0:00This week [music] tech legend Keith Rabois joins us on Verticals to break down where vertical AI is actually headed and who's going to win. According to Keith, the biggest question in AI is not how fast you grow, it's whether you can build a company that lasts or on the other hand a product that the frontier labs swallow. Keith, per usual, does not hold back. We covered a lot of ground, a lot of verticals, legal, accounting, investment banking, and they raised on
0:27entire industries. Plus Keith explains why he's completely changed his mind on enterprise investing and why he's pretty bearish on AI roll-ups in today's vertical software incumbents. If you're building in this space, this episode will challenge you on how you think about your company. Let's get into it with Keith Rabois. All right, welcome back to another episode of Verticals.
0:49Excited to be here today with myself and Nick as always. We've got an incredible guest. We've got Keith Rabois, managing director at Khosla. Needs no introduction, but thanks for being here, Keith. Appreciate it, man.
1:00Thanks for the invite to join you.
1:02Yeah, we want to pick Keith's brain on all things vertical AI, all things application layer today. And so I wanted to start with a quick quote that we heard from you and and just get your take or your breakdown of it. But you were on a legal specific podcast and you know, you talked about application layer companies and vertical application layer companies and how most of them you don't think kind of can overcome this AI moment, the open AI-ification of all of softwares. So if you can expand on that,
1:31I'd love to get your take on like what is a good vertical application look like? What is one that's not defensible?
1:36Maybe we start there.
1:38Yeah, no, I think there's lots of great opportunities in vertical AI to be specific. I've made it several investments in vertical AI companies. We can talk about some of those if you'd like. Our fund has made as well. I think the ones that move more towards autopilot uh are probably more defensible and more durable long-term than those are kind of co-piloting. I so I think that's this one you got to of framework. A second framework is also a channel Sam Altman a bit and basically say that if your
2:06vertical AI company is betting on models improving over the foreseeable future, that might be a good investment. If your vertical AI startup is betting or investment is betting on the models are going to flatline, you know, and stop improving and then you're going to build on top, I don't think that's a great place to be. So, I'd say Autopilot versus Copilot and are you embracing what frontier models can do and
2:33channeling that into an either vertical application where the customer gets increasing value, that I think can be very successful. So, we're pretty disciplined about when we pull the trigger and invest in a vertical AI company, but we find lots of great opportunities and there's wonderful founders uh building in vertical AI companies right now.
2:55Got it. Awesome. And if if we can expand a little bit on it cuz I'm just I'm curious.
3:00Any specific like you guys I think you did Spellbook, right, Keith? How do you view that versus like a Harvey and and which one how does that size up in your framework here?
3:09Well, we'll see it ultimately. I I you know, I think Spellbook you can see there's probably more of a crisper evolution of a difference that you can see sort of in the public domain right now. Spellbook's moving towards autonomous contracting.
3:23Right? So, that's on our core of our website is what we do is we make contracts happen almost instantaneously and with less mistakes, less pain, and less cost than humans. There's no reason 90 some odd percent of contracts in the United States require a human lawyer.
3:39Most of what contracts are are cutting and pasting and shuffling around. Are there custom contracts that are incredibly difficult, incredibly arcane, and require incredible legal sophistication? Yes, there are, but most of us do not deal with those in our daily lives. Like think about this, when I first started investing YC sort of pioneered this concept of a safe note.
4:00And a safe note honestly requires about 3 minutes of explanation to really understand what's going on and maybe 30 minutes to master.
4:11Before that, you used to not have to use like sophisticated counsel on both sides of a transaction and negotiate blah blah blah blah blah and it was very expensive. Like a typical equity financing might cost $20,000. Maybe inflation adjusted maybe back then it was like 5,000 10,000. But in any event, it was a complicated transaction to get $500,000.
4:31And then I was like this is kind of ridiculous. Why are we wasting all this money and time and effort on lawyers when 99% of seed investments can be done with like a few bullet points. And that's what they call the new Combinator these days everybody uses a safe note.
4:44Yes, you may occasionally run line one or so but like almost every seed round gets done on some version of a safe note. So I think all contracts are basically like that.
4:53Most contracts should be like a safe note where a normal person can understand. AI can show you all the options and degrees of flexibility, explain to you the trade-offs and even can counter even negotiate with the counterparty who's also AI and if we're going to speed up the world of contracts it's not going to be done by humans. There's 2,000 contracts that go into putting together one NBA game for the New York Knicks.
