Mike Volpi on Why AI Breaks Traditional Venture Capital | Ep. 52

[00:00] the kind of classical marketing efforts that have been done for brand building, particularly in venture, sound off-key to your average 22-year-old entrepreneur. I think what does resonate is inside knowledge, tips, connections, network, all those things which are kind of indirect but organic ways of building brand. Yes, brand is important, but I think you want to build it organically. All right. I am really excited to be here today with Mike Volpe, who is probably one of the most successful entrepreneurs. [00:29] venture capitalists of the last couple of decades and now building your own new firm. And I'm really looking forward to learning from you. So thanks for doing this with me. Are you kidding? It's a pleasure. And I've already fooled you. Well, that's one. One of the things I want to start with talking to you about, which I thought would be kind of interesting just because, you know, I also gave my shot at it. It's just like building a new venture firm. You obviously had had this incredible run at, you know, a great firm at Index for a long time. And now you're building your own firm, which you've been doing for the last, I don't know, year, year and a half, year and a half. [00:59] thoughtful about how you want to do it and what the trade-offs you want to make are and what the firm design is going to be and what's the strategy. So I guess I want to start when you were thinking about [01:08] creating a new firm. [01:09] What did you sort of start by thinking about? What were like the key considerations that you got going? The first is somewhat obvious, but very important, which is it's very hard to disrupt a market, to break into a market unless there's something macro that's happening. That's an enormous change. Obviously, AI is that. So the first thing is to take a firm that's new and deadly focus it on whatever gigantic wave is hitting the industry right now.

[01:39] spread the peanut butter thing. It's just never going to work. So first, A, identify that there is a massive trend to absolutely focus on it. The third thing I would say is gather people who are not only fluent, but just have grown up and live whatever this new trend is. I do think that there is a big generational shift right now between sort of classic entrepreneurship, classic venture, and this new generation of AI venture. And some people are more predisposed to it than others. [02:09] be very. [02:10] careful about your past success. This is, I would say, very true for individuals, for yourself, and it's also very true for firms. And the more success... [02:22] a firm has had, the more, to use an AI term, reinforcement learning there is of how things were done. And if the world shifts to a place where things are done a little bit differently, that reinforcement could be applied very incorrectly. I actually think this past success reinforcement loop thing explains why a lot of, for example, execs from pre-AI SaaS are like, it's really difficult to adapt because you learned a whole set of things that don't make sense [02:52] is changing. Basically, most venture capital firms are focused on making money on software companies. And that foundation is based on the idea that software is complicated, expensive, and takes a long time to build. And it costs very little to make lots of it, but it costs a lot to make the first edition of it. And so then you're in this world of, I have a high fixed cost thing, which I need to sell to as many people as possible. Every business model starts looking like that. Then you

[03:22] era. [03:23] which takes the cost of making software way down, you're completely shifting the core assumptions on how a business is built. And that then extends into everything from go to market, engineering, fundraising, which customers you target first. How do you target your customers? Should it be a product centric company or a service center? All these assumptions kind of blow up. So if you've had sort of a firm as an investor that's been structured to invest in a certain type of company with a certain type of people, [03:53] that base assumption changes. The cost of making software is now super low. You blow up everything. And so if you get stuck with the old way of doing it, you're probably going to invest in the wrong companies. You obviously invested in a certain way [04:04] over the last years at Index, and you had a certain, I don't know, ownership model, stage you like to do, all the rest of it. When you came in with Hanabi and you're like, I've got a blank canvas now and I need to not train on what I did wrong. I want to learn from the good stuff. I want to not bring the bad stuff. What did you think about from a firm design perspective when it comes to, let's say, the types of deals you were going to do, the stage, the ownership, the size checks, the dispersion of all that stuff? Well, let's start with the people. [04:34] is a relatively complex and deeply technical field. [04:40] And therefore, the people that I have at the firm should be native speakers of that new dialect. The number of people that were focused on AI and machine learning pre-2018-19 is incredibly tiny. So essentially, the volume of people that are, I'll call it, native human beings have a professional depth of five to six years max.

[05:10] 2024, let's say. Reasonably substantial, but not massive, I would say. Like you can catch up. You can catch up. You can catch up. But, you know, you got to have a foundation like how does it work? Why does it work? What are the flaws? How quickly does it evolve? Do you understand compute? Can you talk semiconductors? What's a GPU versus CPU? Why? Memory, high bandwidth, like all this stuff. Like unless you sort of are genuinely fluent in tech, these are not conversations that you're going to easily have with your counterpart, the entrepreneur. So you got to look for people that [05:40] and are pretty damn young. You know, when you build from scratch, the complexity is finding such people in this industry that want to be VCs. It's not the easiest thing in the world. But A, you look for that. The second kind of idea that's very important is in venture capital, we've talked a lot about stage centricity. We are an early stage firm. We are a mid-stage firm. We do growth. We do pre-public. We're a crossover. Everything is defined by stage. I don't think [06:10] I could invest in a company at 10 billion in valuation and three years later, it could be worth 380 billion. In essence, what you have is a suite of companies that we would now consider growth stage. Yeah. But that represent venture like return characteristics. It's also probably the flip side of that is, you know, there was probably a point in time when investing in something that was worth 10 billion. The downside risk was also probably very different than it is today. Yeah. But honestly, like you can't sit around VC and worry about downside risk.

