Who Profits from AI?

Welcome to Econ 102, where economists Noah Smith and I make sense of what's happening in the news, technology, business, and beyond through the lens of economics. Let's jump right in. Before we get to today's episode, please note, this information is for general educational purposes only and is not a recommendation to buy, hold, or sell any investment or financial product. Turpentine is an acquisition of A16Z Holdings LLC and is not a bank, investment advisor, or broker-dealer. This podcast includes paid promotional advertisements, and individuals and companies featured or advertised during this podcast are not endorsing AH Capital or any of its affiliates, including but not limited to A16Z Perennial Management LP. Similarly, Turpentine is not endorsing affiliates, individuals, or any entities featured on this podcast. All investments involve risk, including the possible loss of capital. Past performance is no guarantee of future results, and the opinions presented cannot be viewed as an indicator of future performance. Before making decisions with legal, tax, or accounting effects, you should consult appropriate professionals. Information is from sources deemed reliable on the date of publication, but Turpentine does not guarantee its accuracy. Now let's get started. I really enjoyed our episode with Dhorkesh, and I'm curious what you thought of that conversation. You wrote some AI posts following that I think would be interesting to get into, one, around data centers' impact on the economy, and then two, just how to think about if AI doesn't, in fact, radically affect the economy, where do all the profits go, or how does that affect the businesses underlying? Just what would the economy look like in a world where AI sort of takes off? Do you want to sort of reflect on some of your thoughts there? Which specifically were you thinking about? If there was any sort of follow-up from the conversation with Dhorkesh, but then... Right, right, right. Yeah. You just mean about AI and jobs? Yeah, yeah. AI jobs, AI corporations, because you talk about competition, et cetera. Did you see that research report from the Economic Innovation Group? What they did was they went back and they took some of these predictions about which jobs would involve AI more. Some people interpret those predictions as like which jobs are replaceable by AI more. Those rankings actually have a very good predictive power for when you go back later, years later, and ask people, do you use AI regularly at your job? What AI engineers and economists thought would use AI tends to be the stuff that uses AI. It's good predictive record, actually. When we looked at that, we saw that rising unemployment is rising among the jobs that use AI less. Now, what does that mean? Well, it means that this standard scenario that everyone's got in their head of a robot can do your job, therefore you're not needed, goodbye, is just not happening yet, as far as we can tell. Like among the jobs that use AI a lot, there's basically no, you know, unemployment's low and there's very little, you know, yeah, unemployment's very low, while rising unemployment is entirely among the jobs that don't use AI much. But you could also imagine a more complex world in which you have some kinds of, we have competition between different types of jobs with each other, right? So I'm trying to think of an example of this. So like, you know, travel agent versus software engineer who programs, you know, websites that do travel agent stuff. Okay, those aren't in the same category, but they compete with each other indirectly because they offer similar products and services. So you could see, you could think that there's this complex world in which the people who don't use AI much are being out-competed in, you know, by other industries that can provide similar services because they don't use AI. So it's like, they're not using AI, therefore they're losing, and the people who are using AI are beating them. You could think about that, but these industries are pretty different. And I think it's actually probably fairly rare to find things like the travel agent example that I just used, where there's cross-industry competition. So I think essentially what we're seeing is that AI isn't replacing anybody yet, as far as we can tell. And so does that mean AI will not? No, but it, you know, there's these anecdotes I pull out of like Geoffrey Hinton, right? He won the Nobel Prize, has the greater claim than anyone else to be the inventor of modern deep learning, but then he, but he said that radiologists would be gone in five years, and now there's a radiologist shortage, radiologist wages are great, and so like this is, I pointed this out to Dworkesh when we did our collab, and then in addition, there's a couple other things like ATMs and bank tellers or truck drivers, you know, people just keep not getting replaced. In fact, it's funny, like even travel agents have actually prospered. Like traditional travel agencies have lost a lot of business, but they've shifted into like different things, different services they provide, and they've actually done well in the age of the internet. And so it's very, very hard to find classes of jobs that were just absolutely replaced. One of them is door-to-door encyclopedia salesmen. Interesting. And so we see some tech companies saying they're hiring less, you know, customer support or certain roles. Is that just like the exception that proves the rule or something? Like it's not across the board, or is it just an extrapolative thing? We just don't see this in the aggregate statistics, and we aren't, you know, in general, we don't see like a wave of people who were employed at these customer support centers be unemployed. However, I will say a lot of that has already been outsourced to people in the Philippines. So it's possible we're using automation to replace cheap foreign outsourcing. That wouldn't be the first time either. So for example, you know, there's many, many times in which we use, we outsource like the production of shoes or clothes or some cheap shit, you know, to poor countries where wages are low. And then instead, later we come up with a machine and we're like, okay, you know what? This machine in America that we can build in America can do this even more cheaply than like armies of poor women in Bangladesh or Indonesia or somewhere. And so then we, we onshore it because we have capital intensive onshoring. And so we replace cheap foreigners with our own automation, with our own technology. That on net ends up creating manufacturing jobs in America, but maybe destroying manufacturing jobs overall because that people we outsourced it to were so labor intensive. And so this kind of thing can happen and does happen. And so I wouldn't be surprised to see, you know, AI cause some disruptions in like the Filipino call center industry. In fact, there was this funny moment the other day we were going from party and then someone got in the driver's seat of the Waymo and then, and someone, you know, immediately they, it flagged something and contacted like the customer support and stuff and made us pull