5:19Imagine if that and almost none of those are fairly unique uh super complicated contracts. Maybe a handful of the 2,000. Imagine all that could just happen automatically. So with Spellbook, we did do this for a Fortune 100 company and we'll release the exact name um you know sooner rather later and we saved them $21 million of outside counsel legal fees over 3 years by their own estimate because they were spending all this money on outside counsel to
5:47deal with contracts that are kind of mundane. Part of the art is always knowing what kind of problem you're trying to solve and you know or where on a distribution curve are you. AI is excellent in the middle of all curve distributions. It's like self-driving cars.
6:03When you're dealing with a unique circumstance that's never been seen before, a road that's never been traversed, weather conditions that have never been encountered, humans are actually going to drive pretty well. But, the evidence is incredibly clearer that Waymo is probably 5 to 10x safer than letting humans drive most roads. And so, that is true with contracts. Automatic AI is going to make less mistakes than the average American lawyer.
6:31And there's going to be cases where you need incredibly unique custom judgment, and you should pay for that. But, most people, most enterprises, do not need that most of the time. So, AI can be cheaper, faster, and more responsive.
6:45So, that's just an example. Now, you know, Harvey has a different strategy, and I think it's a very coherent strategy for the way I understand it from afar. I'm not an investor. I actually think what they really want to build is kind of a one-click button that allows a CEO to see the entire world of law and regulation, and allow their company um to ensure it's getting the best
7:13in all those dimensions. They're not That's not what they're doing exactly today. They're starting though with a lot of success with general counsels of large companies, and they'll just continue to satisfy that demand. Both can be incredibly successful businesses. If you take that as like the vision level, they are um fairly distinct.
7:33Do you Do you think that that sort of verticalization, so you know, kind of taking on some of the work that historically a SaaS company would not have, is is kind of the the driver of your interest in a lot of vertical AI these days? Like, do you do you see a similar path for Rego, for example?
7:51Yeah, Rego's interesting, too. It's a classic, very successful vertical application. Uh it has dominated investment banking. 29 As far as I understand, 29 of the top 30 investment banks in the world use Rego. And the other one eventually will. Um we'll get around to them one of these days, but uh in any event, I think the reason why is there is a lot of mundane rudimentary work that goes into what junior investment bankers do.
8:19In AI and technology is excellent at giving these junior investors, investment analysts, and stuff uh superpowers. And so you're turning an average analyst into a super analyst. And over time that may mean that banks may hire less than them. We'll see how that all plays out.
8:38But we've made the work much more delightful. If you know junior investment bankers, they they have pretty difficult lives. Uh so we've definitely made their lives better. And I think we make we're making their clients' lives better because presumably our AI products like Felix and stuff will make less mistakes, give you better outcome outputs than the middle of the bell curve uh distribution of I bankers, too.
9:03So you mentioned, you know, with with Spellbook, the ability to just handle contracts soup to nuts, and presumably there's some sort of um you know, gathered intelligence from clients that over time will make Spellbook better than everyone else at doing this, right? Um at the same time earlier you mentioned kind of building on top of foundation models is maybe not the best path. When you look at things like Harvey's tenant, do you think those are dead in the water? There's no opportunity for kind of
9:32uh stuff after after model training or do you think vertical AI companies they do benefit from building up this this data asset?
9:40It depends what vertical you're in. I think you definitely need to build uh probably your own model if you're in the path of a frontier model. You're going to have to have a faster and better and deeper data loop. So for example, Cursor, it was a heroic acquisition, you know, everybody made a ton of money, founders, investors, etc. That said, their future as an independent company was pretty bleak.
10:07Uh now, could they have built their own models successfully? Possibly yes. They had a lot of talent team.
10:14Secondly, um, they did have a lot of data. Like, they had they kind of pioneered like the true scalable version of self-coding. Um, you know, they were accounting for like 30% of Anthropic's use, you know, sort of. And then, um, I they I think the issue was to build their own models successfully vis-a-vis the found the frontier labs, they probably would have had to invest called a billion dollars.
10:40And the risk reward of that decision first being acquired for a lot of money didn't make sense, you know, to the decision makers involved, which I understand. Had they started though building their own models maybe earlier, that price tag might not have been as crazy and they would have had more evidence also that it would work. And so, it wouldn't have been quite as, you know, steep a bet. So, but that's a classic illustration of I think Cursor
11:09could have survived, but not once they really didn't invest enough, fast enough, deeply enough in their own model.
11:18Yeah. I heard, um, and I could be remembering this correctly, but I heard that Composer, I mean, they only launched it like 6 months before they they got acquired, but um, that it actually had composed some significant share of usage before they got acquired, but it seems like maybe they were betting like, "Hey, in order to compete with all this volume that the foundation models were seeing directly, we would have to just keep on raising forever." Is that kind of what you're saying?
