Benedict Evans: OpenAI’s Moat Problem & the Future of Software

Thank you.

I suppose there's two ways we could talk about this. One of them is to talk about all the stuff that OpenAI has been trying to do. I think the other is to talk about the problem. And the problem is that, as far as we can see, there is no winner-takes-all effect or network effect in a foundation model. So there's nothing that you can do that means other people can't make a model as good as yours, which is what we're used to from the software industry.

What we're kind of used to in tech is that software, by definition, has no capital or very little capital, but it has network effects. And so it tends to produce monopolies or near monopolies, and that produces high margins. And so that's what happened with Windows and macOS and Google and iOS—not macOS—and Google and so on. But the question with LLMs is, like, they're very expensive and very hard.

And so you can kind of fail to get onto the ladder or fall off the ladder, for Microsoft on the one hand and Meta, for the meantime, on the other hand. But there's no lever that you can decide you're going to pull whereby you'll just pull ahead of everybody else and they won't be able to catch up, which is unlike Google versus Bing. It doesn't matter how much money and how hard Microsoft works, Bing will never catch up with Google.

So that means that we've got, like, pick a number between three and six or maybe more organizations that can make a frontier model, and they keep leapfrogging each other every couple of weeks or every couple of months. So that's one problem.

First of all, I'd push back slightly on that comparison and then answer the question. The first is, if you actually look at the market shares, Google Cloud, Azure, and AWS are actually in quite different businesses. AWS is mostly infrastructure. Microsoft is mostly services. Google is mostly scrambling to catch up in a very distant third place, despite having basically invented cloud. So they're kind of not all the same, I think. But the challenge, I suppose, is where do we get to equilibrium on the foundation model itself?

Is it that these things will be able to create ecosystems above themselves? So it will be like iOS and Android. And so they will create value, they will create value capture, they'll build tools and all sorts of stuff on top of the raw foundation model. And so you'll have to choose, right? Or our company is going to standardize on ChatGPT, which isn't what happened in cloud. As a business, if you go and buy a SaaS app, you don't think, well, we're going to log into it with our AWS account.

Certainly as a consumer, I can't remember which cloud Snap uses, and I don't care. And if I install Uber, which cloud does Uber use? Who cares? So it may be that the models are able to kind of build something that looks more kind of like Windows, which is what Sam Altman talked about a lot at the end of last year. Or it may be that they're basically commodity infrastructure, and they're sold at marginal cost. And maybe you barely make a return on the investment, or maybe you don't make a return on the investment itself.

And what you're doing is making a return on everything you built, that you do with this on top of it. And that's where these companies do get different, because if you are Meta or Google, you've got this whole other highly profitable business, which now needs to have LLMs inside it, powering all sorts of capabilities and features. And you probably want them to be your LLMs rather than somebody else's. But you may not necessarily need to make money from the LLM by itself any more than Meta.

I mean, obviously Meta has a cloud, but Meta doesn't even try and doesn't sell the cloud. They just use it to power their own stuff. So those are the sort of the two possible outcomes. It may be that this stuff ends up as commodity infrastructure that's sold at zero margin. It may be that, no, there's two of these or three of these, and they capture a lot of value. There's a sort of price-equilibrium question of, do you get pricing discipline?

Do you have a small number of companies and they hold the prices high or not? As I wrote about this last month, ask your favorite economist. There are words, there are terms for these kinds of questions, named after economists who first asked the question generically. But I think that kind of going back to AGI and OpenAI, the problem at the moment is that the raw chatbot by itself isn't a great product. And most people struggle to work out what to do with it.

You've got a small number of people, mostly in tech and also outside tech, who are very self-optimizing and have certain kinds of jobs that map very well to the kinds of stuff that LLMs are very good at doing. If you look at the usage data, something like 10% of the population is using these things every day, but another 50% are using it every week or every month. So most people who have a ChatGPT account can't think of anything to do with it today.

And so you can ask, well, is that because the models have to get better? And we've had this conversation before. Is this because the models have to get better, or is it because habits have to change? Or at the extreme, could you say, well, people only use Google once a month, once a week too, which doesn't seem like a good answer. Or do you have to build a whole bunch of stuff on top, which is back to this commodity infrastructure point?