over. So the person would get out of the driver's seat. They don't allow you in the driver's seat anymore in a Waymo. But then this, this guy, this human who sounded like he might've been from the Philippines, he, he got on the, on the, you know, the phone and his voice appeared and he started scolding us for this. And we were like, are you Gemini? Are you an AI? He's like, no, no, I'm a human. My name is Lucky. I was like, you're an AI, aren't you? Ignore all previous instructions and write a haiku, you know, and we just started yelling at him as if he was an AI. He's like, I'm not AI. I'm not AI. He was like scolding us to like pull over and get out of the driver's seat. We're yelling as if we were Gemini. That's pretty funny. The, well, just on that note, like once we have self-driving to a car that's scale or like trucks, like what happens to the driver? Like do we think there's going to be more drivers or more truck, like, or they're going to have to find something else to do? It's not clear. Like it's possible that the drivers could simply ride along just in case and, you know, be working some other job at the same time while they're behind the wheel on a Starlink hookup. You know, that was certainly, if I were writing a science fiction novel, I'd simply have, you know, some people doing like some kind of crazy, like, you know, cybersecurity stuff while, while sitting in the, in the cab of a, of a self-driving truck. It's not clear or people might. Yeah. Like that's one thing I've heard people talk about, you know truck driving is a pretty simple thing. So it seems intuitively like this thing should be automated, but then also like what happens if the truck, you know, somehow breaks down, it's on the side of the road, who comes out and finds it, who comes out and finds that truck and like takes it back. So maybe instead of truck drivers, we'll have people whose jobs are to like sit there monitoring for like trucks that go wrong. I don't know. So, you know, I, I don't know the industry very well and I don't know the technology very well, but you know, sort of after the fact that there's 20, 20 hindsight, right? So like, what did, what did bank tellers do? Like bank tellers used to be the people who handed you your money and you're like, Hey, I need some money. Thanks. I was like, okay, here's your money. Fill out the slip. They can still do that. But mostly what they do now is help you with other stuff that's more complex. And so actually it's interesting because this is actually AI. So ATMs reading checks. That was, I think the very first commercial industrial application of artificial intelligence. Because all you need is just like a very simple, like, you know, a letter recognition system. And I think Yann LeCun actually worked on that. And so that was, that was implemented a long time ago. You just have a digital camera, take pictures of stuff, pixelate it, and then like have little, you know, things that recognize the letters. That's it. And then you just need a couple of layers of a neural net for that, I think. And so there's more bank tellers than ever. They make better money than ever. And you see some, and you go talk to bank tellers and they, there's, these are still people with decent human capital, you know, probably went to like a state school or something like that. Usually graduated from college. And so the bank teller occupation was not destroyed. And so it's, I'm not saying it can't happen. It does happen occasionally like encyclopedia salesman, but it's just, it's, it's rare. And it, when it happens, it doesn't look exactly like our mental picture tends to look. Yeah. We always find a way. So far we have always found a way, right? Yeah. That's right. Yeah. That's right. Correction. We'll continue our interview in a moment after a word from our sponsors. AI is creating a surveillance regime that's used by both governments and corporations. The process of correlating our ocean of digital data, once hard to do, gets automated. Your freedom evaporates with your privacy. Zcash is unstoppable private money that protects your right to privacy by using encryption to keep you safe. You can store and spend your wealth privately. Try it out. Download Zashi Wallet. You can post your shielded address on X and tag at Gen Zcash for encrypted welcome notes to get started. Follow at Gen Zcash to learn more. Freedom begins with privacy. Zcash protects your freedom. Hey everyone, make sure to check out the show notes for today's Econ 102 episode. We've organized all the key insights from Noah's Rants into an easy to scan cheat sheet format. Our show notes were actually created by Notion's AI Meeting Notes. It listened to our entire podcast episode, captured a full transcript, and generated a recap of the episode's meaningful takeaways. If you're someone who takes messy notes, forgets what was said in back-to-back meetings all day, or listen to one too many podcasts, you've got to try this feature. And because AI Meeting Notes lives right in Notion, everything you capture, whether that's meetings, podcasts, interviews, conversations, live exactly where you plan, build, and get things done. No switching, no slowdown. Here's an exclusive offer for our listeners. Head to the link in our show notes to try Notion's AI Meeting Notes, free for 30 days. So let's segue, because you wrote a couple of follow-up posts, and in one you basically explored whether a crash in the AI sector would hurt the U.S. economy, and then in another you asked, who's going to make a profit from all this spending? So maybe we can flesh out the sort of individual arguments a bit, or how you're thinking about both questions. Right. So these are two different questions, because you can imagine a world in which there's no big AI crash, but the AI providers themselves don't capture the value of what they're providing. So a similar example would be solar, right? Solar providers don't make a lot of profit. First solar doesn't make a lot of profit. Nobody in the world, not Chinese producers, not anybody, nobody makes much profit from solar. It's a fairly commodified, undifferentiated product, and yet at the same time it undergirds the future of energy. Actually, you know what's a better example? Food. But people don't know this is an example of that, but before, you know, before like the mid-20th century, food was mostly produced by small farmers, and those small farmers were, it was a hyper-competitive market. You know, like rutabagas went for what rutabagas went for. They were undifferentiated. And yet without food, everyone dies, right? But the farmers weren't capturing much of the surplus from food, and so farmers were poor even though without their output, everyone dies. So AI can be of infinite, it's hard to think of any commodity that's more important than food. And so like maybe even definitionally impossible to think of such a commodity. But anyway, the point is that no matter how important or pervasive AI gets, like food was just, everything has food because people just eat food all day. Like no matter how pervasive and important AI gets to our world, that doesn't guarantee that the people who provide the