11:45If you have a loop of data creation like Cursor or our code, if you're if you can compound that loop it and do it extraordinarily well, it's really hard for someone coming from behind to catch up cuz you're just going faster and faster and faster. Now, there may be diminishing marginal returns in some verticals and maybe not diminishing marginal. Like, in driving, self-driving, there's not really diminishing marginal returns cuz there's so many roads times so many conditions uh, that whether Tesla or Waymo has the most data, they're going to keep acquiring data and
12:15getting better and better off a very long time. Maybe encoding, maybe it's some diminishing returns to the last second or third players to catch up or something. Maybe, I don't I don't know. There are verticals where like corner cases matter. For example, law. I think in date let's talk about contracts just generally.
12:33By vertical, you want data. So, like how is a typical NBA contract work? Maybe very different than how does a contract work in let's say consumer package goods or or let's take space. You know, there's a lot of companies in space now. Very complicated contracts putting somebody, you know, putting a rocket probably has hundreds to thousands of contracts. I think we've actually published the data on it.
12:57So, I think you need to cut cut cut the data and slice it to have enough data. So, that's one dimension is can you stay ahead just by making a loop really run really fast?
13:07You have to keep You have to keep slicing and slicing. It's really hard. We don't have a lot of ground. Second is you have proprietary data, which can be very real. In vertical like Rubicon for example, a lot of the most important data that I bankers use is not in the public domain.
13:21I have a company also mentioned that's kind of a vertical AI company called State Affairs where we allow large like Fortune 1000 companies to monitor all across the US particularly in at the state level specifically what's going on in terms of legislation, pending legislation, how it affects your business, etc. And this tool didn't exist before. And some people have intuition that this data is publicly available cuz it's stuff the government's doing. The truth is it's public but not accessible, which is a difference. Like there are states like
13:51Wisconsin where you if you don't attend the hearing, you can't get the data. You can actually send somebody to the hearing and he or she can write it down and it's publicly available. But there's no online reservoir of that. So, if you don't have somebody in that hearing, good luck. So, you can find another advantage vis-a-vis a model with non-accessible data. So, compounding loops, non-accessible data, cutting by small cells and slivers, corner cases, all that stuff, gives you lots of potential.
14:18Then the question is also, does the last mile matter? There are some verticals where you can have sort of the right answer, but unless you're standing behind it, and maybe law kind of works this way, maybe medicine works this way, where you're actually delivering the full service, like I'm saying this is the right legal answer, this is the right prescription you should take, and I have malpractice and other, you know, sort of regulatory requirements, getting 99% of the way there may not be
14:47the same product as the full solution. And that that also depends by vertical.
14:56Do you Do you think that given now, you know, there is a path some defensibility, at least in some, you know, verticals through, you know, data or maybe network effects as well, and they can do more, that the the market size opportunity in some of these spaces is actually much larger than it was in SaaS, and, you know, are you seeing that play out?
15:18Yeah, potentially. Like So, I'll give you a like a case. Like So, for example, okay, in Ramp, it's I think it's easy to describe. They're not like a classic vertical AI company, but by metaphor, I'll apply them to some classic illustrations. So, Ramp's goal has always been, we're going to be the app that a CFO has on his or her phone and controls everything that the finance organization needs to do.
15:41And 6 or 7 years ago, that felt like almost impossible. Now, it actually feels like next year we should be shipping like this mobile app that the CFO sits around, and from FP&A to closing your books, to procurement, whatever it is, enforcing policies, changing policies, all of that just happens. Token spend, token management, token optimization, token ROI versus human ROI, all of that should just happen in app. So, the future organization should need just one CFO, which is legally required as a Delaware corp.
16:11Think about now, like apply that to Harvey. Could I envision Winston and his team building the equivalent across all law and regulation that someone CEO, COO, CFO, GC sits around and has everything they need across the planet to both monitor impact law and regulation, yeah, that'd be a massive business if you could pull that off.
16:37Cuz also like looking at um you know, Comp like there's a core um I believe gave you you did Comp, right?
16:46Comp's a great illustration.
16:48Yeah, like they have and they have, you know, they it seems like they're starting with, you know, this almost like four deployed services and then maybe automating that over time. That's something that talk about Comp. Yeah, maybe we should give a little introduction cuz most people are not going to be as familiar with Comp as some of the other names that we talked about. So, Comp basically replaces humans in HR with AI. And they launched originally in Brazil and have a lot of traction in Brazil with very large companies like
17:16public companies, etc. They're launching imminently in the US uh and, you know, will do the same. So, if you don't want to scale your HR group or don't want any of the people in HR and want AI to do everything from Comp to benefits to whatever, you know, whatever you rely upon for HR, AI can do that for you and that should be a pretty strong value prop, especially if you can do the breadth. I think if you can only do a sliver, you have a couple issues where maybe that's not what the customer
17:45wants. So, you're not doing the proper job to be done, so to speak, and you might be a little bit more dependent on the frontier of law than you want to be. If you can replace humans that run HR, A, most CEOs I know are going to be absolutely thrilled. And then B, there's no way Anthropic's like replacing HR departments.