And so it seems like a lot of what the question now is: are we all just going to use ChatGPT or Anthropic or Claude, or is that going to be like an invisible API call buried underneath that's used by some other thing, and it's the other thing that makes all the money? The really kind of interesting comparison here would be TSMC, because what happened in chips was that there's no network effect, but with each generation, it got more difficult and more expensive, and the numbers dropped with each generation.

And now there's basically one company at the frontier, and two or three that are a generation behind, and then five or ten that are a generation behind that. And TSMC makes lots of money, but they don't capture all the value of the tech industry. As a consumer, you don't know that TSMC makes a chip in your iPhone, and they don't get a cut of every App Store sale. So that's sort of the puzzle. This great strategic problem for OpenAI is you've basically got commodity technology.

You don't have a network effect or winner-takes-all effect. You've got 900 million weekly active users, but most of them are not using it every day and can't think of anything to do with it, and only 5% of them are paying for it. So the usage is like a mile wide but an inch deep. And so you've kind of got to swap the mindshare and the momentum that you have for something more durable. So there's a story yesterday that they're trying to do deals with private equity firms to get into the private equity firms' businesses.

They did this video-sharing app at the end of last year that didn't work. And there's an app store. Oh no, there's another app store. No, are we on the second or the third app store now? I've kind of lost count. And there's an e-commerce integration. But guess what? E-commerce integrations are really, really hard and complicated because there's millions of merchants and billions of SKUs. And Google and Meta have failed at that twice. So it turns out that OpenAI went, eh, no, we probably can't build that ourselves right now.

So there's this sort of challenge of how do you get from having the mindshare and one of the good models to having some kind of a durable platform business where you've got developers, users, corporate accounts, something that's locked in that doesn't just depend on you having the best model next week and the week after that and the week after that.

Is there something I wrote like a year ago, I think, where I said, what do you mean when you—something like, what is a better model? Which sounds like a crazy thing to say. But I always remember years ago, when iPhones were the hot new thing, one of the New York late-night TV shows, they would go out onto the street and they'd show people last year's phone and say, "It's a new phone. What do you think?" And they'd go, "Wow, it's amazing!"

Because you couldn't really tell, actually. And if you're doing very specific, very hardcore stuff, and you're pushing these models to the limit, then you'll be able to see, oh, it didn't do that, and now it can do that. With most stuff, the models are jagged in lots of different ways. You don't really know whether it will be able to do that or not, or how well it will be able to do that. And most of the stuff that you try, it can sort of do to varying degrees.

And now it can sort of do it slightly better, or maybe slightly less sort of. But if you've got a bunch of use cases where you need the right answer, as opposed to sort of the right answer, then saying that the model is better doesn't mean anything. I mean, literally, it is literally meaningless. Because what you're telling me is, I asked the model to compile 50 things, and last month it would get 10 of them wrong, so I'd have to check all 50.

And now it will get eight of them wrong, or 12, but probably eight. So again, I'll have to check all 50. So actually nothing's changed. And the point that things change is not when it goes from 90% right to 91% right to 95% right. The point is, it's either right or not right. And if you don't have that, then you've got to kind of sit and think much harder about what you do with this thing and how you use it and how you build tooling and software around it.

Part of the challenge here, I think, is, which actually we could kind of segue to talk about, the SaaS apocalypse. A lot of the challenge is working out how to ask for what you want. And so, for the sake of argument, even if the model had general AI and was actually really superhuman—not superhuman the way that branding people talk about it—it was actually really AGI in all senses, it's still quite hard for you to describe what you want.

Well, so, two or three things. Firstly, I think very obviously it only works if you're using it a lot already. And so if you haven't worked out a bunch of stuff to do with this, and most people have not, then you don't see the memory. OpenAI did this data marketing at the end of last year where they gave everyone a cute little graphic of how many messages they'd done and told everyone what percentile they were in. And so I went on Reddit and grabbed like 300 screenshots of people posting these.