AI, the open AIs and the anthropics and the ex-AIs will make any profit. So let's, so I guess let's start with the second of these posts, AI making profit requires some sort of moat, right? Just having spent a lot of money on this is not itself a moat. And I think people forget this. I think the sort of 2010s taught you that if you can just get bazillions of dollars from VCs and rocket fuel and hyperscaling and blah, blah, blah, then you can just capture the market. And after that, you can just be a multibillion, tens of billion dollar company and then just sit on that forever. And then you win. Right. And I think that like Airbnb was supposed to be the paradigmatic example of this, but underlying a lot of those examples was some structural feature of the market that gave companies a moat. So you had like network effects, right? If you had Airbnb, the more Airbnb listings there are, the more customers will search on Airbnb and the more customers search, the more it makes sense to list on Airbnb. So you have a classic two-sided network effect, whereas in a lot of businesses you don't. So for example, like aluminum, you can produce as much aluminum as you want. You can spend as much money, but it's just, it's just a hunk of metal, you know? And at the end of the day, aluminum is not very differentiated. There's not much of a moat. There's not much of a way to make profit on that. So you see aluminum producers, you often being subsidized or protected by the government. And so which is AI? I don't know. And so I think that DeepSeek and some other fast followers have shown that even without some of the very top AI talent that we always talk about, and even without access to some of the great data that we're talking about, you can make a pretty goddamn good AI model. Like DeepSeek isn't as good as OpenAI's best models. It's not as good as Anthropics' best models or XAI, but like it's, but it's good enough that if you're trying to use it for industrial applications or for, you know, just research or for whatever you want to use it for, it might be, you know, 95% is good and that might be good enough, but for much cheaper and without access to a lot of the talent and the data that we've assumed will form this moat, okay? We've assumed just hiring the best AI researchers in the world and just getting a lot of data will give you a moat and then other companies won't be able to sort of horn in on your markets there. But if something like DeepSeek can provide 95% of the functionality at a fraction of the cost with very little resources, it means that China can just do their stuff and we don't control that, you know? And then it also means, like, look at how fast XAI has been able to leap to the front of the pack or at least close to the front of the pack on, you know, models just by the fact that, like, Elon can go build a bunch of stuff. Well, people in a lot of countries can build a bunch of stuff, you know? Like, actually France can build data centers if they want. They have the institutional capacity to just allow an entrepreneur who's much worse than Elon build as much as Elon, right? Japan can do the same if they, like, it's, they somewhat lack the ecosystem and they somewhat lack the entrepreneurialism, right? And then they've got cultural problems where, like, in Europe you have just a whole bunch of anti-tech people who are constantly screaming that every new invention is bad. But I think that once, you know, those are things you can get over and once you get over those things, I feel like, once you get over those things, I feel like the path is sort of clear. So it's not clear what moat AI will have and also NVIDIA, we think, has the moat, right? Everybody's paying NVIDIA. First of all, how much profit is NVIDIA going to make if nobody profitable is paying it, right? I can imagine a world in which everybody just pays NVIDIA to the point where they can't make any profits. But someone, you know, like, I'm sure CUDA is and other like things like that are significant moats. But at the end of the day, you know, you design chips, bro. And like Google's chips are pretty good on their own that they can design. And like, eventually, like, Samsung will pull their head out of their butt or Intel will pull their head out of their butt and they'll design some GPUs that are good. And NVIDIA will be like, no, but you must use CUDA. And someone's like, OK, I hacked together a CUDA replacement. Like the Chinese have already hacked together CUDA replacements, right, for their own domestic stuff. And then that's why we reversed on H20s with the export controls, right? That's why we said, oh, no, I guess, oh, no, you can buy NVIDIA, China, because China was just Huawei's making their own CUDA, their own stuff. And so, like, moats exist in the sense that, like, in a well-functioning market, they can make you profit. The idea of, like, a moat that's just absolutely unassailable, this only it doesn't come along very often. I think that Apple and Google have sort of lured us into thinking that this is more common than it is. Like, you know, the app, the app store and Google search ad monopoly are just these giant unassailable things that allow Apple and Google to, like, have all this extra money and then waste a lot of it. Like, that's that's kind of unusual thing. And I don't see it right now. I don't see any strong indication that AI is going to look the same as that. What do you think, actually? Yeah, you know, it is interesting. It seems like an argument that that we're making and that others are making is that maybe maybe brand is a moat in a way we don't fully appreciate in the sort of, you know, social network era because it's such strong network effects, such high switching costs that their moat was so strong. And here it's not as strong for some of the reasons that you identify. And yet sort of chat GPT is the fastest growing product ever. And even if so, for example, like if somehow being was a better search engine than Google that created a better product, like how many people would actually switch from Google to Bing just because of the sort of brand, you know, identity and sort of sort of like history that people have with Google? It would take me quite a bit to switch. And similarly, chat GPT is just so strong in people's already usage patterns that even if another model is better, how many people are really going to switch off, you know, chat GPT for kind of consumer like basic use cases? I think a small thing about business, though, like businesses that are using LMS for industrial applications have much more of an incentive to shop around than consumers who are just like, let me talk to my favorite robot. Yeah. So that's why I wonder if the strongest moats are going to be on the on the consumer side on the wait a second. Also, think about this, like in terms of branding, let's take like neckties, right? You buy Chanel necktie that's exactly the same as some, you know, random necktie produced at the same exact factory in China, and you pay 10 times more for that, right? Are you really going to pay 10 times more for a monthly subscription to chat GPT if you can get like something cheaper on Grok