[06:40] you're going to lose your money. No, I'm just thinking about, you know, there were growth investors who at one point in time could, you know, they could, I'm going to get my three X's when it works and I'm going to get my money back when it doesn't. And now it seems like there's still venture outcomes from much later stages. Absolutely. I mean, when I, when I got started in this industry, growth investors would be like, oh, I'd like a three X lit prep and I might triple my money and whatever. And now I think you can genuinely go into a company at, you know, you [07:10] So I think you have to be ignore stage, observe the size of the opportunity, the magnitude of it and so forth. And in some cases, you're finding two people in a garage and you're helping them build the business. And in some cases, you're piling in Anthropic at three hundred and eight billion. And they they have literally all of those have 10x plus. [07:30] possible outcomes. So I think you just don't pay attention to that. I want to stick with that topic for a second. I think it is an idea that is maybe still not the mainstream idea, but definitely more people are operating in this way where you can sort of choose what boundaries you create. The stage one has gone away in favor of other things. Do you think that you're able to do that as a result of experience or do you think a early career investor is also equipped to both traffic in pre-seed first money investments and also make [08:00] Let me phrase it this way. On the assessment side, so being able to [08:04] that effectively value very different stage companies can be done by the same person. It's not that hard. Like, you know, looking at a spreadsheet is not rocket science for the average engineer. What is trickier? One of the things that I still think in this era and in prior eras is important is the proximity or affinity you have with the founders. If you're investing in two kids coming out of OpenAI right now and they're 22 years old and you're a young new investor, you can definitely do that.

[08:34] it's not likely that you'll be able to establish that relationship with Dario. So I think the advantage I have is that I can play that, [08:43] you know, relationship game with some pretty senior folks because of my experience, but also appreciate that I need to build that relationship with a 25 year old. Now, some of the younger people on my team don't quite have the ability to go out to Daria or Sam and establish a relationship with them, which is understandable. My job is to start to open up that pathway for them so they can. But I do think the assessment of the fact that, you know, Cerebris is [09:13] Or that, you know, open AI might still be that anybody can do is just look at the math and see what it looks like and expect some level of multiples. It's not that hard to make the decision. I do think it's hard to get in at a growth stage if you don't quite have the reputation. When you thought about building this, not just, you know, creating great funds, but also sort of building a firm. What are the other things that you've thought about? Like, have you thought about, you know, we talked a little bit about team, the stage and focus and things like that. [09:43] sizing of the funds over time? Like what are some of the other attributes when you think about building like a new real venture capital firm? I still believe brand is super important. Brand is probably... [09:55] Even more important, when your purchaser, i.e. the entrepreneur, is not incredibly well informed. It's like, you know, what is brand? Brand is basically a way of summarizing the value proposition that a product has to the prospective buyer. Because the prospective buyer is an expert. You know, you buy a Mercedes-Benz, why? I don't know anything about the engine or the reliability or whatever. You just know it's good. You just know it's good, right? That's what brand is.

[10:25] entrepreneurs being younger, [10:28] and having less network in the world of capital, brand is very important. So I do think you have to be super conscious of brand. [10:38] However, the creation of that brand has to be done very differently, which is to say, in the old world of brand, [10:47] You know, you could use marketing good stuff like, you know, you put out super professional content and banners and you sponsor this conference and so forth. I think in this modern era, there's an enormous amount of transparency. And so brand gets conveyed in a much more organic way, person by person, reference by reference. [11:17] interact with, which spreads and then people get to know you. Yes. Right. So I think the kind of classical marketing efforts that have been done for brand building, particularly in venture, sound off key to your average 22 year old entrepreneur. They're just it doesn't resonate. Yeah. [11:35] I think what does resonate is inside knowledge, tips, connections, network, all those things which are kind of indirect but organic ways of building brand. So, yes, brand is important, but I think you want to build it organically. That's not the right thing. I heard somebody did like a survey of college students and what BC firms. And interesting on this point, Thrive was ranked really highly. And they don't do a lot of, you know, it's not loud. There's not tons of blog posts. There's not all the other stuff.

[12:05] Josh and his team over there are probably the prototypical built a brand on the down low versus the, you know. Totally. There's also when you feel like you're in on a secret, you know, when you when you know about it. And if you know, you know, brand that's worth a lot more to people. A hundred percent. But how does that get cultivated? I guess that's one of the questions. You know, like, you know, over time success, you know, Hermes doesn't exactly do big brand ads. Yeah. You just know it. I also think I'm sure you've probably felt this just over the course of your career, but. [12:33] probably over the long term, making good investments is probably the number one thing. And, you know, being able to sit with founder and say, I've invested in these companies. I think that helps. Look, I think at least from my own experience, what I find is like having made some good investments, i.e. I know what I'm doing is a good one. But also just people saying, look, so and so was very, very helpful in these particular moments in my company's evolution. When this happened, they were helpful in this way. When this happened, they were helping. You know, you're on the board of somebody and they say, who's your best board member? [13:03] that person. And the reason it's, I think, particularly tricky in VC is because people are always like, okay, tell me about what your value proposition is as a venture capitalist. And you're like, well, I do recruiting and we do marketing and we introduce you to customers. In that voice. Yeah. You know, this kind of stuff. And that's what every single venture firm does. But the real... [13:22] The real in my experience, the real thing that you do with entrepreneurs is whenever they get stuck. Yes, you have an answer for them and they get, you know, entrepreneurs, you've been one. You get stuck in all sorts of ways. I don't get along with my co-founder anymore. What should I do? I've got the series B coming up and should I do it as a two stage series? Be like all this. And you can't sell this, though. It's like one of those things that just has to come through references. I think exactly. Like you can't be like, absolutely. Oh, I'm there for you when it's hard. It's like, OK, that reference.