And it turns out that if you did 1,000 prompts last year, you're in the top 20%. So basically 80% of people hit return less than 1,000 times last year. So an average of less than three times a day. Obviously ChatGPT's growing up in the year, so you can't really average it across the year. But if you typed return 1,000 times in the whole year, you're not really using this every day. And so there's no memory there.

Then a subsidiary point would be, okay, is that a network effect or is that stickiness? It's probably more stickiness. And how sticky? And what happens if you just ask ChatGPT to tell you everything it knows about you and then paste that into Claude, or vice versa? So it was kind of unclear whether that was a thing. I mean, the whole thing does feel very sort of 1997 in a lot of ways, in that it's clear this is a huge deal and it's already starting to work, but that doesn't mean you know how anything's going to turn out.

And one thing that occurred to me looking at ChatGPT as a product is it's kind of like trying to differentiate the web browser itself. You've got an input box and output box, and how can you make them different if the whole point is that you can type in anything and get anything out? You can make the underlying rendering engine better, like Chrome has a better rendering engine, fine. And you can make the underlying LLM better, but the actual product that's presented to the user, it's an input box and an output box and a couple of buttons around the edges.

I mean, it's the same with the logos. You've seen the joke that all the chatbot logos look like buttholes.

There are all these little spots. But again, what are you supposed to show? Because how could you—is it even possible in principle to make the UI of a chatbot different? Because isn't the whole point that it's completely universal and there's no UI? So you've got this kind of, back to the thesis, seemingly OpenAI sort of has this problem that, first of all, the model itself is the same as everybody else's, more or less. And if you're, again, if you're the kind of person that watches this podcast, you're probably thinking, no, no, no, it's really different.

But, like, for somebody who's only using this once a week, they really don't see the differences. And then it's not clear what you're supposed to build on top of it. And is that going to be you? And even if you worked it out, why wouldn't everybody else build it too?

Well, this is so—Kevin Weil and Mike Krieger both made this point on stage, like, 18 months ago or something. And then Fiji Simo made it in, like, the second hour of a podcast I was listening to at the gym. You know the joke that, like, anything—the most secure encryption on Earth is anything you say in the second hour of your podcast, because no one will ever hear it? Well, I actually listened all the way through. And she said, which is also what Mike and Kevin had said, is the way it works is you turn on your computer and you look at your phone in the morning, you've got an email from the research group that says, "Hey, guess what?"

We've got this cool thing. And then your job is to go and do something with it. So it's like, you get the email and then we've got a new voice model. It's like, okay, well, I guess we're adding a microphone button today. That's not—

You don't control the product strategy, which is, of course, how science works. But you don't know what's going to happen. You don't know what's going to get built. Obviously, you've got Sam and Dario and so on setting the fundamental research strategy. But the thing comes back in six months, and it may work, it may not. You don't know what's going to work. You don't know if that will happen. So, fairly or unfairly, I opened the last essay I wrote taking that quote from Fiji Simo and then comparing it with Steve Jobs famously saying, "You can't start with the technology and work forward to the product."

You've got to start with the user experience and work backwards to the technology. And this is sort of, it's not like anybody doesn't know this, but that's sort of inherent in where we are, is that this technology is continually shifting and changing and evolving, and you don't know what it'll be able to do next week or next month. So you kind of don't know what it is that you're trying to build. You're a product taker, not a strategy setter.

Well, so you eventually get to a point that people have actually built a different product, as opposed to an input box and an output box to a model. It's not at all clear to me that building your own agents is a consumer-facing thing. It seems to me either you shouldn't have to know what it's doing. All that should be submerged inside some product. Now, I think Stripe did, like, they copied autonomous cars and did level 0 to level 5 agents.

And so level 5 is that the agent knows that you're out of dog food and buys you more dog food, and you don't even know. And it just kind of appears, like fully autonomous ordering and fully autonomous purchasing, fully autonomously solving the problem. And where level 0 is like, here's a picture of a coat, work out what it is. And the agent goes and your model goes and calls a bunch of agents to solve that question for you. You don't have to know that that's what it's doing.

