How AI is opening up new markets and impacting the startup status quo with Sarah Guo and Elad Gil

0:05Okay, hi listeners. Today you just have me and a lot shooting the what's the appropriate term here. Shooting the breeze. Shooting the breeze. But I want to start this shooting the breeze session by talking about this Goldman Sachs report that everyone's reading which essentially says, I'm just going to get on my soapbox for a second here, that the title is something like calling the top on AI. So for obvious reasons I don't like it, but I do think it's worth decomposing for a second. I do encourage everybody to go skim this thing. So there's a bunch of interviews in it and And two of the core ones are from this guy, Darren Asimoglu and Jim Covello, the respectively like MIT professor and the GS head of global equity research.

0:52And Darren is arguing essentially that AI is going to impact less than 5 % of all tasks and the trillion dollars of CapEx that people are spending on training models is a waste because AI will be unable to solve the complex problems. It's not built to do that. And Jim argues, you know, he argues that in contrast with the Internet, where you are disrupting something expensive from the beginning, even early on versus having a very expensive solution that then becomes democratized. You know, AI is very expensive from the very beginning. And then the other argument he makes is that any efficiency gains from AI will be competed away anyway.

1:36And so, like, you know, none of the companies are going to gain from this. And so if we just like talk about Darren first, Darren's arguing about something he doesn't understand. Like he his claim is, you know, how do we know scale works? More data won't make customer support reps better. I think like that's just a fundamental misunderstanding of the technology and also objectively of what has happened over the last several years of scale and data improving capability and quality of model outputs. I think a lot of these folks too, by the way, are just kind of stuck in the old AI world. Like I haven't read the report, so I'm not talking specifically about these authors.

2:16But a lot of people are treating this like old school ML. And they don't seem to realize that there's been sort of a breakthrough in terms of these transformer based models or other architectures that effectively are both highly scale dependent, but also provide different types of functionality and features. then you're sitting there and you're munging some data and effectively doing fancy regressions in some sense. So I think that's the other issue here in terms of a lot of what I hear. This happens a lot in healthcare. In healthcare, they always talk about how data is the new oil. And you're like, data is not the new oil.

2:49Sometimes data is useful and well-labeled data can be extremely useful, but a lot of it is also about the model and the application and everything else. And so I think there's just this broader misconception in terms of how this stuff works and what it means and all the rest of it. Yeah, I agree with that. I think this is actually a case of like this time it's different and also people lacking, you know, even the market state of what's happening on the ground. Like there are absolutely things that are cost effective to do today. That's why you get actually a series of companies that are democratizing capabilities and really more on the prosumer side.

3:20But, you know, beginning to see things even in, for example, healthcare, traditionally really slow industry, where you go zero to five or 10 million of run rate in your first year. I think like, as you said, one sort of problem with this framework of thinking is, you assume AI is like what it has been in the traditional ML world. The other is this assumption that the tech won't get much better fast and it won't get cheaper fast. And I think the willingness to predict 10 years into the future of insignificant improvement is ludicrous when literally all of the people working on this tech are unwilling to, like you and I are probably unwilling to predict two years into the future, much less 10 years.

4:01Well, I mean, the thing I'd predict two years into the future is that there's going to be even more broad spread applications of it. So I think it's almost the opposite thesis, which is this is the early days. And if you look at enterprise adoption of AI, most large enterprises are very early on. They think of it as three things. What sort of vendors can I buy AI related tools from? What are my internal tools and how do I to adapt into AI. And then third is how do I think about it from the perspective of external customer centric products? Almost everything that's happened to now, up until now has been like vendor buys, Decagon, Hervey, et cetera.

4:33And then the product side is very early and it's really the AI centric for consumer companies, ChatGPT, perplexity, et cetera, that are, that they've provided functionality at scale so far. And so the big wave of enterprise adoption hasn't even happened yet, right? It's very early days. So all the impact is in the future. Yeah, I think the predictions I'm willing to make with strong confidence is models will get better across a bunch of domains. And like we're very early in exploitation. And as you said, like there's a sequence of enterprise adoption where we like haven't really gotten there yet because the planning cycle takes so long and, you know, all these natural frictions.