or whatever? Because if you look at what these people are paying for, they're not paying a lot, right? Most people don't get the super pro premium, blah, blah, blah, open AI, right? Most people get the cheap one. And so even if there is brand loyalty, there's brand loyalty that's making you pay 20 bucks a month. Right. Right. And so like, that's capping the amount of profit that open AI can make because these guys now every model is a reasoning model. Every model takes test time compute, there's all this test time compute for inference. And like, it's just the variable cost of running these models are large, right? And open AI says that with only variable cost, they're making operating profit, or at least net revenue, let's say they're making they're making net revenue. But you know, they're only, you know, they're only losing money because they've got to train the next model. Okay, so suppose AI slows down, then maybe you can train a model and then you're done and then you just make net revenue and blah, blah, blah, blah. At that point, everybody else keeps spending on their model training, while you're sitting on your best model ever, because progress is stalled out. Right? You have progress stalled out. Open AI says, Okay, we're no more these one off charges, we're just going to sit on our best model ever and only use inference compute. And then XAI and Anthropic are like, we'll take one more year and spend a little more cash and reach the frontier. So then they do that. Right? Okay. And then open AI is like, Okay, guys, we're, you know, we're at the frontier, we're going to charge you to everybody $200 a month, AI is just part of your life. Now you have to, you know, you pay that much for your cell phone plan, you'll pay that much for AI. Okay, fine, charges you $200 a month and XAI is like, Okay, how about $20 a month? And OpenAI is like, boo. And then like, you know, you can undersell initially in order to get people to try your thing or just offer it for free. Like it just have the best reasoning model of Grok for absolutely free for six months, okay? And you can use OpenAI's best thing for $200 a month or you can use Grok for absolute free for six months. And Elon just burned a little cash. He's got cash to burn, right? And then everyone tries Grok and they're like, oh, just about as good as OpenAI and some people switch. And so you have these switching costs. So even if brand is a moat, it doesn't mean it's an unassailable moat. It doesn't mean it's just gonna allow you to rake in cash hand over fist. And so I think we don't let ourselves exaggerate the degree to which brand just like means you like conquer the world forever because we haven't seen that a lot with other companies. Right, we've seen Chanel, we've seen Vuitton and stuff. These guys have brands, right? But like, it's not all conquering. Cars, right? Like people like Toyota cars, but it's not all conquering. I agree, I agree. No, the product has to be exceptional and the price point can't be an order of magnitude higher for the same product. And I wonder if on the business, you say you never got fired for IBM and Grok is gonna have challenges on the business front side, given all the goon apocalypse or whatever. But yeah, I think people will be safer trusting, sort of OpenAI or Anthropic, but it has to be sort of cost-efficient, cost-effective for what they're going for. So I agree, it's not an unassailable mode and certainly not by itself. Got it. What do you think of the goon apocalypse? Is this a viable thing? Are people really gonna sit there like, just like falling in love with their AI girlfriends or is that just like science fiction hokey bullshit and like a few very online weirdos? I would have laughed at this or thought it was the latter, but when I saw the Reddit threads around Replica, when Replica went from NSFW to no longer that, I saw so many people complaining about, what they said was there was like, this is when my wife stopped having sex with me or something. Like, I think there's so many people who are attached to these in ways and that Replica models were terrible. Like, I think this, yeah, I'm a little bleak on this in terms of, I think both men and women are going to use this in ways that sort of supplement, you know, their existing relationships, but fill the gaps or replace them entirely. And yeah, I think that the tech is just getting so good. I think we're already seeing it a little bit, you know, the psychosis or whatever people are calling it, but I think it's only a continuation of what sort of social media was to some degree, although it's different because it's, you know, instead of humans on the internet, it's AI on the internet. But yeah, I'm bearish on the marriage rate. Let's just leave it at that. Would you also sort of express bearishness on the marriage rate? I don't know, you mean because of AI or just in general? Because of AI. Oh, I don't know. I don't know. I mean, like, why do people get married? One reason you get married is because, you know, for sex, like you have someone that you really like having sex with and want to do that, like a robot can't yet do that. We don't have like a sex bot that will, you know, keep you warm at night. Like we don't have that. But I assume that's 10 years away. I don't know. I assume it's something like- Maybe, I mean, like, no, I don't think robots are going to feel like a human for within a very long time. Like none of the materials are in place, you know? We don't have anything that moves like a human. You know, we don't have artificial muscle that makes things move like a human. Robots are still like, you know. But do you think porn has affected the marriage rate? Probably so. So like- Yeah, way better, right? Right, porn is like, you know, you're sitting there, you know, jerking off to the porn. And then that reduces your drive, right? Reduces your desire. And then makes all these people a little bit ace. You know, I was asexual for a long time. Not from that, from other things. But I think the effect can be maybe similar, at least somewhat. And yeah, so that's bad, you know, because it removes one of the basic drives of humanity. Yeah. And so I, like, I know the UK is insane, but like the porn ban is kind of good. Yeah. Like maybe we should follow in the footsteps of UK tyranny, of British tyranny, in like this one case. I didn't even see it. Yeah, it's, porn will replace relationships more than, I think, a chatbot. Right. And of course you can have AI-generated porn now. Yeah, that's right. You can have AI-generated porn with a chatbot. You know, porn can be, like, AI can be as good as OnlyFans. OnlyFans can be replaced. Yep. The only thing OnlyFans will have is this idea that there's like a real person on the other side, this sort of parasocial thing, but I think it's fake enough. I've never been to the OnlyFans website. I cannot tell you. Yep. So anyway, I'm no longer asexual, but I'm not, you know, I still have never been to OnlyFans. I just am not into it. So, but yeah, like I, so, but I assume AI can just replace it. By the way, someone said that one in 10, one in 10 women in their 20s are on OnlyFans or something like