18:02It almost seems like if I zoom out, it almost seems like there's there's two buckets emerging in vertical AI. You basically have like industry specific GPTs or industry specific agents, right?
18:13And and those are either those a lot of times will look like ChatGPT or or or Claude, but they're industry specific, right? I think Harvey probably looks like that today. Rogu kind of looks like that today. Obviously, their vision is grander, but at the surface level, that's what it kind of looks and feels like. Um and how those evolve and change over time, you you gave us some some insight into.
18:34And then you have this AI native kind of AI native services bucket, and that still feels kind of that still feels fresh to me.
18:42Like I I see a lot of vision and opportunity on, "Hey, look, we're going to just handle this, you know, from A to Z for you." But are you seeing any like breakout AI native services businesses, or is it all really like it seems to me all the fastest growing vertical AI companies today, all the zero, you know, whatever zero to 100 million in 18 months that you're seeing with like Legora announcing, they're all really just like the industry GPT.
19:06Do you agree with that?
19:07I agree that descriptively, but I you know, we focus and I try to focus, but I'd say we as a fund focus on durable, sustainable value creation. So, we start with the question, what could this be in let's say 2030, and work backwards. Is this going to be a sustainable, independent, successful company in 2030?
19:26So, can you grow really fast right now in some verticals? Absolutely. And maybe there's a path of least resistance that leads you to do that. I'm not as enamored with that strategy, typically. I think Rogu is a special case where I think reading Peter Thiel's Zero to One is more canonical take with a very defined market and dominated than it is uh similar like the Harveys and Legoras.
19:53So, I think it it is distinct to um but I think you can grow really fast, but the question is are you building a sustainable, independent, public company? And I'm pretty dubious per my example of Cursor. If Cursor probably couldn't have following the Cursor strategy, if you sell at the right time and you have the right acquirer, could could it make everybody look good, but that's not really what the business I want to be in, and it's not the business our fund really wants to be in. We want to be the
20:20business of building um high-impact permanent companies. And I think there you have to deviate from that strategy.
20:28Now, you still have to get traction. So, there is a gap. You know, there is definitely Like, for example, okay, so Waymo took the strategy of we have an infinite money. Google mints all this money. Let's go map every damn road, you know, for years and lose a lot of money and that's really going to work. And, you know, it kind of is really working now.
20:44X years later. And lots of people on the Waymo team along the years didn't even know and weren't even super confident that it would work. Mhm. Tesla had a different strategy cuz they didn't have billions of extra dollars sitting around with nothing to do with them. So, they basically said we have to build a car that actually works. It's pretty damn good car. you know, Consumer Reports rates it the best car.
21:03We're going to use the car being driven everywhere to create the data at like basically effectively like effectively no marginal cost. And then eventually we'll get to enough scale where we can deploy the feature. So, I think most startups need to follow Tesla's playbook. They can't really follow Google's playbook. And, you know, [snorts] so therefore you definitely need to also have a wedge, you know, if you're like a proverbial way to express this into the market that gets you the scale, gets you the data, gets you the accumulating advantages,
21:31whatever. Or, you need to be one of the very few founders that can raise you know, hundreds of millions of dollars kind of pseudo one-shot. That's going to be rare. And, you know, that's also it's both a micro like what founders have the credibility and background where that might be possible. And then be what macro environment is going to tolerate investors taking that risk.
21:53Do you see any industries like if if I zoom in on that industry specific kind of GPT strategy, whatever we want to call it, and you and you take that strategy and you go after a vertical that isn't in the in the way of the frontier labs. They're not focused on it. They don't really care about it. They'd probably rather make money off of it via API credits.
22:11And you assume that that's the case for and, you know, a very significant portion of time. Call it decade plus cuz maybe it's too niche or it's too regulated or like whatever reason that they don't want to go in and do it. Do you think that strategy is enough to stand on its own two feet to build the durable company you guys are focused on or is it hey at some point it's just going to turn into AI native services, too?
22:31Maybe. That's a It's an interesting question for the founder. So, in account like we didn't talk about accounting, which is uh there's some interesting examples here. So, we funded two companies in accounting, one called Basis, which is more what you would think of as the Harvey Lagor etc. Rogo classic companies. We sell to most of the top 150 accounting firms. We make closing the books like automagical.