5:08But I do think one of the flawed assumptions in Jim Covello's picture of the future is that the way you apply these advantages is standard across companies. That's like saying, oh, everyone will use the Internet and therefore there's no economic gains to be had by companies when the Internet happens. But like you clearly get Amazon, you get borders and you want to be Amazon or you want to be Klarna or, you know, whoever is actually changing their cost structure dramatically here. Yeah. By the way, this is not a new story during the Internet era. There are people who said that the Internet was meaningless.

5:42And the CEO of IBM in the day had this quote, the Internet companies were fireflies before the storm. And the storm was companies like IBM adopting Internet and et cetera, et cetera. And then, of course, now the trillion dollar market cap companies are Amazon and Google and the like. So this is an old story. It's old wine and new bottles. Time to break the bottle. What else is like inspiring you or bothering you from a markets perspective? You know, I think the really interesting thing about AI is that it's opening up markets in unexpected ways. And I think there's three sort of drivers of that.

6:17One is we have new capabilities. And in particular, we have new capabilities across multiple fields, you know, biology and robotics and obviously language and image gen and video and things like that with a lot of emphasis right now on language and image gen. And so these things are very generalizable and you train one model to do lots of stuff versus doing bespoke mini models, which was sort of more of the norm before. Second, you can access it through an API anywhere in the world, which means you don't have to build all the ML ops. You don't have to do your own model training. Like you can use it out of the box with a simple API call.

6:48So suddenly anybody can use it. And third, a lot of companies and organizations have a mandate to do AI. and that means the markets are suddenly open in ways that they haven't been in two decades when the last time that happened was with the internet but they said oh we have to do something on the internet we have to be like an e-company i think is what people said you know those three things are driving unexpected behavior in terms of how you can address markets and um what i and you know the the small team that works with me have been looking at is um market by market say you look at services and ai transforming services and i think i talked about this before where you know um Software spend in the U.S.

7:27is something like half trillion dollars a year. There's probably five trillion in headcount costs for services industries that could be transformed by AI. You know, things like legal or accounting or, you know, you name it, sales, et cetera. And as we've gone through market by market, we've basically been looking at, you know, are there companies that exist? If a company doesn't exist, should you incubate something? or other situations where you actually want to do a buyout that's AI driven because it's better to have somebody take over the asset or the company and then dramatically change the cost structure or the leverage per employee using AI.

8:05And so I think for the first time, two things really make sense. One is incubation, which usually is a terrible idea, but there's a few things that I and my team have started actively working on. And then the second is AI driven buyouts. And I've now backed one or two of those with the idea that you can suddenly do things with AI that you couldn't before and you can kind of change the game in a market. And so and that's beyond, you know, obviously all the really exciting stuff happening in terms of model architectures and infrastructure and all the rest. But I just think from a market's perspective, which is sort of our topic today, there's really odd things you could suddenly do.

8:37And part of it is just the buying behavior has shifted because everybody has an AI mandate and they're willing to consider products that they wouldn't be able to consider without AI. So it's a very exciting time from that perspective. Yeah, I want to come back to incubations in a second, having done a few of these. With buyouts, like just to articulate like, you know, what I think is the premise here, because we've also looked at these as you are shortcutting the change management process for industries that can be like dramatically automated with these new capabilities by just controlling them.

9:13Right. You're actually shortchanging two things. You're shortchanging adoption of technology, which in some industries may just not happen very fast. And in some cases, there's an incentive not to adopt it. You're shortchanging change management to your point. But third, you're actually taking over an asset and you'll be able to entirely rework the way that a subset of the organization functions relative to AI. And Florida kind of did that to itself with this customer support team where they reduced headcount there by 700 people by effectively using OpenAI and some custom workflows to do better customer support.