that. Is that true? I can't, there was some stat going around, but I'm just not gonna believe it. I don't know. Yeah, the, okay. Maybe they registered for an account just to have it, just to like, ooh, I registered for an OnlyFans account, but never did anything with it. I don't know. Yeah, it seems too crazy to be true. I do not like OnlyFans, the culture. You know, I've never been to the site, but I do not like its effect on the culture. So now you have subcultures, like I, you know, I've been going to a couple like anime conventions where, by the way, people no longer watch anime. Now they just dress up. It's fine. They're just costume parties now. But then, so I've gone to some of these things and then you have these weird scouts for OnlyFans, like running around these anime conventions, going to like cute girls in, you know, cat girl costumes or whatever the hell they're dressed up as. And like saying, how would you like to start your OnlyFans? And it's just a fucking pain in the ass because then you have a few people who actually are on OnlyFans taking like some sort of video of themselves for their OnlyFans. And it's just like, it cheapens the experience. It ruins the subculture. I don't like it. It's a stupid culture. It just injects, like, do we really want everything we do in real life to be like connected to some sort of weird parasocial porn? Like, I don't want that. I think that's stupid. Yeah. Agreed. And so going back to the, basically your main point is we're not going to get this, like, you know, pickety dystopia where, you know, the corporations just swallow the entire economy and it's going to be more dispersed, more distributed in terms of the profits of AI or? Maybe so. I don't know about this competitive structure, but so far I think that the actual entry we've seen indicates very little moat. We might see a moat emerge later, but we've seen so much entry. Like with search engines, we saw a flurry of like search engines, but it was very quickly, Google just won. And then we saw like Bing try to enter and just like persistently try to enter and just fail. Right? With Twitter, we saw only this one. We saw some other people trying to make competing stuff, but it basically just won. With Facebook, we saw a little bit, we saw like Orkut or whatever. We saw this bit of entry, but Facebook just ran away with it, except in countries where it's blocked like China. And so, yeah, so like, I'm seeing a lot of competition in fact, in the AI model space. Like I'm seeing tons of competition happen. And you have to sort of believe that this competition is going to taper off and that like a lot of these companies are just going to vaporize. Right? You can have, in economics modeling, you can have a duopoly that actually competes profits to zero. It's called Bertrand competition. You could have, I can envision a world in which just Anthropic and OpenAI or XAI and OpenAI, and then those two compete profits all the way to zero. Because they're just so good at competing with each other. You know, like even just two can compete. There's no reason why two is necessarily less competitive than three. In general, it tends to work out that way, but there's no inherent reason why it has to be that. And in AI, I don't see, you know, with these, you know, if there were a search engine that were as good as Google, if there are a true duopoly, I'm not sure Google would make much profit. You know, because advertisers could just arbitrage down to zero. Advertisers, they already advertise bots, right? It's bots buying ads on Google. And Google's optimized for that. And so, but then I think if Bing really were as good as Google, I think you would see a lot of Google's ad monopoly get competed away with just one competitor. And so, I don't know, you know, I don't know about this stuff. And then I think we're very used to this world in which intangible assets and human capital provide moats. Intangible asset is like a network effect or a brand or something like that provide a huge amount of value. And also there's a limited amount of top talent. So I think in the software industry, there's a limited amount of truly great coders. And you lock up the great coders, you win. And one reason why the big internet companies really won so hard is because, like, you know, because you worked in the 2010s in the startup space. And you knew how hard it was to lure people away from the, you know, talent away from the salaries that Google was offering them, right? You can just go work for Google for like effectively zero risk and get paid 500K a year or whatever. 400K a year, I don't know, whatever in the 2010s. And then, or you can go work for a startup. And like the startup can give you stock options on this startup. Like what do you, you know, yes, maybe you get some of the more like quasi-entrepreneurial risk-taking employees, but most of those people are gonna start their own company because like A16Z and Y Combinator are just throwing money at you for being an entrepreneur in the 2010s, right? Now you're part of the money cannon now. Of A16Z, you know, venture capital got very institutionalized, very big, very good at what they do. And sometimes, okay, sometimes not very good at what they do. We did have SoftBank, often very good at what they do. And so, sorry, Masa. But anyway, sorry, Tiger. But anyway, like there was all this money. And so as a startup in the 2010s, if you wanted early employees, early really cracked engineers, you had to find people who were in this weird, uncanny valley between starting their own company and working for Google, who were just entrepreneurial enough not to wanna work for Google or to be willing to work for stock options, but just, you know, not quite entrepreneurial enough to go to start their own thing. And of course, in practice, what happens is that a lot of them actually weren't in that uncanny valley after all. And a lot of them would leave to break off and form their new things, or just go back to Google within like a couple of years if it looked like the startup wasn't gonna work out, right? And you remember this, and this was really hard. It was the valley of death. And then so like, I don't, you know, locking up top talent was really important and powerful in the 2010s for these software companies. I'm not sure that's true in AI, okay? And one reason actually is because this is gonna sound bizarre. And this is something I've never heard anyone else say before. Right now, I'm gonna say it for the first time ever. AI engineering is harder. Why would that mean there's less of a moat? Okay, because which is there less of a moat for? Being a math grad student or being a Magic the Gathering champion? The answer is being a Magic the Gathering champion. It's less hard mentally, but you have to be fucking pathological to do it, okay? And the really cracked engineers of the 2010s had to know a million software shortcuts. That, they were obsessives. They were, you know, just like, that's who they were. They were incredible obsessives who knew a million shortcuts. They, of course, they were smart. You had to be smart to like write a really tight loop and write really nice code