[13:52] investment success becomes brand. And then it's highly unassailable because you can't throw banners at that. How do you think about leading in board seats then in contrast to participating and, you know, maybe having like a slightly less engaged relationship? Do you do both? Nabi is a small fund. [14:08] And, you know, we manage one hundred and seventy five million dollars. So leading a series A in this day and age, which probably ranges between 15 on the low end to 2530. [14:20] Not broadly generalizing, very hard to do with my size fund. So for now, we generally lead seeds, co-lead a few series A's, deploy some capital against big companies that we think are going to be. [14:32] even more successful. I do think that at some point in the future, maybe with a little more capital, we'll try to lead some things. But [14:42] I also think that the idea of classic lead may not be as relevant in the future, meaning, you know, again, this is classic VC rule. I want to own 15 percent of the company or 20 percent of the company at Series A. OK, that's that's nice. First, it's probably incongruent with what the entrepreneur wants. But then again, would you own 20 percent of some schmo company or one percent of Anthropic? So I think there, too, you sort of need to relinquish this need to like lead, not lead, whatever. [15:12] deploy capital into good companies if you can deploy it as a, you know, 4% owner or an 8% owner, just deploy the capital. And then these companies create such disproportionate value over time, just pour more money in, more money, more money, more money, even at higher valuations, just put more of it in because they're great companies. And so I think that's one of the kind of historical VC rules that we have to let go a little bit of like, I must own this much in

[15:42] And then buy more over time. Now, what does that mean in terms of board membership, not board membership? It's a badge of honor for a VC to say, I'm on the board of that company. [15:51] In my case, you know, I've done this for a while. That's not I don't need that badge of honor anymore, but I will commit to spending a lot of time. And so basically almost every investment that I've made as part of Hanabi, I will do weekly, biweekly or at least monthly check ins with founders, go on a walk with them, spend an hour with them. In some cases for the younger companies is literally like weekly. A, I find that a lot higher bandwidth in board meetings. [16:21] times at this point and I don't find them as productive for knowledge extraction. And I think it helps the entrepreneur more with what matters to them in the moment. A lot of it is managing their exacts and things like that a lot of the times. Yeah, there's so much stuff. There's all sorts of stuff that's not ... Everybody's got, like, you know, when you're a founder, like, "Okay, this week I got this issue. Next week I got a different issue. Next week I got..." So, yeah, once a quarter you prepare a deck and you sit around and read the deck and all the three guys that are observers want to say something useless. I agree. Once they're big, they're useless, I think. A small board meeting is a good for you. [16:51] I completely agree. I have done some board seats. I don't make it in any way mandatory, but I do make it mandatory to invest the regular time one-on-one or one-on-two, depending on how many founders they have. When you think about going from… [17:07] you know, a good early fund, couple funds. What do you think will,

[17:12] allow for a new firm to break through into the sort of like multi-decade this is now a truly established [17:20] firm? You know, like, do you think that'll be a function of the brand ascending to a certain place? Is it a function of getting to a certain size? Like, what do you think is the thing that leads to, all right, this is now one of, you know, the institutional firms? When you take this job of a venture capitalist and really distill it down to its [17:38] It's barest bones attributes. I think there's four things that matter. You need to find opportunities. So discovery sourcing is important. You need to have good judgment to decide whether something is good or not. You need to convince the entrepreneur to take your money. So you've got to be a salesman. And then lastly, you've got to help them develop their business so there'll be a reference so that the other things work. Those four things matter. [18:08] become where you have like 600 people at a firm to do spreadsheets and by the way nobody's going to do spreadsheets anymore that's true too that's over i think you can have actually very compact organizations [18:20] of people that do those things very well. Doing those things very well take actually a fairly diverse skill set. So once you start [18:30] trying to over-specialize like, you know, this VC is a good sourcer, but they have bad judgment. This one's a great salesperson, but they can't source. You know, this one's a great operating partner. You're trying to stitch together like five people to do one. No. So I would say the firms of the future are probably leaner and more concentrated to groups of people that are well rounded at that suite of four skill sets. On that point, the way you described it, you know, was the loop of the strong reference from being a great board member loops back. You can't then

[19:00] next deal. Like it almost has to be one person. No, I agree. Venture capital has gone from this sort of boutique business to the scale business of like I administer $20 billion. So if you want to call that venture. Administer is a funny way to say it. If you want to call that venture, okay, then you have all sorts. But in terms of actually helping companies grow and develop and in true [19:30] demonstrated by as many people in the firm as possible. I suppose also there's going to have to be something around the sort of like the generational transition ability of the firm has got to be like the thing that makes it last beyond the founders. And yeah, but I mean, you [19:43] trickiest thing at every single venture firm. This fact doesn't change, which is the loop in VC is very long, right? So you're a brilliant investor at Benchmark. You're making investments today. Benchmark will raise a fund the next fund you do, not because of the work you did, but your predecessors. So is the credit due to them or is the credit due to you? And this is, well, I don't know, because you're going to generate value for the capital that you're raising now. So I think [20:13] on the economics in a firm. [20:16] Things go poorly. That is the quintessential failure mode of every venture firm, which is legacy partners who no longer do useful investments take the disproportionate portion of the promotion. It seems to me it's almost like it was clearly the last generation's work that earned the next fund. But you have to give it all away anyway if you want the thing to live on. Yeah, you get a bet on the future. And you have to just say, I made enough and now something was handed to me, I'll hand it to the next people kind of thing.

[20:46] Obviously, I know you're fully AI-pilled, the whole firm's around AI. So I want to talk about a few of the ideas here. What I want to start with is just sort of level setting on your overall... [20:56] I guess, assessment of the state of play of AI. And maybe let's start with kind of like, [21:01] what you think is happening at the labs at the sort of like core intelligence and you know i want to talk about sort of where you think the you know advantages and edges are what you think around sort of like compute and capital and scale of these things like what's your read of sort of like you know the state of ai at this sort of central lab level first the first thing i would say is my mental model of ai is a stack of things that start at the bottom with uh foundries and [21:31] models and the infrastructure that's associated with building those. And then there's sort of, you know, middleware and application layer software that sit all stacked on top of each other. I think you were referring to the big labs themselves, which sort of sit right in the middle of that pie. My perspective right now is that the winners are the winners and we already know. [21:51] I think, you know, clearly OpenAI, Anthropic, Google, I would throw meta in the mix right now. [22:01] - Yeah, those are the five. - Those are the five. Why those five? And that's because if you think about what are the vectors, what are the forces that make one model company [22:12] better than the other one. Right now, compute and accessibility availability of compute and chips dominates. Compute is very highly correlated with capital. And you're talking about a group of companies that spend a lot of money.