9:46And suddenly it was 24-7. It was, I think, almost 20 languages. It was higher NPS than the customer support reps. It was faster response time. It was lower repeat incidents. And there's lots and lots of industries where a lot of the cost is associated or leveraged on time. In other words, you could keep the same employee base. You could just make them 10 times more effective in terms of the set of customers they can serve. Where a lot of the work is basically what people call email jobs, right? You're copying and pasting data from one spreadsheet to another. You're responding to emails. You're in a CRM, whatever it is.

10:20And some aspects of that are now suddenly accessible or automatable via this new type of Gen AI. And just to articulate the case against it, too. Like, I think it is, you know, managing services companies, identifying assets and the operational intensity of, you know, doing acquisitions as well is a different skill set than most, let's say, software engineering heavy startup teams are. And so I do think that is the question. Like, you know, can you get the right people to run these things? Yeah, you kind of need both because the traditional playbook on what used to be called like tech enabled buyouts, which largely didn't work, is you'd have a PE person come in, they'd buy a bunch of companies, they'd roll them up, and they'd put this really thin veneer of software on top of it so that they could claim it as a technology company.

11:12And effectively, what you're doing is you're arbitraging a tech valuation in order to buy other sort of EBITDA rich assets. You're looking for cash flow, right? But you're buying it for cheap at a normal private equity multiple. and then you're raising money on the back of that revenue or cash flow using a tech multiple, which often is, you know, a few X higher. And so effectively, a lot of those early things were arbitrage and the early investors and the founders of those things tended to do well. And the late investors and late employees tended to do poorly because as the thing increasingly became recognized as a private equity play versus a software company or technology company, the market cap and multiple would come down and normalize, right?

11:52It would lose value at the later stage it got because the investors became more savvy about what was really happening. That's different here where the leverage on the technology is dramatically higher than just adding a thin software veneer. But, you know, one of the things I backed is actually driven by a software founder who's hiring in a PE team versus the other way around. And obviously there's very good PE people who are hiring in, you know, or working or co-founding with ML people or AI people or software teams. But you need to make sure that you have the right mix and that you actually know what you're really doing.

12:25Because otherwise you're just doing a roll up, but you're calling it something else, which, again, is fine. Right. Yeah. But you just need to have some roll up like you need financial engineering DNA and investor DNA as well. I think one of the things that makes this like idea resonate with me is I have some friends running, let's say, the current generation. You know, these people, too, of like sort of tech eating services companies and traditionally very fragmented industries. and like in a small room, they would say, man, like Sarah, Elad, the automation works and the distribution is the problem going and acquiring the customer relationships.

13:00Yeah. And so you're buying the distribution. Yeah, exactly. You're just buying the whole thing. OK, let's talk about incubation. So my thinking on this historically has been like, OK, when does this actually make sense at all? It's when like you really like have visibility into an opportunity set or a customer set that others do not. And the expertise set required to make a company work, like the DNA is unlikely to get together organically. And so that's the alpha. How do you see it? I think incubations usually are a terrible idea. And most firms or groups that incubate companies tend not to work very well.

13:42And there's counter examples of that, right? There's a handful of firms that have actually done very good incubations. and often, but not always, they're kind of vertically specialized. And so they really understand the dynamics of DTC or they really understand the dynamics of healthcare or they really understand the dynamics of X thing and they have proprietary access to customers or product ideas that they know will quickly take off and resonate or they have a captive customer base they can sell to or they know how to roll together assets, right? And so I think the way that incubation has tended to work historically is if you have some form of deep expertise and relationships that you can leverage.

14:17And most people just didn't have that. They would just go and try and start a company. And most, you know, the reason to some extent the startup ecosystem works is you have 5 ,000 different founders simultaneously doing a parallel exploration of multiple markets and multiple technologies until a small subset of them actually hit something that works, right? So there's probably like a handful, you know, three to 10 companies a year that actually matter out of the thousands of founders, right? And so, you know, a lot of these approaches traditionally have not worked very well. And those counter examples, you know, Snowflake was famously an incubation by Sutter Hill Ventures, etc.