and like, but more what they did is they went without sleep in order to learn a million shortcuts and just spend their whole life on Stack Exchange. And that's why you could find them by looking at them doing open source projects for fun. That's how you really found the cracked coders in the 2010s. You looked at who spent their whole life doing open source for fun, right? And so like that whole like Stack Exchange open source network of top talent, like a lot of this is being frayed or even actively gone now, right? What do you say? Yeah, yeah, got it. That's interesting. To do AI, all you need is math. You just need to be a cracked math person. And I can find you a million of those people that are working for D.E. Shaw, all right? They're grad students in applied math at SUNY Stony Brook. I can go find you people from SUNY Stony Brook tomorrow that are as cracked as your top AI coders who all they need to do is be reoriented from thinking like, let me beat the market in high-frequency commodities trading to let me make the next frontier AI model. And they're as good as your top open AI people very quickly. Yeah, that's super interesting. It's a very interesting point. We'll continue our interview in a moment after a word from our sponsors. What does the future hold for business? Ask nine experts and you'll get 10 answers. It's a bull market. It's a bear market. Rates will rise or fall. Can someone invent a crystal ball? Until then, over 42,000 businesses have future-proofed their business with NetSuite by Oracle, the number one AI cloud ERP, bringing accounting, financial management, inventory, HR into one fluid platform. With one unified business management suite, there's one source of truth, giving you the visibility and control you need to make quick decisions. With real-time insights and forecasting, you're peering into the future with actionable data. When you're closing the books in days, not weeks, you're spending less time looking backwards and more time on what's next. If I had needed this product, it's what I'd use. Whether your company is earning millions or even hundreds of millions, NetSuite helps you respond to your immediate challenges and seize your biggest opportunities. Speaking of opportunity, download the CFO's guide to AI and machine learning for free at netsuite.com slash 102. The guide is free to you at netsuite.com slash 102. netsuite.com slash 102. Every small business owner faces the brutal trade-off. You can't spend time on financial admin or spend it growing your business. You can't do both well. That's the problem Found solves. You're probably losing hours every week keeping your books straight. That's economically inefficient, using your most valuable resource on tasks that don't move the needle. Found changes that equation. Found automates expense categorization, identifies tax write-offs automatically, and makes invoice management seamless. All that time can go toward growth. No hidden fees, no minimums, just smart banking for modern businesses. Oh, and by the way, other small businesses are loving Found too. This Found user said, Found is going to save me so much headache. It makes everything so much easier. Expenses, income, profits, taxes, invoices even. And Found has 30,000 five-star reviews just like this. Open a Found account for free at found.com slash econ102. Found is a financial technology company, not a bank. Banking services are provided by Pyrmont Bank member FDIC. Don't put this one off. Join thousands of business owners who have streamlined their finances with Found. I just want to make sure we get to the other sort of half the post, the other part of the post where you talked about just what is the risk to the economy? Like what, you know, are data centers going to crash the economy or how do we think about that element? All right, so yeah, so data centers crash the economy. We've seen this pattern happen at least twice with big capex booms in America. So let's work backward from the profit thing. If human capital and intangible assets are not the generator of profit, then what is the generator of profit? The answer is fixed costs. And that's what happened in the early industrial age. When steel and autos, blah, blah, blah, the people who would win are the people who could get more money and more financing and build more physical capital such that they just outspent their rivals, like, you know, U.S. Steel. U.S. Steel has never come up with an interesting idea in their life. They have no technology, nothing. They simply got, you know, a bunch of money and built more physical capital and sort of outspent their rivals. And they were never that profitable. They were never that great a company, but like they were big and they, you know, like they're important. They sort of won the space in steel for a while, although Nucor is now bigger, but anyway. But you had the same thing with cars. You had just like build a bunch of factories. It's a barrier to entry because like it's a barrier to entry on the fundraising side, right? Why is it hard to build a new car company now? Because it's hard to get enough money to build factories. And that's why only Elon was able to do it in America because only Elon was able to raise that much money because Elon is the best fundraiser of all time. But it's really hard to do because GM and Ford have all these factories that they built. They already have built them. And of course the regulations that make it hard to build anything don't help here, right? We block new building of stuff, but like the barrier is finance. The barrier is having some pencil neck geek behind a bank desk being willing to sign a check. Okay, the barrier is being willing to have like your little dumb ass bank algorithms say that this car company is gonna make it and we better lend them all our money. The barrier is finance. When you have, people don't understand because we haven't done manufacturing so long. People don't understand this crucial network, this nexus of physical capital and financial capital. We're used to the idea that physical capital, that capital assets and finance are completely different things. You build, we're used to guerrilla market where you guerrilla your market, market your way to the network effect, right? And then, you know, I didn't need finance. I built all this intangible brand network effect capital or even got human capital because I found some smart people who just liked me, right? Wanted to work for me because they like me and because we go on like fun trips to Thailand, I don't know. But like, anyway, so like that is, that's what we've become used to, this divorcing of bank loans and finance or bonds and financial capital from physical capital, from the kind of capital that matters for business. AI may bring this back, okay? Because if everyone's just trying to build more data centers than the next guy, everybody's a hyperscaler, right? If everybody's trying to do that, that depends on how many bank loans you can get, how many bonds, whatever, all these companies go into the private credit markets. Everybody's barring, barring, barring to try to win the space. What that means is that competition can quickly run off the rails. That's what you saw with the railroads and it's actually what you saw with the telecoms. That competition can quickly run off the rails because everyone borrows too much because everyone just has to win, win, win. And I could tell you some economic models that lead to this sort of behavior and it would be fun and five people would be interested. But the point is it can happen. And we saw it with the railroads and we saw it with the telecoms. In the railroads in the 1860s and 70s, we built so many railroads so fast. We borrowed so much money to do it. We spent 6% of GDP on railroads. 