[22:26] 50 to 100 billion, conservatively, a year on compute. Now you show up and you're a new company X and you're like, hey, you know, I raised 2 billion. And it's like, yeah, nicely done. You're about an order of magnitude. [22:40] if not more, at a disadvantage than your competitor. So I [22:44] Don't see a scenario. [22:46] until model architecture has changed and they may change it's the bitter lesson to me right now even if you came up with a far better model and you could argue that the way open ai and anthropic do their things are let are a little more clunky they can just throw compute at it and blow you away i don't see that landscape changing they are the dominant force in llms now there are other kinds of models that could be kind of interesting but at least in that world game over okay so two [23:16] about then open source. And so like, obviously, in the wake behind the frontier, you have open source, you know, let's call it six, whatever, nine months behind. Obviously, smaller models, cheaper, you know, served by, you know, inference companies. And there's obviously a lot of utility there. Is that going to be an important part in your view of the next few years? Do you think that that's going to sort of be at the edges no matter what? Open source is a [23:40] is a relevant phenomenon, but not a business in AI. [23:45] If you look at the overwhelming amount of prompting that's happening, it's against the most advanced models. [23:53] And that's the most monetizable. [23:55] So advanced models being queried over and over again is where the dollars are, and that will sustain.

[24:03] Open source can exist in my view in the context of somewhat far behind and somewhat reliant [24:12] on various forms of distillations and techniques that the big guys have already done. So what that tends to do is if you think about the monetization curve, it commoditizes the tail end of the curve, but not the front end of the curve. And the big dollars are still in the front end of the curve. So I think there will be. The asterisk there is that if you're creating an open source model that's very close to the front end, you're by definition using a lot of compute [24:41] You need to have money to get that compute. And you're seeing even the most staunch open source firms not be so open source anymore. The latest Quinn model is closed. MuseSpark is closed. That's not a coincidence. They're spending a lot of money training those models. They're not just going to like open them up. Do you think the frontier labs will stop allowing third party to use their frontier quite as much because of the compute crunch stuff? It is likely that that will be the case. [25:11] if the big labs start to say, hey, you can access our old models, but the new stuff is only first party? Well, I mean, what it'll do is it'll cause the open source stuff to be even further behind. And by the way, temporally, you can have blips. So like Jensen can get the idea that he really wants an open source model. So you can go to Reflection or whatever and give them $10 billion. One generation, they can catch up and then they'll fall behind. But it's a one-time thing. It's a one-time thing. I don't think it's continuous. I do think that you'll have, I mean, open source will be relevant.

[25:41] It's relevant in one more context, which is at least in today's world, and I don't know if this will sustain, but in today's world, if you post-strain a lesser model enough, it will perform a small number of tasks just as well as the big models can. [25:58] Right. Essentially, open source can exist as the underlying layer of companies that are post training them for specific. [26:06] personalized or business or corporate use. That exists. And there may be a business model somewhere out there. If these [26:13] Post-training kinds of companies are quite successful that they can provide financing for open source to be closer in to the big models. But open source out of the kindness of your heart, which was the Quinn, the Deep Sea, Kimi, Lama. Doesn't make sense. No. I mean, who's going to go spend $50 billion to give it away? I mean, this doesn't make sense. It made sense for a minute? Or what do you think was the argument for it at the time? I don't think people forecasted how much money it would take to train these bigger models. Yeah. [26:41] I don't think people expected it. I think, you know, decisions like why was Lama free? My guess is the folks at Meta looked at all these people and said, we can't exactly monetize this, but we'll try to commoditize their market by putting out something that's free. And I think that that's. [26:56] past. [26:58] That's past, clearly. Yeah. I mean, it seems like once we saw that it was tapping into labor budgets for real instead of like software spend at that point, it's not really about price. It's about value. Yeah, exactly. That's the open source side. And then what about the Neolab side? Because that's like obviously the other potential vector here. And I would argue that's probably like one of, if not the hottest venture segments right now that is getting backed at, you know, huge dollar amounts, huge prices. Like what's your read on why that's happening?

[27:28] thesis that the big lives are going to win. It is difficult to imagine that there will be [27:33] an algorithmic improvement. [27:36] that will allow a new lab to be competitive with the capital that the big labs are throwing at it. Not to mention the fact that the big labs, they're not sitting there clueless, like thinking like, oh, yeah, there's no other algorithms here. They're doing their own research of what the next generation will be. I don't see a scenario where there's a more clever architecture to a model that outperforms dramatically the big guys. Now, there are a couple of areas that I find... [28:03] Somewhat interesting. And those usually sit in pockets where there is a set of proprietary data that the model can be trained on that is not available to the big labs. [28:14] So, [28:15] If you think about the core of the training that happens for the large models, it's the internet. The internet and the internet is largely available to everybody. If there's pockets of the internet that are not available, you can pay for it and open it up and get that stuff. Any model that sort of bases its training on internet available data is likely to not be as competitive with the big labs. However, there are pockets where you need or have proprietary data. [28:45] Why is that interesting to us? They create their own data through their own labs and they do training and reinforcement based off of that. We are interested in robotics. Why? Well, robotics data is not broadly available on the Internet. However, you can either tele-op it or have remote workers create data, but you have to basically generate the data that your model trains on.