22:57Uh there's an accounting shortage, so accounting firms love it. They take on more clients uh with less accountants and there is a shortage of accounting junior accountants in the world. And we've created the you know an incredible experience that works really well. It's accurate. And then we also funded a AI native accounting firm. Built from the ground up on first principles that actually uses Basis actually in some ways as a foundation. So, it's a It's an interesting hybrid.
23:22Um so, it's very compatible with Basis, but it starts with we're going to do everything through AI and we're going to deliver whatever, you know, the client wants through AI, but we'll have the license and status to do taxation audit, you know, and accounting has some of these interesting dimensions, too, where even in a frontier lab can do some versions of accounting.
23:46Is it going to want to submit taxes and be liable, you know, for that penalties and stuff? Is it going to want to do an audit? Well, actually can't really do an audit. There's like requirements to be an auditor and stuff. So, I think actually accounting has all of these dimensions in one place. There's closing your books, which is a little bit more tractable and arguably frontier lab, if it had context, could do that.
24:08Then there's audit where you need a third party like by definition the whole point of having an audit is a third party. And there's taxation where you're you're kind of have this noose around your neck if you're wrong. Um so, accounting there's all these different angles, but we're really happy with both of the investments that we made in the vertical so far.
24:27Seems Keith, you you've made a lot across both, but I know you know, 10 years ago or so you kind of had a um you know, now now now canonical view on on markets that are interesting and you talked about, you know, focusing on fragmented markets with low NPS, um verticalizing into those.
24:46Do you today see SMB versus more concentrated enterprise to start as more or less interesting or you know, does does that view still hold up?
24:57No, I've changed my view completely. So, I used to hate the idea of me personally investing in enterprise. Personally, it's just self-awareness. So, I'm like, I know nothing about selling enterprise. I've never had any success doing this. Why the hell would I invest in something? There's lots of good enterprise investors.
25:11What has changed though in the age of AI and I I was very good and had a lot of success personally and then investing in people who would scale through targeting SMBs and even smaller, like micro merchants. Think of PayPal, Square as classic illustrations. So, I I'd love being able to use marketing and product instead of sales people. I had this allergy to sales people, which led to an allergy to enterprise.
25:35In the world of AI though, many of these companies uh are selling top-down. And the motion is very different.
25:44And that is leading to very different results and actually very different tactics and potentially very different org structures. So, some of the companies we named, and I won't go through all their strategies in detail, but many of these companies sell top-down and they sell top-down very very successfully. The CEO, co-founder goes to senior executive, CEO, board member, one level below the CEO and basically creates the sale. And yes, there may be an AE moving the paper across the finish line, which is a lot
26:13of work in enterprise sales. But then they're also coupling this top-down motion with brand marketing and you see this play out on X. Like if you could take a step back, I I find this absolutely amusing as a former lawyer. If you said, "What's the hottest topic about an X going to be in 2026?" Really technology. Other than maybe you know some snafus around anthropic I'd say legal AI is the hottest thing on X for technologists.
26:39Whether it's Harvey and benchmarks and you asked about Casetext model and and then Mural and the fights between them and then you have like you know by dimensional spell book native AI and there's podcast that are vertical like just about AI legal AI it's getting crazy and so this brand marketing angle and being able to scale company through top down sales plus brand marketing is a very different equation than the old school playbook of I hire this many AEs and they ramp at this pace and then we
27:08can predict our revenue because we just added this cohort blah blah blah blah and you layer it and boop slam dunk. I don't like to do that. I don't really want to invest in these companies. There are a few in the AI age that do follow that formula. I did invest in profound and profound they were not really a vertical AI although maybe they're shifting into that but they weren't originally a vertical AI company but they do follow a more traditional playbook and they've done very well with it. But most of the companies that people think like think Casetext another area we haven't talked about but
27:37Casetext has got a lot of top down sales and a hell a lot of brand marketing going on.
27:41And Keith do you think that changes the not only just the kind of early hires that are more important but also the the profile of founder you think is suited to win?
27:53Yeah, I think the founder either needs to be adept one of the founders doesn't always have to be the CEO but one of the founders you need to have founder credibility to to really do have the cachet to do a top down sale. Uh one of the founders has to be able to hold his or her own with high business vis acumen execs. Top down sales don't work the same way and so that's an important ingredient and then understanding what used to be called brand marketing it's it's as simple as storytelling
28:23but basically be able to cut through the clutter and become the default like there's a reason why we use shorthand to describe a bunch of startups already in this conversation. And uh they do so much storytelling that the average listener to your podcast probably has reasonable familiarity with at least half the names that we you know, we just recited.