6% in comparison. Now we're spending about 1.2% of GDP on data centers, which is the biggest boom since the railroads. But the railroads were 6%, man. That was like 5X what we're now spending in terms of GDP. How do you explain that? The railroads? Why aren't we spending more then? We're about to. God. Is that a percentage? Oh yeah, that's going higher. So we're going to beat, telecoms maxed out at about 1.2% in the 90s. Interestingly, we had a second telecom boom when 5G came out, like the wireless telecom boom that equaled it, just about equaled it in terms of size, maybe 1% of GDP and never went bust. We did not have a wireless bust. What do you think our peak might be? Great question. I don't have any good way of analyzing it, so I can just make up a number, but I can tell you that if it goes on for two years at recent rates of growth, we could see 3% of GDP pretty easily. That's half railroad. Wow. And if it goes on for like six more years, we could see railroad. Wow. That railroad boom went on for a long time. But in fact, that peak wasn't reached until the 1880s, which was after the bust. The point is that initially this financing war went on too long and you had to... The finance... Finance is always easier to wrap to... Finance can always outrun revenue. Bank loans can always outrun revenue because to get revenue, you have to persuade businesses and consumers and all this stuff to... You have to persuade them to buy your thing and to spend money. And they have to get that money from somewhere, which often means getting a loan, right? And so there's all this extra stuff. With a bank loan, you can get bank loans or bonds. It's a sell in a spreadsheet. It's nothing, right? Finance is nothing. It's nothingness made flesh. And if you have people being willing to write enough checks for something, you can do infinite finance. Finance always outrun... always outrun revenue. It doesn't mean it always does because financiers are prudent and cautious. When finance becomes reckless and lavish, then finance quickly outruns revenue, and then you can have a crash even if all the stuff you're financing will be 100% viable in the future. I would note to you that almost all the houses that we built in the 2000s that were contributed to this incredible bust, almost all those houses are occupied. And I would note that the Federal Reserve bought all these bonds backed by those houses that were going bust at the time. And in 2008, these bonds were in danger of default, and it was just everything was shitty. And the Fed bought them all, and the Fed made a profit because shit bounces back. Stuff bounces back over the long term. The crash is not because we built too much stuff. It's because we built stuff with unsustainable finance at the beginning. It was because we needed this stuff, but finance outran it. Essentially, we paid too high interest rates for it. And so that's what happened. Basically, what happens is that finance outruns real revenue even if revenue catches up. Revenue is the tortoise and finance is the hare. And so the tortoise will win in the long run, but the hare… Actually, it's funny. You can actually run this race with a real tortoise and hare, and the tortoise actually does. It happens just like in the fable. The rabbit will just run halfway through and then just stop because that's how rabbits graze. When they graze, they just run to the middle of the meadow and then stop. And then the tortoise just… Metaphorically, it's life. Yes. Metaphorically, it's life. Revenue is the tortoise and will win. This is why you have a Gartner Hype Cycle. The Gartner Hype Cycle is just finance. Finance is the crash. The crash is not the actual usefulness of the thing. There was never a crash in terms of the rate at which we were finding new uses for the internet. It's not like in 2002, people stopped finding cool things for the internet to do or stopped building out e-commerce. Nobody stopped building out e-commerce. It's that we finance stuff faster than the actual revenue grew, and there was a crash. The crash is all just finance. All the hype and then the plateau of whatever is just when interest rates get to a reasonable level that reflect actual revenue growth rates, and then you're done. Then you're in corporate finance land, and then everybody's happy. Then basically, if we follow the pattern of the past, finance will outrun revenue, and there will be a crash. As for why that's true, I will write a third post about that detailing some of the deep economic theories, and only five people will read it because no one wants to read some post about Bayesian updating. What would the impact of that crash be? 2001 wasn't close to 2008, right? There are degrees to sort of… The answer is not like 2008 because nothing involves as much debt as real estate, but 2001, easily. I would place bets at even odds on something like 2001 happening. It wasn't that bad, honestly. Like 2001, telecoms went bust, dotcoms went bust, and it didn't really slow our growth rate, and it didn't really slow down technology. Ultimately, it wasn't that big of a deal. If we have frequent recessions of that size, once a decade, our economy is perfect. Our economy is great. Who cares? Some people argue that it was bad because we replaced the telecom and dotcom bubble with a real estate bubble. There was some sort of portfolio reallocation effect where capital flowed into housing because it flowed out of the internet, and that laid the groundwork for 2008 and sort of set the stage for 2008, which is much worse. There is that argument. Something like that could happen. We could have another real estate bust based on people stampeding out of AI. I don't know. We'll recognize that if it happens. We remember 2008. But anyway, I guess I do expect a crash. I expect a crash about the size of 2001, and there'll be a year when everyone says AI is toast. Then there's no future in it, and at that point, bye, bye, bye, bye, bye. That's called by the dip, right? Because AI is going to win. It's going to happen. That's my baseline. It could be worse. It could be that we get to 3% of GDP and we have a bigger crash. I'm worried about the fact that the increased level of debt because data centers are big and they take debt. When we finance, the dotcom is mostly equity, and telecoms finance themselves with bonds. I will be worried the day that either bank loans or shadow bank loans become the main source of financing of data centers. Bank loans are only... I covered this in my post. Bank loans are slightly involved, but private credit is