[29:15] strategy can actually stand out. So there are pockets where I think different kinds of models can be used for different kinds of orthogonal directions than the classic core models. But the Neolab that comes along and says, you know what, I'm really good at RL, right? And, you know, because we do RL better, we're going to, [29:33] beat the other guys. And I'm like, you know, I got the memo on RL. They didn't miss that memo. Spend a lot of money on RL. On the robotics example, let's say so, you know, just to run with that, like data is like the thing that could differentiate it. How do you then think about investing in, let's say, scale for robotics data? That's going to be a data company selling to the big labs versus I'm going to invest in a new lab that is going to get that data. [29:58] Or do you think both are workable? As a longtime investor in scale, what I would say is the business of providing... [30:08] labeled data to large labs is a tough business to be in. At scale, we were there early and we served the labs well, but every single time a new contract will come up, it's a dogfight. It's scale, it's surge, it's Mercore, it's Handshake, it's Turing, whoever it is. Because ultimately, the job of annotating that data isn't that hard. It's a low barrier to entry job. And what the labs want is that data provided to you at low cost. I think essentially the same will be true for [30:38] that is provided to the labs. It's going to be very competitive, low barrier to entry, and it's going to be largely dependent on who bids lowest. So then what would you need to believe to back a new robotics model? One is the architecture of the post-pre-training is important to understand in robotics because generally, like all other things, if you do heavy doses of post-training in robotics, the robot will do one task really well, but that doesn't generalize. In order to do have robots that

[31:08] generalize, you need a lot of pre-training data and that pre-training data is very important on the specific embodiment of the robot. So like what physical shape that robot takes is very important. So companies that generate a lot of their own pre-training data, often created in-house, [31:26] is actually a material differentiator, in my view. Purchase data is nice, but not a material differentiator. So if your robotics company says, I get all my data from scale, I get all my, not that you should buy some of that data, please do, because I'm still on the board at scale. But I don't think that's the differentiator. Differentiator is how and what kind of data do you collect in-house? And you don't share with anybody else. And you created largely, I think, around embodiments, which are explicit to your robot, your architecture, particularly [31:56] Since the hardest tasks in robotics right now are largely manipulation tasks or dexterity tasks, like what kind of gripper do you have? Is it three fingers, four fingers? Is it a little claw? Those things matter a lot in the pre-training. Do you think a company that makes robots [32:11] would be a good source of that data for themselves? Like, is that a good way to back in? Like, if somebody's building a robot, they have them out in the wild, [32:18] They're presumably collecting a lot of potentially proprietary data. Like, is that one of the vectors? Absolutely. Yeah. I mean, one of the advantages I'm pretty convinced that Elon will have with Optimus is that he has data collection mechanisms all over. I mean, look at self-driving. It's just like once you're in the wild, you start getting the flywheel going. Absolutely. And, you know, we're investors in a company called Mind Robotics, which is the robotics spin out of Rivian. Same thing.

[32:48] differentiated than other people. What's your read on the overall compute situation? So like one framing of this is that basically like the demand for AI is just going absolutely vertical much faster than can compute can get online. What do you think is like the likely ways this plays out? You know, if like the labs, you know, Anthropic adds 10 billion of ARR, you know, a few more months, it's just like, how do you think this ends up shaping, you know, maybe down through the stack? You know, we were talking about the labs, but like, well, what will the ripple effects be? And like, how do you think about this dynamic if demand really is surging at this [33:18] It used to be that the demand for GPU compute was largely on training. Now, with this explosion in inference, you have the twin effect of bigger models being trained and tons and tons of inference running on the same systems. So that means that the load or the demand for compute is kind of skyrocketing, and it's likely that it will continue for some time. The supply side of that is basically TSMC wafer starts. [33:48] There's only so many of those. So anybody that bought [33:52] compute historically, [33:54] probably got it cheaper. [33:56] that anybody that will buy it in the future. And I think one of the very smart things that OpenDII did is they bought a lot, a lot, a lot of compute ahead of time. As a result, I think, [34:06] Google aside, because we talk about TPUs in a second, but Google aside, they probably have the lowest cost of [34:12] compute of any player in the market right now. Yeah, there's the price of it. And there's also just the availability of it. It's just like, you know, money doesn't get you more wafers kind of thing. You just got to wait. Yeah, exactly. On the other hand, the value, particularly on the inference side, that the value that people are seeing from the inferencing of these models is very high. Like, you know, you sit there and you do a spreadsheet and you ask Claude to do the spreadsheet for you and you go, wait a minute, I'll pay triple the amount of money that I'm paying

[34:42] You'll be able to sell the output of that compute at a much higher price point in the future, given how productive and useful it is. Today, we live in a world where NVIDIA is the gatekeeper to all compute. [34:52] Again, TPUs aside, because that's a whole interesting side conversation of what Google decides to do with TPUs. I don't see that sustaining. And I do see more specialization in the kind of compute. You don't think NVIDIA's dominance will sustain? No. What do you think will happen? I think you're going to have compute that is orient towards specialized tasks that are not necessarily dependent on NVIDIA. So, for example, I'm a big fan of Cerebris. [35:18] Why do I like that one? It's not a particularly good training chip. [35:22] will argue with me that that's not true. That's not true. But, you know, let's leave that as where it is. In inference, it's a godsend. It's super fast and it's made certain compromises that make it very, very good for that. I think that's one. I think we're going to see more of that style of architecture that's very good for inference, for example. I think you're going to see more of these. There's even... [35:43] further evolutions of, you know, back in the old days, we used to call them ASICs. But if you look at Etched or Talos, these are companies that are fundamentally baking in some parts of the neural network weights into the chip itself, which will make it even faster than Cerebris, even less flexible than Cerebris. So I think that you're seeing this evolution towards mission-specific silicon rather than general-purpose silicon. And that will happen and will take the load off of