28:45In a normal era of enterprise software the chance that up-and-coming startups would have names that would be recognizable to half of podcast audience would be like zero.
28:58And so the traject- the sequencing of building a brand, storytelling, and the leverage you get out of it is just different in the AI world. And that is actually exciting to me because I'm more comfortable in that world. Because we used to always have to use brand marketing um as part of our playbook. And so as the way these companies grow is being reinvented now I don't feel as dumb as trying to compete with a bunch of enterprise software like investors
29:27who really knew what they were doing because those playbooks aren't what people are following. So at a best case we're both back to zero and we're all learning, you know, from the ground up, watching and learning every day, which is really challenging and really fun. Uh so now I'm much more willing to invest in companies that start with hey, we're going to sell to enterprise.
29:45Do you think that strategy is a long-term strategy though? Do you like the in what I mean by that is this brand kind of up, you know, founder-led top-down That's a it's a very good question. I think it's top-down part may not be permanent. It works right now because the reason why top-down sales works is every normal organization in America sits around and says, "Hey, how do we adopt AI or how's AI going to threaten us? Blah blah blah. What are our competitors doing with AI?" So it's a CEO level conversation, board level conversation. And anytime you're on the
30:14agenda for a board meeting or CEO conversation uh CEO like priority list you can do top-down sales. If you're not talking about something that's on the top three priorities for the CEO, top-down sales is not going to be your friend. It's probably your enemy actually. The last thing you might want is the CEO paying attention to this newfangled technology startup that we're working with somebody we're going to pay them an extra million dollars a year.
30:37You may want to like bury it somewhere in New York. But if the CEO is under pressure to make their organization relevant in the future and ensure that the organization stays relevant top-down sales could work very well. And so that may be a p a point in time when the AI wave has already crested where organizations think okay, we kind of understand it. We've made our bets. We understand how it's working and not. And so the top-down might not work as well.
31:05The brand marketing and storytelling might work a lot longer which is if you're an investment banking, there's no chance now that you don't know what Rogue Wave is. Which would have been crazy for a 5-year-old whatever startup to have an entire industry where everybody knows who you are. But that part that strategy piece may work may have more runway to it as people learn how to use these techniques. I think you'll see more innovation. Like for example, it's not always paid. Sometimes when people talk about brand marketing like oh, we're going to put billboards out there. We're
31:35going to sponsor, you know, the Knicks or whatever or Lakers or the Seahawks hopefully. They know it's more money. So there's other things people are doing. I'd say Harvey is like arguably the most creative about this.
31:50All their benchmarks, frontier intelligence and you know, I glossed over your question but I think actually if frontier labs have different cutting edge abilities sometimes people call it like jagged intelligence then building your own model that finds for a particular task with the not just cost-efficient but performance-efficient choices actually could be very sustainable. So it's unclear to me as not being an AI researcher do we continue in an age of jagged intelligence forever? But right now we
32:19seem to have it. Like if you're going to write this kind of autobiography there might be a better product to use than if you were going to write poetry.
32:29If that continues then a harness-ish thing that's tuned to the whatever the task is using different models under the hood might have a long way to run and you can lot of fancy AI people I know debate this whether it's all going to smooth to one one or two models and there won't be these edges or the edges continue for very long time. I don't know, but that's kind of the bet but Harvey's pioneering a lot of the research and you see people following but I think it's helped them cut through
32:58the clutter pretty considerably.
33:00Maybe just slightly different angle on this, you know, we've talked about maybe how you know, company building go to market is changing. Um we seem to just see round sizes get bigger and bigger. Do you think this makes sense given you know, what's going on or you know, do you think this is sustainable or good?
33:21Is the party going to end and when?
33:24[laughter] All all parties The lessons of psychology is all parties end. And you know, I saw Ben Horowitz who's been through many cycles talking about how it ends last week. I [snorts] saw you know, Ken from Citadel's comments today or yesterday.
33:41At some point there's definitely going to be a transformation. My personal prediction is if there's any challenges with the Anthropic's IPO either timing or ultimate trading valuation that is even a blip for them will be felt as an earthquake for everybody else.
34:00Because part of what's going into these calculations is people are putting a zero on the expected value and saying, "Oh, if our office is worth trillions, great. Well, the expected outcome could be or or an outcome could be trillions and so we can pay billions." That's really basically a simplification of what everybody does every day in venture.
34:19Um if that's not true, then the zeros go away really fast. Now, what the the one thing that is let's say fundamental and not bad is there are times when people have an ambitious founders have an ambitious vision and they want to build apps. And what we do is we always say like what are your two three four five inflection milestones.
