. But, you know, there's a lot more capital coming in, so it's hard to even maintain that kind of ownership. But we're in an outlier business right now, right? Like, I think for a long time, I mean, you know, I've been in the business for 25 years now. You know, you were kind of saying like, 'Hey, great outcomes are 300 million, 500 million, a billion.' Like, so you're like, 'Hey, you have to own 20% to get a 100 million or whatever.' Like, no, that's not, those are like, and I know you talk about it a lot on your show with Rory and Jason, all that. That's not how the game is being played anymore. It's like you have to be in the big outliers to drive great returns, and you're better off being in them at a very small percent than owning a large percent of a company that exits for 3 to 500. Those just aren't going to move the needle.

Well, I mean, we don't, you know, we announced our new funds, so we're not, you know, we're pretty full stack. We can take big concentrated positions. Fortunately, we've got LPs who like to co-invest with us, but you know, we don't have a 10 or 20 billion fund, nor do we aspire to have that. So there's some quantum of capital that's like, hey, that's for somebody else, the next round. But I don't know that it's as much of a valuation thing. I think it's more because I would rather, you know, to be in the most amazing company. I would rather be in than not.

I think it's pretty rare, as you know, to find companies these days that don't end up raising a lot of capital that way outside of the frontier company. Look anywhere in the AI stack, even the application companies. I mean, there's part of it, the companies are growing faster than ever. So they want the capital to, you know, really be able to play offense. And there's also kind of a part of this dynamic in the market right now where there's this signaling effect that every X months or a year, you know, you raise capital that's, you know, employees want to hear that to keep up with the labs and, you know, some of the retention, you know, you have to do more secondary. So there's just the landscape is just very, very different than what I grew up with.

I know it's terrifying. Can I ask you on the second round that you mentioned there where you're like, 'Okay, we really sized up.' How did you think about that one and how did that come to be?

I mean, like if there was a playbook that I would love to repeat, it was this. I mean, so we basically built a relationship, guided the company and said, 'Look, we need to go all in.' Menlo style, you know, our recruiting team, our BD team, and just get close to the founding team, see, build relationships, see how we can have value. And there's a lot of examples of that that we probably don't have time to go into, but we got to know them, and we got to see them operate, right? So let's say that the round closed in something like March, the model was launched in April. So you start at zero and then sometime, you know, through the year, you know, you'd see them adding 10 this month, eight the next. So the revenue started to build. In parallel with that, you had Amazon and Google come in, both with the big investments as well as technical partnerships around Bedrock, Vertex, and then distribution relationships. So you're like, okay, let's take a look at from when we invested to now. They've got a capital partner, a distribution partner, a technical partner, two of the biggest in the world. They're the alternative to OpenAI, who's kind of tied to one cloud with Azure. So it's like, hey, this is the multicloud provider. And then you just saw this kind of revenue drumbeat start. But the seminal event was we held our LP meeting in November and we had an Anthropic executive named Nero, who's kind of a jack of all trades and a very valuable one at Anthropic, come and present. And he blew everyone away. Like after the meeting, our LPs were like, 'This is crazy. Like this company is amazing.' Even my partners were like, 'This is so amazing that we're in this company.' It was just the description of like the power of the models and how it was impacting so many applications already, human behavior, all that. And we had had a bunch of inbound leading up to that. So we literally came out of that meeting and said, 'All right, we've got to do this. We've got to figure out a way to lead the round.' And two weeks later we signed a term sheet. We, you know, aggregated all the kind of demand from our LPs and folks we knew, and, you know, the rest is kind of history.

Yeah. I mean, I think it's really like what guardrails or kind of parameters have you set on your fund in terms of how much you want to put in a fund. So if you know you've got a billion dollar fund, you might say, 'Hey, we only want $100 million max in a company.' But look, we feel super, we maybe we did 50 in the first round and we want to do 100 in the next round, so we can't put it all in the main vehicle. So let's do an SPV. So I don't think you have to do it. I think oftentimes it's valuable to be able to do it because you can play offense if you need to write more capital to win a round, and obviously it can be helpful to a company that you come with more strength. You know, I mean, there's a side of it where you can say like, well, look, it's kind of extra economics at times to go outside your fund mandate and be more full stack and not let somebody else take it. I think for the most part for us, it's just like let's kind of keep our fund size at a level that we think makes sense for the environment, and if an amount of capital per company goes outside that, then let's bring in our LPs.