And so there's certainly sort of a spectrum of what do you mean by agent? Do you just mean that the LLM can call different tools, but you don't know about it? Or do you mean you can actually get the LLM to go and do a thing for you, with or without multiple tools? It's a slightly fuzzy term. It's a little bit like saying metaverse. You didn't really know what somebody meant when they might have meant VR. Well, VR is real.

They might have meant games. Games are real. But when they said metaverse, you didn't really know what it was they were talking about. And it's kind of the same with agents. It kind of gets again to this sort of—it almost kind of deepens the problem in that the more the capabilities expand, the more jagged the thing is. And so the harder it is to know whether it will be able to do X or Y, the harder it is for you to kind of mentally map: is this the kind of thing that it would or would not be able to do?

So, for example, these models struggle massively to read PDFs, but that's not a thing that you would know from looking at it, that it will be completely unable to read a PDF. Because why? Why? There's not some intuitive reason why you could kind of deduce, well, of course it won't be able to do that. And some of this is just familiarity. It took a while to work out, you can't—you could ask Google that, it's not going to be able to answer this.

But it is this sort of puzzle of how do you build product around representing that? At what point does your model of what it does—the frontier—get smooth enough, or your mental model map closely enough to the jaggedness of the frontier, that you can kind of understand what it could or couldn't do? And at what point do you have to map—no, you actually have to sit and map that to a particular use case. But there's another way of thinking about this, which is to say, like, with every new thing, you start by getting it to do the stuff you're already used to.

And it kind of takes time to build new things that are native to the new technology. And you can probably split that apart further, just because obviously every slide needs to have three bullet points, and say, like, there's sort of a third step, which is at what point can you kind of pull the whole thing inside out and do something that just doesn't bear any kind of connection to the way that you might have thought about the old thing. The sort of obvious example would be the progression from, okay, Flickr having a mobile app to Instagram to Snap or TikTok, because Instagram is taking the desktop experience.

Now, indeed, arguably still, Instagram is still kind of taking the desktop experience, except that it adds filters, which no one uses anymore. But Snap says, but this has a camera, so why aren't we starting with the camera? And then TikTok says, and it's a social network, so why aren't we starting with that? And so you're sort of pulling the whole thing out. And TikTok isn't like a mobile version of Flickr or indeed a mobile version of YouTube anymore. It's something else.

And we're still at the stage of taking a PDF of your catalog and putting it in your company website.

As we try and work out what we should do with AI. And of course, that was a really big deal. And it's still a really big deal that lots of companies where having a PDF of your catalog on the website is really good, like 30 years later.

But the jaggedness of what these things can do is almost kind of mirrored by the jaggedness of where this is important in different industries and for different kinds of jobs, like which jobs get affected more or less, which industries get affected more or less.

The problem is that's self-selected behavior. That's people—what can people think of? Whereas, it's again, a Steve Jobs quote: what you want is to say, "Aha, I've realized a thing that you could do with this that hasn't occurred to anybody." Because that's where you create billion-dollar companies, trillion-dollar companies. You change the question and you realize that maybe I'm not going to do it like that. I've realized I can solve that. Here's this problem. You didn't really realize that problem existed.

And I'm going to go and work out a way of solving it in some way that doesn't even look like the problem either. And then that's how you really change things. Yes. Like, step one, OpenAI can look at all the desire paths and they can pave the desire paths. The problem there, of course, is that they also have to map that against trying to make money. So I think when you look at the usage data they released last autumn, very little of it was actually e-commerce.

There's loads more people using it for porn, but you can't make money from that. You can make money from e-commerce. So they tried to do e-commerce and advertising. Also, they hired a bunch of e-commerce and advertising people from Meta. So that's what they know how to do. I mean, this comes back to my point: is this commodity infrastructure? Which is—and it's also, this is a quote from Craig Federighi at Apple. So people having a go at Apple that you haven't got any LLMs.