34:42What do you need to achieve that shows you're on the right track that you're going to succeed, and then you kind of sequence your financing around validating those inflection milestones. There are times when your goal, the slope is not gradual, and there is no inflection milestone except in a large step function.
35:00Those kind of companies, you kind of either need to invest call 20, 30, 40, 50, 100 million, whatever, or you just can't get there because you can't hit an inflection milestone. That can be a very rational, very thoughtful fund deliberate fundraising process by the founder and the VC.
35:18Most companies, though, that you see raising, you know, a lot of money front-loaded are not doing that analysis, but there are times when that is the correct conclusion. If we don't invest 50 million, it it makes no sense to invest anything less because there's no way they can hit any milestone for sub 50 million.
35:36And so, those those those kind of investments can be great, you know, if you can achieve the milestone because it does unlock more capital at significantly lower cost of capital.
35:48No, that makes a ton of sense. I mean, it it seems to me like when you know, when zeros get added, if you will, for some of these opportunities that are perceived as the biggest market opportunities, it sucks oxygen out of the room for some of those deals that may still very well be 100x's. If they just may not be 100 billion-dollar markets. And um So, I don't know. I'm I'm curious like obviously this is as a as a investor, very dependent on, you know, fund size and how you're generating returns, but
36:18like, I don't know. How How do you think about it? Are you guys saying we only want to go for, you know, the biggest swings or um How do you think about what market sizes are attractive in this environment?
36:31It does relate to your fund size, no doubt. Um the size of the magnitude of the opportunity that makes sense to pursue because to some extent, there's this ex- existential sort of tension in venture. you get measured on dollars you return and dollars on dollars, dollars invested, dollars returned. You do also get measured on an IRR calculation.
36:54But the magnitude of dollars on dollars is probably net net, at least within some constraints of IRR, more important to how LPs evaluate you. And so, yes, you can produce a 100x on a million-dollar investment. But then, if your fund size is not trivially you're not very small, it just doesn't matter. It's a 100x and IRR on the investment is awesome.
37:19But you need to move the needle by billions if your fund size is reasonable. And then, if you have LPs who are large, they need billions of returns. And so, you how how you what LP base you have, what's your fund size, has massive impact on what strategy you is it coherent to have a certain strategy and are certain types of investments prudent or not. And so, it has to be aligned, just like a business
37:48strategy, just like you're running a company. Your market and the customer segment you're going after and the value proposition of the product and the pricing and your competitive whatever advantages you have have to be aligned or the business strategy makes no sense. It's true in venture. It's just and this part of tease all these things out from afar.
38:07Yeah, no, 100%. Just does seem to me like a lot of um I mean, investors and founders alike are kind of looking at the top of the market, you know, whether that's Anthropic or, you know, the things that Khosla can do may not necessarily be the things that everyone can do. So, like abstracting that to, "Hey, I should just throw ownership out the window because companies can be two trillion dollars now." It's you know, it seems questionable.
38:38Uh Yeah, no, I I think that's more right than wrong, but it is related to what is your fund, you know, structure, size mostly, and then B, um a little bit around what's your competitive advantage, too. Like, what's your alpha as an investor? Like, how do you want to compete with the entire set of investors? So, if you're going to compete on some dimensions, ownership might be more important. Like, for example, if [snorts]
39:07uh you're going to be a high value-add, a company-building partner for the founder, consigliere for the founder, ownership matters a bit cuz your time is uh there's a significant opportunity cost that's probably more expensive than the capital cost. On the other hand, there's very successful funds that are momentum investors, and they'll try to be helpful if they can, but that is not the core of their DNA. So, they don't have the time opportunity cost constraint.
39:32So, for them, the ownership might be a distrac- might even be a distraction in a kind of red herring sort of thing. So, it does depend upon how you compete. Like, I compete as the consigliere to the founder. That's what I do.
39:44And so, for me, the scarcity is not how many dollars I can invest. I have too many dollars sitting around that if I want to invest in any given amount of time, I can I can definitely invest.
39:55But, I do have significant scarcity on high-quality time. And so, I want to make sure that I invest it in the highest potential people with the goals, with the highest potential impact, and that if we're right and we're successful together, that it generates enough returns to offset the opportunity cost of all the time. And that that means I need to think through at least ownership constraints.
40:21All right. Well, Keith's got a hard stop here, so let's end with I want to end with a couple rapid fires if that's okay for you, Keith. We Yeah, fire away.
40:28We like to do the a few bull or bear questions. So, first first one, uh bullish or bearish, AI roll-ups? What's your take?