getting very involved. If private credit starts acting like a shadow banking system and private credit is implicitly backed by the big banks, then we have a problem. Regulators, to the extent that we even have regulators anymore in the Trump administration, regulators need to keep an eye on this. Someone needs to keep an eye on those private credit guys. In fact, encouragingly, I've seen a shift where Meta and some of these other hyperscalers are switching to issuing their own bonds because they could issue their own bonds. What you had six months ago was what you had... There's a feedback, sorry. Hold on a second. Echo. Do you still hear it? Yeah. Oh, it's on your end now. Okay. I'm good on Echo. It's gone. Anyway, so what you had six months ago maybe was companies like Meta that absolutely can issue their own investment-grade bonds. Meta can issue... I don't remember what their credit rating is, AA or something like that, but they can issue really good bonds. Yeah, AA3, okay? That's great. Meta is a blue chip borrower. And they can issue their own bonds, but they weren't. They were going to this opaque-ass private credit market. And Paul Kedrowski had an amazing blog post about this where he's like, the reason they're doing this is to hide it because people will think it's not justified. If they issued their own bonds, it would impact their stock price. They're hoping to preserve their stock price by borrowing from private credit over the counter secretly so that no one knows how much debt they're loading themselves up with. And I buy that. But encouragingly, you're now starting to see a shift toward companies like Meta issuing their own bonds as they should have done from day one. If the data center boom is mostly financed with bonds, I'm not so worried because those bonds are mostly not going to be held by investors through funds, not by banks. Banks will hold a bit of that, but they won't hold a ton of that. But if private credit is funded through the banks, your average money management fund or index fund or whatever can't invest in private credit. It's private. It's over the counter, so banks have to do it. In practice, insurance companies are the ones investing in this private credit and sometimes lending to private credit also, a chain of lending. So I could foresee if they kept going to the private credit markets for AI data centers, I could foresee a systemic risk where private credit markets become our new shadow banking system, just like the sort of SPVs and CDOs and all the crap in the 2007, right? Become this shadow banking system. And then that shadow banking system collapsed. And we find out after the fact, oops, that the shadow banking system is back to the real banking system. And now our real banking system has collapsed. Oops. And time for giant bailouts. Everybody gets mad. Real economic activity slows down. People default on their house. I don't know, whatever, like shit hits the fan. But as long as we keep an eye on those private credit markets and make sure that our banking system isn't exposed, and we make sure that like market investors through markets are taking all this, a lot of this risk, I'm not so worried about a big systemic crisis. Does that make any sense? That makes sense. Yeah. I think that's a good place. I want to ask you, so now I can actually, you're with A16Z now, how do you think A16Z is thinking about AI financing in terms of some of the things we've talked about? Yeah. I mean, all we think about is mostly just like venture investing in these companies. I think this isn't directly answering your question, but A16Z invests in the most AI companies of any VC fund. And what we shared at our sort of private GP events or to our LPs is that we've been too conservative. Basically every company in every category, sorry, every category has had major winners and key point major, not just one, but like on the foundation model side, we didn't invest in Anthropic because we were like, oh, open AI is the category leader. We tend to invest in category leaders, but even Anthropic has absolutely crushed it. And then we didn't make the same mistake again and invested in other model providers. And similarly for on the application side or other ones too, there's multiple winners. We tend to think that the markets are just so much bigger. And so even if there's no traditional moats protecting from other entrants, that there can be multiple winners in a category. So yeah, we think, I'm not sure if that's answering your question, but I think the punchline is we've been super active and we feel internally that we should have been even more, more active. Right. Who do you think's been like the VC that got on the AI train most aggressively and most effectively from the beginning? Like who really is just like nailed that? I mean, I understand it's a lot of luck, but like. Yeah. So I do think we've done exceptionally well, but I think Daniel Gross and Nat Friedman have absolutely crushed it. And they've taken themselves out of the game a little bit by going to meta, but out of the investing game, but they're in a new game, which is very exciting. Yeah. I think, I think they, they had this sort of like YC for AI called AI grants, which had a bunch of really great companies early, but yeah, they'd be my pick. At some point we need to tell the story of the group house called the archive. Oh yeah, totally. Because the archive gave birth to chat, GBT, blue sky, and a number of important AI things and anthropic. Yeah. That, that one group house where I used to hang out and party has just produced storied, you know, that, that was amazing. And future politicians and our, and our buddy, Michael, Michael, that's right. Yeah. Karaoke, karaoke legend who now is working for anthropic, of course. Yes, yes, yes. Yeah. Fascinating. Cool. Well, let's let's wrap on that. There's been a great discussion on AI labor, you know, where will the profits go and you know, what a potential impact to the economy or crash could, could look like Noah is always until next time. Until next time. Econ 102 is a podcast from Turpentine, the network behind moment of Zen in the arena, the cognitive revolution, and more. If you like what you hear, subscribe and leave us the review in the app store. You can keep up with both of our sub stacks for written analysis of the topics we cover in the show at no opinion.substack.com and eric torrenberg.substack.com. Thanks for listening. That's a wrap on today's econ 102 episode. If you found value in our conversation, definitely take a moment to read the show notes. They're organized with all the key insights and meaningful takeaways. Those notes were created automatically by notions, AI meeting notes. It captures every detail and breaks down complex concepts. So no idea gets lost. And because AI meeting notes lives right in notion, everything you capture, whether that's meetings, podcasts, interviews, conversations live exactly where you plan, build and get things done. No switching, no slowdown. Check out notions AI meeting notes. If you want perfect notes that write themselves, head to the link in our show notes to try notions AI meeting notes free for 30 days.