[36:13] think that we're going to do a lot of work in that bottom layer of the infrastructure to liberate ourselves as an industry from a stranglehold of chips. Yeah, I feel like the U.S. probably can't invest enough fast enough in this. Already a bunch of the GPUs are being built in Arizona. I think some of the Cerebris chips are made in Arizona. I know Intel is trying to come online with their fabs to do independent silicon. I don't know that that's still worked yet, but I do think we're [36:43] the geopolitical issue of Taiwan. And, you know, I can't imagine that that same situation will be true in five, six years. That would be great. Yeah. [36:51] What about now flipping to the other side? You know, we've been at the bottom of the stack, like the application layer. Obviously, over the last few years, there's been a bunch of amazing application companies built. Do you feel... [37:01] Like there's, [37:02] durability there and if so what do you think the source of it is or you know what are the places that you think are risky not talking about pre-assess i'm talking about like an ai native you know seeming application i guess let's start there in a given business there are two things that are fundamentally proprietary [37:17] One is we talked about data. Every business has a bunch of data. [37:21] And it's their most relevant data. And the second one is business workflows. So how you do and conduct business. I think... [37:29] Companies that can capture... [37:33] the data and the workflows that exist inside of a business or in a given industry vertical that the big labs are unconcerned about or don't have access to those things can survive as application companies. If your basic concept, though, is, you know, I take a document, I send it over to OpenAI API or Anthropic API, it comes back and I show it to

[38:03] is over time, the OpenAI, Anthropics, and Googles of the world will look at the most interesting TAMs and saying, wait a minute, I'm going to do that. So there's Anthropic for legal or Anthropic for finance or whatever. This is not their core business. So they're going to have varying degrees of success. But it'll do 80% of the job. But it'll do the damage. Business verticals, if you're in construction, there's a bunch of workflows in construction that it's not likely [38:33] I personally, I don't know, maybe this is something I need to let go of, but it would be surprising to me if people no longer wanted UI's. [38:40] But we'll see if everything just became chatting. No, I'm in that camp. You know, it was interesting when you were talking to Brett, he was talking about like systems of record and how systems of record are important. I actually think one of the reasons why Salesforce sustains is because every salesperson knows how to use Salesforce. So actually, the human interaction with the system is the single highest moat that anybody has from breaking in. That might change over when agents do the work versus humans. [39:10] as I am, I think that takes a while. Yeah. And so that that retains some of the UI does retain some level of sustainability for some period of time. What do you think about AI like services and software fully do the work type of things where, you know, instead of selling you, you know, accounting software to an accounting firm, we're just going to sell you completed accounting work kind of thing. Like, do you think that's where stuff ends up? Are you more excited about that? Are you more excited about the software? Yeah.

[39:36] looking companies? Well, I answer the question two ways. One, I think it's a high bar today for AI systems today to actually offer accounting as a service. You still you may not need the entry level, lower level, white collar work, but you still in today's world need a lot of contextual and other forms of knowledge that accountants have about their clients, about the systems, about how all things work. So I think we're still in an era where [40:05] We're purely delivering a sort of service from AI with no humans involved. I think we're a little early for that. Particularly because in some of these, you need the verification of a human stamp. It is actually part of the value. You know, I think customers obviously will want some human to have supervision over it. And then humans... [40:23] We have some very interesting sporadic contextual knowledge about things that's actually surprisingly helpful to the point earlier about VCs and what makes us different than introducing you to customers and employees. There's just like random little bits of knowledge here and there that we don't know what it's going to be. I think that exists and persists for a while. [40:42] I will say, though, that SaaS business models, like classic product models, are going to go through a very important transformation, which is when you spent a lot of money developing a piece of software and then sold it to many, many people, you architect your business to be a product business. And VCs for years have looked at things and said, like, oh, I want a product business, not a service business. But now, if the cost of software goes way down, I think you go in one of two directions. Either you're a low-cost provider. You're like the Amazon of software.

[41:12] margin available to everybody's, you know, good delivery and so forth. Or what you offer is a more customized edition of the software that either has agents associated with it or is really plugged in and integrated with your system in a particular way. I think what that has done is sort of recast, you know, 10 years ago when I was doing VC and we said like, oh, they have a professional services business model. You're like, oh, no, no, no. Pass, pass. Not that one. Now they're like called FTEs. [41:42] Cool. And the interesting thing of, you know, what I what do I think of an FDEE? [41:46] is it is a job that is a go-between between a business problem and a technical problem. That's the fundamental value of an FDE. Now, your business problem tends to be pretty specific to your business. Your technical problem is something that's built on a stack and structured and so forth. And somebody needs to say, Jack, I understand your business problem. This is the way how we solve it. And that's the software company of the future. In some cases, you will solve it by populating a framework or an infrastructure with agents. In some cases, [42:16] - Mm-hmm. [42:18] But that's what a software company of the future, I think, looks like. And one of the interesting things, just from a business model perspective there, is these deal sizes are so gargantuan that an FDE just doesn't matter. [42:30] for the for the cogs relative to the compute costs and other things. If you're doing a 10 million dollar enterprise deployment, who cares? 100 percent. But, you know, keep in mind, like I started my career peddling routers and switches. Right. And that's when you used to go to the network engineer on the other side. It goes like, well, my switch has 24 ports while you're his only has 16 ports. And therefore, and like for that difference, you could charge one hundred thousand dollars a year. But when you're talking about the CEO of T-Mobile saying