Now, you know, you can argue you might have some LPs, some, you know, if it's an older fund, some dynamics like that where you want to give liquidity, but that would be like maybe you take 10, 20% off the table. But for the most part, if we're in a winner, we want to run, we want it to run, we want to put in more capital. And then at some point, you know, you feel like the company's maturing or maybe they're waiting a super long time to go public and you'd like to say take some chips off the table, but it's not something we spend a lot of time on. We spend more time obviously on hopefully making great investments and then being a great partner to those companies as they scale in all ways, including capital.

Yeah, I mean, I'd say like, well, maybe I'll go back even, I'll expand your window to 24 months. Like, you know, when we did the SPV, it wasn't, Anthropic wasn't a household name yet. Like we saw everything going on and like how amazing this company was, but from the outside it wasn't quite as obvious. So, you know, even to get, you know, the whole syndicate that we pulled together and I had to give my friend Ravi and Byron a call to bring them into the round as well, which all worked out. But it wasn't, it was just that was very nerve-wracking because Menlo had never done an SPV before.

It just happened to be over 500 million. So you can imagine like what I never, and by the way, it gives me great empathy for entrepreneurs, which I have anyway because I understand how hard this is, but like being on the front lines, having to be the person kind of, you know, capital raising, talking to these investors, getting, you know, an occasional turndown, having to answer second and third order questions sometimes annoying, no offense to anybody, that's tough, man, that's really tough. So that was my most nerve-wracking, but at the same time coming out the other side of it, the most exhilarating and obviously all that work was very worth it. I'll run through a couple other deep, the DeepSeek moment, you know, that was like, oh my god, what's happening now? You can't even remember that. Then there was the, then there was the DAO, the DAO moment and, you know, it's just like this environment is so dynamic, right? Like everything's moving so quickly that there's just like a new challenge and opportunity, both crisis and opportunity seemingly every six months or so.

Oh, you mean the thing that came out recently around people doing SPVs, not my...

Just to be clear, we partnered extremely well. It was great. I think the problem is, you know, it's secondary markets, SPVs, they've just become too annoying and aggravating in the market to founders, and someone else is basically like, I don't want you marketing my stock. I want to be the one who's figuring out who's in the cap table, who's an investor. And, you know, I think that there were a lot of people, you know, claiming they had access who would kind of round up people to invest in their SPV and then they would try to go get access. There's just a lot of bad actors out there. And so I think it needed a bit of a, you know, a salvo across the bow to just kind of be like, 'Hey, settle down everybody, because if you're not directly in our, you know, in partnership with us, you shouldn't believe this is real.'

Mhm.

Yeah. Well, I mean, you know, that was another wild story where you see a company go from zero to something like 300 million in a year. I think we intercepted them around, well, we kind of tried to get in when they were around 30 of ARR, but the round we did was around 150. So, I mean, look, you're kind of looking at like, this is a phenomenon. So there's numbers and then there's the market and then there's the founder, right? So the numbers were just like ripping and you're like, 'All right, so this company is going to go from 0 to 300 in a year.' Even if you assume it decelerates to whatever, you know, a 3x growth rate, that's 300 to a billion. And I'm talking about when we first made the investment. And then, you know, you compound out from there and you're like, how, I mean, never, just certainly in the first, let's say, 23 years of my venture career, you never saw anything like that. Now there's a few more examples, but clearly this was an outlier even amongst outliers. I think the thing that we also really gravitated to here, aside from like, you know, Anton, he's very visionary. He's kind of the voice of the category. I think he's got some very unique and distinctive plans about why, you know, this kind of 99% of people, as he likes to call it, everybody who was never a coder and programmer, but making everyone become creators. So you had like this massive vision. We felt like an iconic entrepreneur and then like crazy numbers that you could, you know, you could do whatever model you wanted and you're like, look, if this thing keeps compounding and this is really the company that we believe this will be, one of the most valuable companies of all time. So, you know,

They do a lot. And, you know, we're in this kind of tricky period as investors where right now a lot of great companies have low margins and, you know, let's say like 20 to 30% margins and, you know, they all probably have a path to get to 60 or 70. Well, you know, a lot of companies just because the cost of compute and inference, it's harder to say you're going to be an 80, 90% gross margin company anymore. But, you know, great companies are, you know, 60, 70% gross margin. But, you know, the path to get there is like, hey, I'm going to do some optimizations. You know, I'm not completely tied to, you know, to inference around, you know, my cost structure and I'm probably going to do something complementary to the leading labs with my own data and build a model that kind of gets my gross margin up.