14:54But, you know, most of the time these things haven't done great. And there's other positive examples, right? Right now, because of AI, the odd thing that's happening is that certain subsets of the market are incredibly crowded. You know, you have a dozen companies all doing the exact same thing. And then you have these areas where it's really obvious to build stuff and they're just wide open. and nobody's doing anything. And then relatedly, there's a lot of very large enterprise interest in building things or adopting AI technology. And so there's all sorts of forms of incubation that I think are possible now.

15:26And so there's a handful of things that I'm working on right now from an incubation perspective, which again, normally I wouldn't do very much. The last incubation I did was about a year ago where I worked with Ankur Goyal, who had started a company that Figma acquired. I worked with him on a company called BrainTrust, which is sort of eval and, and, and, uh, Prompt Playground and a few other products all kind of, um, rolled together. Um, and that was just driven by the fact that, you know, uh, he and I had been sort of riffing on like, what does a, what does every enterprise need to use in order to adopt, um, AI?

15:56Um, but you know, it's, it's, it's rare for me to do those sorts of things, but, but right now there's just a lot to do because there's so many just clear market opportunities or customers to work with, or, you know, so there's, there's that customer pull or market pull that normally doesn't exist. So I think it's a very exciting time. One of the things that I think makes this make sense right now is the the set of people who understand the technology are not commonly the set of people who understand the domain. Right. This is why you have this like mismatch that you describe of like some really obvious open markets.

16:30And so I think if you can get like great engineers and technologists who understand what's going on in AI and the capabilities that are possible to actually apply them to where the enterprise problems are, that's exciting. And, you know, the first incubation I ever worked on was a company called Awake, where it was actually, you know, applying last generation machine learning techniques to network data for security use cases. But I think be it that or other – it tended to be like, hey, a technologist from a different domain is looking for the use cases that best match, which can be very dangerous.

17:07But if you know what it is, then it becomes a unique match. And so I think that is exciting. What do you pay attention to just in terms of like new applications you think should exist, stuff you want to incubate, whatever it is? Yeah, there's a ton. And if people are interested in working with me on something, they should obviously feel free to reach out through the network kind of thing. But there's one thing that I've been looking at very seriously on the healthcare side. One thing I'm working on that is sort of incubating is these large enterprise assets that I mentioned. And that's very exciting because then you have instant customers.

17:45and we're in the process of pulling together a founding team for that. If anybody, again, wants to ping me or apply for it. And then there's one or two areas that I'd really like to work on, but we just don't have bandwidth. One is in the services world and one is more like a new type of model or a large scale model for a specific application area. Are you working on anything in the area? Yeah, well, we just did a healthcare thing. So I'm very excited about it after, you know, many, many months of casting about. Sorry, I guess we talked about the really early stuff, you know, which is incubation and all that kind of stuff.

18:23The flip side of it is public markets, right? Like, and one question I've been increasingly getting is like, what do you buy in public markets given AI? Or would you change how you think about portfolio construction given AI? Or how do you think about, you know, existing companies and the risk of AI to those companies and things like that. So do you have any sort of thoughts or picks in terms of things that you think are exciting that are much later companies relative to AI or how to think about that? I mean, I obviously focus on the early stage, but I'd be curious for your point of view on this.

18:56I do think there's a, like, if I could short something, I think that there is a mid to late stage private company, small public company that does not have the speed and institutional and founder or leadership will to go change the business when they see the writing on the wall. And I think that that's not certainly not every company. I mean, we've talked to a number of like amazing founders who have attacked the opportunity really aggressively because they see the way to grow revenue or make the businesses higher quality. but I think that isn't going to be most companies. That's sort of the mid to late stage private set.