[43:00] 2% and you do that for them, that's hundreds of billions. That's right. Right. And no surprise, the contract values that these people are willing to offer are huge. Why? Because you're actually solving a business problem for them. You're not providing them with the technology. And that's why the FDE stuff makes so much sense. So that's a bit about like the sort of and native stuff. Just quickly, I'm curious your sort of pulse on the pre-AI SaaS. [43:30] year or so and every time the labs release some new thing you know it's like the stocks get traded down further do you think it's like oversold do you think it's undersold [43:38] How do you feel about it? You know, I'm not a very good public investor, so it's hard for me to say if things are over or undersold. But I do think all of those companies have like one of two paths forward, right? Which is one is hang on to what you've got. Grit your teeth. Fire more people. Just get out of the way. Increase your EPS and just, you know. [43:56] Make the most profit available for the duration. Fire half your staff and increase the EPS. Private equity, essentially, to yourself. But there's going to be a set of companies that will be able to embrace the AI thing and do... [44:11] kind of a transition of their business into more of an AI-centric business model. I don't think the market has really differentiated that because if you look at the multiples, everybody's sort of sitting at five times, four times. And also everyone's saying that. Yeah. And I mean, look, I think there's a big difference between... [44:27] I don't know, Figma and Workday. Yeah. Right. I think the chances that Dylan's going to figure something out are much higher, a lot higher. And I think his well, I'm an interested stakeholder, obviously, but I do think that some of those stocks have been oversold. Now, I don't I can't tell you if they should be trading at a 12 times multiple or a six, but it just feels a bit oversold right now.

[44:50] Given the capability of the founders, the beautiful thing about founders and leaders is that however the business looks today, it doesn't have to look like that. [45:02] in five years. Yeah. Right. If Tesla traded at car company multiples, it would look different. It would look very different. The reason it doesn't do that is because people look at Elon, even though they suck to be the same, of course, even though there are no cars that are well, I mean, there's a handful of cars that are self-driving and there's no robots in the market. People go like, [45:19] He will successfully transition the company at that stage. Therefore, the company is worth more. I think SaaS is like the car business right now. Everybody thinks the car business sucks. Everybody thinks SaaS sucks. Some of those leaders will be able to transform their companies by hook or by crook into something like what Elon did. And I think you have to look into the management of that company and say, do they get it? [45:43] And then I think you'll see differentiation of the good ones and the bad ones in the SaaS universe. I completely agree. And then how about outside of AI? Is there anything that you're... [45:52] investing in that's not AI? Right now, it's like 90% plus AI, but I would say defense tech is an interesting area. In defense tech, too, there's a lot of AI into it. But I do think that we're seeing geopolitical shifts that are [46:10] are [46:12] reasonably permanent for some time. Take, for example, the dynamic between the U.S. and Europe and defense spending. I don't see European defense spending being reduced. It's probably going to be dramatically increased. And so there's opportunities there. It's the right tailwinds. I do think that with companies like Anduril and to a degree Palantir being successful and breaking into

[46:42] and inspirations for both sides. I was going to say it does both things. It both, it unlocks capital and talent, but then it also sort of warms up the DoD to believing that, you know, these companies can work with them and take their understanding seriously and all that. Absolutely. And I don't think that's going back, you know, this sort of like the, you know, [47:00] Pandora's box has been open on that front. So I do think that that's an interesting sector. You know, it's interesting because there's always these philosophies like, oh, do you invest in weapons and not invest in weapons? And when I was at Index, uh, [47:11] We didn't do defense investing at the time. And I remember looking at Anduril and thinking, this is going to be a winner. And I had a conversation with Trey Stevens. And I was like, hey, I don't think we can do this because we don't do weapons. And he was like, I don't think you get the mix here. Weapons are a system that deters violence because the more weapons one side has, the other side have it, the less likely as they are to go to war. Now, this president has disproven that. But leaving that aside, I think the logic was pretty strong. [47:41] And so now I'm more of the belief that I think the right kind of defense investing is the right thing to do. Yeah, I think so, too. Before you joined Index, you were like a founder, you're an operator. Do you feel like that mattered looking back? Like, was it valuable to you to have done that? Do you think would you have been an even greater investor had you just been an investor the whole time?

[48:11] than just more experience would have. First of all, I really enjoyed being an operator. [48:16] It was super fun to be at a fast growth company, to see things explode and grow. You know, I joined Cisco. We were. [48:22] A few hundred people. I left. We were 55,000. Oh, my God. You know, like. In how long? I was there for 13 years. Wow, that's crazy. It grew to that size in about 10. Wow. And by the way, you feel somewhat like you participated in something that's reasonably momentous. Like, you know, I got there when nobody had the Internet. And because of the products that we made, everybody got the Internet. Wow. I mean, that's only you. You count the companies on one hand that have had that headcount growth. Yeah, yeah, yeah. Not just the headcount growth. Of course. [48:52] impact of what was created. Yeah. Just because of the pure enjoyment of having been there, I'm happy I did it. How does that how is that relevant in the context I do the rest of my life? I do think that the old Steve job things applies, which is you've got to connect the dots looking back now that I understand VC better. And by the way, I still get I still learn every day, but I understand it a little bit better. I didn't actually realize how important having that card [49:22] was because there's a lot of VCs, [49:26] Everyone's smart. [49:27] Everyone's got a degree from a good school. Everyone's got a big fund. [49:31] You've got to differentiate yourself, right? Like when an entrepreneur is selecting a VC, I am the product. [49:36] What makes this product different than any other product in the market? And in my case, having had that experience of growth, recruiting, customers, relationships, all that was enormously useful for me. I'm not of the camp that you have to have been an operator because, you know, you look at Peter Fenton, the biggest thing he's operated was a lemonade stand. Well, shout out, Peter. Yeah, this is for you. But he is one of the greatest VCs of all time. He is. And so...