40:38I'm pretty bearish. More business fundamental value is created on growth than cost reduction. I think most people who are thinking of roll-ups are gravitating to using AI to eliminate to reduce costs. which does produce better cash flow, isn't it? The But I think more than value creation in terms of enterprise value creation is can you create unusual growth stories?
41:00This is why Ramp is valued so highly. We're excel We're accelerating growth seven quarters in a row at over 100% annualized. As far as last time I did the research, and it may be same data, but I did the research six months ago, in the public markets, there was one stock that was over 100% with break even or profitable, and it's called Nvidia.
41:24At the time I did the research, Ramp was the only equity you could buy that had the same profile. There may be a couple others now, but the fundamentally, it's a very scarce asset.
41:34So, I think if an AI roll-up was concentrated solely on how to use AI to propel growth, and it had evidence that that would work, I'd be more excited, but that's not 90% of what I hear when either my colleagues or founders are pitching me on AI roll-ups.
41:49Okay. And then last one for me, uh bullish or bearish vertical software incumbents? So, like Toast, Service I'm I'm pretty I'm pretty bearish. Um I do You know, there's a timing thing, and it may be people over rotated like initially, but I think I think there is going to be a new a new wave of companies in every vertical built from the ground up, a little bit like mobile. I think it's a fair metaphor is probably where there's some incumbents from the web days that
42:18made the transition to mobile. There were a few. But if you think about, you know, take your home screen, pull it up, and like look at the apps you actually use, and the associated time, cuz if it's on your home screen, it's going to be a disproportionate amount of time, not many of those were like what were called internet companies. So, I think it ultimately plays out like that.
42:37Two quick ones for me. First one, um you know, maybe maybe a little loaded from the semantics, but uh bullish or bearish deep tech?
42:46Oh, very bullish. I mean, that's what Khosla Ventures does. I mean, I wouldn't be working at Khosla Ventures. We've been doing this for 20 some odd years. We We love the more technically differentiated and all sort of the more difficult but promising, the better.
43:00Mhm. And what do you I'm just quick follow up, like what do you consider deep tech and are there portions of that maybe you're not as excited about versus are?
43:10I It's a It's a great question. I think we're the the way I defined it is is the only market where is the only risk in this investment, the primary risk.
43:19Can this actually be built?
43:22That's a tech risk company or what I call deep tech. Like is this actually possible to do? Is it invention, nobody's ever done this anywhere before in the world?
43:31Or is the issue can we actually communicate the value prop to customers or will they buy it at a price that we like?
43:37That's market risk, that's standard stuff. When we make a deep tech investment, we're we're sure that if the team can deliver, there's billions of revenue. And that and so that's deep tech to me.
43:51Is like there's no doubt about it. You deliver the value prop the way you described it on your slides, there's infinite customer demand. Like so I have a 5-year-old kid. I'll give you an example, it's really funny. He told me he wants he wants to be he wants more money than Elon. I was like, well, that's I'm surprised you even know who Elon is at 5. This is literally a true story. And I was like, well, you know, the only way to get more money than Elon is you're going to have to invent something. And he's like, okay. He's like, it's going to be called super fast airlines. I was like, okay, what's it do? He's like, you can go from here to
44:20New Zealand in 1 second. I'm like, yeah, that might work. Now you just got to just Now you just got to decide it. I'm like, you solved That's a deep tech investment. That's great. Now you just They can hop in and I'm not worried about having your ability to make money.
44:35[laughter] For 5-year-old, that's actually pretty damn good.
44:38That's pretty solid for five. That's awesome.
44:40I'd I'd buy a ticket, you know, if I I'm like, yeah, I'm like, how do you even know He's never been to New Zealand. I'm like, how do you even know New Zealand exists and like that's far away?
44:51I certainly couldn't do did I was five.
44:54[laughter] Oh, man. Elon, someone's coming for you.
44:58Very very very last one. Bullish or bearish podcasts?
45:03Oh, interesting. Um pretty bullish. I think, you know, there's a there's always the question of cutting through the clutter with a unique voice in any form of content, there'll always be an audience for. In writing, in podcast, the more the more you have though, the harder it is in some ways to cut through the clutter. But, like you saw the announcement this morning that, you know, Stripe is sponsoring Stratechery.
45:30It's still the best. Like I've been reading Stratechery, you know, for like a decade plus. And by far the most consistent, original, and useful thinking in tech.
45:42[snorts] So, if you could do that with a podcast, I'd be very bullish.
45:45Excellent. Well, we're we're working on it. Um but, Keith, this was awesome. I know you you got to run, but thank you so much for joining us.
45:53A pleasure. Thanks for the invite.
45:54Thanks, Keith. Appreciate it, man.
45:56All right, take care.
45:57All right, talk soon.