19:39So I think that is, that's something I would really think about in terms of portfolio rationalization. What about you? What in the public markets are you paying attention to? You know, I think of it as less just public market specific, but more like what are things that, you know, are durable in the coming world? And there's almost two versions of this question. It's like, what do you want your kids to be able to own someday or whatever, if you can afford it. And then there's what portfolio would construct today in tech. And, you know, the crazy thing is if you go back 10 years ago, there was kind of this era where people were talking about Fang, right?

20:17Those were the four companies you were supposed to just buy and hold. And that was like Facebook, Alphabet, Netflix, and Amazon, right? That kind of morphed now and Vidya is part of it. And they, you know, know, they keep kind of changing the rubric. But to some extent, you know, one question is, what is that next set of companies that have multi-hundred billion or trillion dollar potential? Right. And so there's some basket of like, whatever you want to call it, new thing or new, yeah, I can't remember what they call it, the great eight or the something seven or I can't remember what it's called.

20:51Yeah, it's the magnificent seven right now. The magnificent seven, right. So what are the, what is the next wave of that? And maybe that's SpaceX and Stripe. And, you know, you can kind of think through what do you think of these companies that may just keep compounding for the next 10 years. So that's one segment. Second segment is potentially some of these pre-existing big tech companies that will just keep going. Like I wouldn't be surprised, for example, if Apple ended up with like another iPhone replacement cycle just for AI. Oh, we have a new chip that runs these, you know, they mentioned that they're going to be working with smaller models that run these smaller models locally.

21:25You need to upgrade your phone, et cetera. So at least for some period of time, Apple may have a bump from that. So there may be some big tech. Third is what I call like AI durable companies. Like what are the companies where AI just doesn't matter for, you know, is it railroads? Is it certain SaaS companies, right? Where it just doesn't matter. You add AI, who cares? That actually shows durability. Like if AI doesn't improve the product that much, it's not really a competitive factor. so you can imagine a basket of stuff there. And then obviously there's like, what's the new AI index besides NVIDIA, right?

22:04NVIDIA is a core piece of just a way to index the market, but there may be other companies, maybe that's scale, maybe it's something else, right? But I think I'd kind of think of it across those four segments, you know, big tech subset, what's the new Magnificent Seven, what's durable in the face of AI, and then what's your AI index? You know, as we build out this whole AI wave, how do you think about staffing AI companies and post-AI teams? And what does all that mean? A while back, you said that, like, you think the founders right now, both, you know, existing and new, they've been like high quality, motivated, ambitious in a way you haven't seen in a while.

22:45Just really, really good. I think one specific trait that keeps or a specific belief that keeps happening is like especially second time founders. But both types like I've seen founders be really ambitious to have like a great pure metric of revenue per employee. Right. They're ambitious to have really small teams. And I think that's interesting. Like people talk about like the one person billion dollar company or whatever. But I do think this is kind of the purest version of like actually doing startups, right? Because you never quite had enough money to begin with. So you're always trying to solve the problem with the people in the building.

23:26And so, you know, solving it with fewer people in the building and more technical capability feels like just an extension. Yeah, it feels like we're a couple of years away from that, though, right? Like, um, there's some things that we can become dramatically more cost efficient on, but, you know, at least the existing coding copilots and other things give you some efficiency, but not, you know, 5X and maybe as some of these, um, you know, AI, uh, coding agents really start to work at scale. That's when you have this shift in terms of cost structure. Obviously there's cost structure shifts that we talked about in terms of buyouts and customer support or BPO or other things.

24:04But, you know, I think there's one of the questions in my mind is like, when do we feel this really hit? And what is the GPT generation equivalent at which it happens for enough of the economy or enough of a tech company, right? Like what portions of the tech company for a startup can be automated today? Because if you think about it, many startups are six to eight engineers and a designer or product engineer and then, you know, a founder or two and then sometimes like an early business person. Right. So there just isn't a lot to get rid of in some sense for very early companies. I don't think it's at that stage.