Yeah, I mean, first of all, I think like Anthropic is a fantastic partner to Lovable and vice versa. But like this market is so big. So there's really two dimensions that one, you know, people, some people worry about Lovable and Anthropic tripping over each other. I think Anthropic always comes at things a little more like the technical user and Lovable comes at it more from the lay user. And sure, there's probably some overlap in the middle, but I think there's plenty of big space for each one to do extremely well. And, you know, look, Cursor was about as in the crosshairs of Anthropic as possible and they, I think they still had a pretty darn good outcome. But the whole open source topic, look, it's like any market when you start off in a certain way and it's just like, look, I want to get something running. I want to get it out there and just prove I've got, you know, a cool product. And so you just default to the simplest thing. Over time, you do more optimizations, right? And so I'm also on the board of OpenRouter, a company that you all talk about quite a bit, and I love hearing you guys mention them. And, you know, that's kind of like the north star there is like, hey, you ought to have some intelligent layer that intercepts an API call from any application and basically says, what's the best model for me? Like across whatever efficiency frontier I'm trying to optimize for, is it price, is it reasoning, or is it performance, you know, latency, things like that. And at scale, like that's the kind of stuff you need as a company to manage and optimize your business. And so wave one of AI is like, let's just get it going. Wave two is like, let's get a lot more sophisticated about what we use and when and how.

No, I think the foundation models, especially, well, let's say specifically Anthropic, have such special models, performant, intelligent models. This can be hard for somebody to just kind of say, I've used open source with my data. It's going to be functional and positive for some amount of what you're doing, but I just don't think it can be powerful enough to really, you know, displace it. So my mindset generally would be like, you're going to use multiple models. Let's say if you're someone, pick a company that maybe you use 50% Anthropic and 50% open source in your own model. I don't think it goes to that, you know, well, you were talking more cost, but I don't think it goes to that 96% because what's happening is companies see this like, yes, I can get lower cost, but if I use Anthropic, it actually increases my customer retention. I generate more revenue. I get users to engage with the platform more. And that is what the data is suggesting now with a lot of application companies. But there's certain API calls that just don't need that level of functionality. And frankly, it's good for everybody. It keeps Anthropic on their toes to keep innovating. Most innovative company around. So they'll keep innovating, not stay still. And then startups innovate in their own way with open source.

Well, I mean, I think it's like any product. You know, you can argue that the cheaper it is, the more it kind of opens up the market because it's, you know, you can do more for less and that all, you know, those economic curves always spark activity. But, you know, I mean, look, even within the Anthropic family, right? Like you've got Sonnet, you've got Opus, you've got Haiku. So I mean, even Anthropic itself is innovating around, you know, hey, it's not one size fits all. So I think, you know, you're going to have the combination of something like that, a family of models from Anthropic and then a set of open source models and things that you train with your own data. And you're going to look across that whole tapestry and say, hey, I'm using 50% this, 30% that, 20% this. And those are the kind of optimizations that happen at scale. And that's the stage of market that I think we're just entering into, which makes it a lot more fascinating, frankly, because there's going to be so many kind of second and third order companies that spike and take off versus, you know, the whole market being concentrated.

Well, I think it goes back to, you know, what I said about, you know, optimizations. I mean, you know, Google with their TPUs a long time ago, Amazon with their Trainium. I mean, just at some scale, you look at your bill and you're like, I'm paying somebody way too much, you know, and you say, well, I'm willing to pay that for some part of, you know, my COGS because that's just so much better and different and I can't compete with that. But maybe there's some other types of activities they're doing that I can really leverage my own technology and bring my cost structure down. And, you know, I mean, the chip business is hard. Good luck wading into that, right? You know, it takes a special team, especially if you're going to compete with Jensen and a lot of other options out there right now. But, you know, these companies are smart and they're looking at like, hey, look, there's some very specific thing that we do in our model that if we had a chip that just behaved like this from a, you know, I don't know, from a memory cache, whatever, like it would make us so much better. And I'm sure for some percentage of the workload, whether it's in training or inference, that could be a big deal. So that's probably worth the swing. You know, if you're a hundred billion dollar revenue company,

Now in the routing business.

Well, I don't like, first of all, what OpenRouter has is like they've just got this groundswell of organic activity with developers who come to them because they trust them. They know it's a great inference marketplace. They love their intelligence. Like I don't think a ton of developers flock to Nebius. Like if I'm a developer, I don't wake up and be like, 'Hey, you know,' so they're kind of in the wrong part of the conversation. But if you're on Nebius and they're your underlying provider and they provide routing, okay, fine, you know, but if you're a company that's building and thinking about multiple cloud platforms and you want to kind of even obfuscate that, then, you know, OpenRouter is a great solution.

I mean, their trajectory is insane. I mean, I forget what they've publicly announced. So I better not say anything, but like this company, wildly profitable. And, you know, at a scale that would probably shock most people before this whole open source model, alternative model, model optimization market really takes off. I feel like we're just on the cusp of it. And this company is already a beast. So I have massive and very high hopes.