[50:05] We all, as a venture capitalist- And there's many other examples like that, too. Absolutely. I mean, Michael Moritz was a journalist. Yeah. John Doerr was a sales guy. So we all come at it from different angles. I think the important thing is to recognize [50:17] that you have to have an angle, and that that angle has to be relevant to the entrepreneur. And so for me, having been an operator was mega, but I don't view it as like it's a must. It's an approach that happened to help me. And most importantly, it's fun while you do it. Yeah, absolutely. When you think about what types of founders are thriving today, [50:39] Has it changed versus before this AI thing? Are there different mindsets, different types of educational backgrounds, different work experience, different age? Have there been updates that are material to you that change what kind of founders you're looking for? [50:52] interested in? I think the biggest difference is the level of maturity that some of the younger founders have relative to even 15 years ago. There is so much content available, so much [51:05] peer community available, that you take the average [51:09] 20 year old, 21 year old, and they're showing up with so much knowledge compared to a generation ago. I know. I was 21 15 years ago. And when I meet a 21 year old today, who's like a founder of, I'm just like, oh my God. Like I wasn't there when I was 27. Dude, I was 21 in the stone ages. So yeah, there you go. But I'm just like, it's crazy what people know now. And I think it's going, I think it's accelerating because I think, you know, now a 12 year old today is going to be using AI tools. By the time they're 20, they're going to be eight years into this. And it's not

[51:39] like book knowledge or AI knowledge, there's also a lot of just commercial knowledge. Yeah. What do you think that's about? Is that from the online content? Where is that from? Yeah. I mean, it's all of the above. You know, we're absorbing content from all sorts of sources. It's peer groups, it's online, it's stories that we tell, podcasts that we listen to. Here we are. Especially yours. That's right. But I think people accumulate knowledge from all these sources at a very young age. It's incredible. It's extraordinary. I mean, it's absolutely extraordinary. Can I ask you an [52:09] You're not old, but you're not a young VC, exactly. And you're doing it very successfully. What is the mindset? What do you have to do to be not a 25-year-old VC? First of all, thank you for calling me old. You're not old. If you were, let's pretend. Look, I think the... [52:26] I think the challenge [52:28] that [52:29] we old dudes run into, especially when we've had some amount of success, is that we think we know better. You've got to approach these things with a beginner's mind because every 21, 22-year-old knows a whole lot of stuff that I don't know about. If you approach the relationship with the like, hey, I'm the all-knowing grandfather. Let me tell you, kid, how this is. It's not going to work. [52:54] It's not going to work. You just have a conversation with people and you say, that's interesting. Let me give you my first principles answer to what you're saying. I think this is the right approach. But if you don't agree, tell me why. And let's have a conversation as equals. So I would say, forget your age, treat the person like an equal. And sometimes you might have an experience that you can pull out and say, well, you know, five years ago, I saw this. It might be wrong today, but I'm just going to throw it out there for you to consider. And if they like it, great. And if

[53:24] biggest thing. And this is so hard for people to do because I mean, one of the things I think makes that particularly hard by the time somebody is, you know, been working for 30 plus years, they've had successes. You don't want to relearn everything. You know, I think it takes a certain mindset to do it. But when you look at some of like the best tech executives, you look at a bunch of the best investors. I mean, I look at like, you know, take Jeff Bezos, take the node. You take a lot of people. I do think that if people are able to reset in beginner's mind all [53:54] extremely experienced and, you know, are able to think from the beginning again. Totally. I mean, not to get philosophical about it, but we all start as younger people. Maybe you didn't because you're a very confident man. But I started with a lot of lack of confidence, like, you know, can I do this? Will I be able to do this? And then over time, you sort of like, oh, I figured this out. I figured that out. And then you get to be 59 like I am. And you're like, hey, I want to feel confident on this foundation. I thought you were 49, by the way. I didn't [54:24] But, you know, you get on this foundation of self-confidence, which is built on the fact that you think, you know, at that point, being able to tear it all down. [54:33] and say, actually, I don't know. [54:35] It's a very hard thing to do. It breaks the entire structure of what gives you comfort to be the person that you are today. Do you feel like you've got that successfully broken down? [54:45] Not always, but I try hard. I would imagine it feels good. [54:49] Like I would imagine it feels kind of freeing. It doesn't always feel good to people. It doesn't feel good to feel.

[54:55] Dumb. New. New. Yeah. It's a it's very naked, right? It's a very uncomfortable feeling. But it's permission to like explore and have it all be light again. I agree with you if you can get there. Yeah. But I think for a lot of people, that's a lot of breaking down of layers and layers. It's there's just stuff that's been built up. I do think it's the ultimate expression of self-confidence is when you can accept the fact that you don't know something. That's what I think, too. I think being able to like raise your hand in a conversation with a group of people say, what was that word you just said? Yeah. Super confident thing to be able to do that. People don't want to. [55:25] That's a tough thing for people to go through. Does it change for you the types of people that you want to [55:30] build your firm with at Hanabi when you think about [55:33] the sort of dynamics of the market? Does it make you want a wide range of [55:38] teammates' experiences and ages and backgrounds and all of that? Or do you think it's more about this mindset at the core? I think this mindset is absolutely the core forever. But I do think that a modern day fund is structured with a lot of people that are cohort relevant with the current entrepreneurs. I am not. [55:56] And I try my hardest to be cohort relevant through the things that we talked about. It's much easier to teach somebody of the current cohort a little bit about VC rather than teach an old dog how to how a 25 year old thinks. Well, Mike, this was a total pleasure. I really appreciate you doing this with me. I learned it done. Thank you. Super fun. Thank you. Have me back anytime. Great.