And Craig says, "Well, we don't have Uber or YouTube either." He didn't actually call it that, but obviously—

We don't have a video-sharing site. We don't have a taxi service. We provide the platform for other people to do that. And you can't, as a platform, you can't invent all of those things. Apple and Google could not invent everything that was done on the iPhone and Android, and Microsoft couldn't invent everything that was done on the PC. The trap in that, the counter to that, of course, is that you can also say that about Microsoft on the web.

Like, Microsoft didn't invent any of the stuff that we did on the web. I mean, that's not quite true. They invented Expedia, and they tried to do all sorts of stuff. But in the end, all this interesting stuff on the web wasn't just—you still did it on a Windows PC, but it was done with other people's tools, other people's software, other people's capabilities. And the PC, in effect, was commodity infrastructure that you used to access the web. And it became a Chromebook, which is actually what a Mac is now, mostly, for most people.

So you've got this on the one hand. On the other hand, can you possibly create all of the billions of use cases that people are going to come up with? There's still millions of use cases that people are going to come up with this stuff. No, but you want to be the place that you want them to build it for your thing and not be commodity infrastructure for everyone else. You want to be kind of somewhere in between those two.

The puzzle is that, this is my point, is you don't sign into your SaaS app with your corporate AWS account. You sign into it with Okta or something. But no consumer has an AWS account. No company has an AWS— that's not the right level of abstraction or aggregation. It's just infrastructure. And it does feel at the moment like the LLM is infrastructure. And another historical analogy is, I mean, I'm just old enough to remember the early '80s with PCs.

People said that what you should do is you should buy a PC and then you should buy a database program, or you should buy a software development program and you should make your own software. So you're a retailer and you want some inventory software, you should buy a PC and you should buy C++ or whatever it would have been, Turbo Pascal. I don't know. Whatever the programming stuff would have been, and you should program your own inventory management software. Or you should buy, like, a database program and you should make your own inventory management software in the database program.

And neither of those were the right way of thinking about this. In fact, I found this wonderful quote from The New York Times from 1980 talking about VisiCalc and spreadsheets. And this guy says—of course it was a guy who says—"Well, I used to code DCF models in machine code and it would take me 20 hours. And now with VisiCalc, I can do it in 15 minutes." So tell that next time you hear a software developer saying, like, AI is a completely different thing and nobody has ever abstracted software like this before.

Like, yeah, we've been doing this for 30 years, 50 years. So this is this sort of question: where do you create the applications and the use cases? Who does it? How close does that go to the raw chatbot? How far up the stack does the chatbot do it? So now, it was ChatGPT apps, which failed like two or three times. And of course, it's OpenClaw skills as well, or whatever OpenClaw—yeah.

Not really. I mean, this week they've got all the fire, they've got all the juice, whatever the word is. I don't know. This week, this week; next week, next week it'll be something else. OpenClaw was kind of interesting to me because it looks a lot kind of like desktop Linux in that suddenly you've been having to use this monolithic thing from these big companies where you couldn't really get your hands on the metal. You were just like, you had to get your API key and you type stuff in and they tell you what you could do.

And now you can, like, build it yourself with your own hands. It's like the Homebrew Computer Club in the '70s, and you can make it yourself. And then the other side of this, of course, is it just needs to be a little bit more polished, and then it'll be ready for everybody, which, of course, Linux people have been saying for 30 years.

And, like, they don't understand, like, the whole culture of it is just wrong to do that. And you ask it to tidy up. The other interesting thing about OpenClaw is it shows you why Google and Apple haven't shipped this. Like, "Tidy up my inbox." "Okay, I deleted all your messages."

Yeah. So there was this huge craze around OpenClaw in China, which obviously, I don't speak Chinese, is interesting in its own right, but it gets to this sense of how much pent-up enthusiasm there is and how many possibilities there are, but also paradoxically how hard it is actually to make a real thing out of this. This is kind of the funny thing about AI assistants: this is really a great use case, except it's much, much, much harder to actually do it and be sure that it won't delete your inbox.

It's funny, this is like, it's almost like a straw man, except there are presumably real, rational, sentient people who actually say this stuff. I don't think it's particularly interesting to explain why people aren't going to code their own Stripe. I think what's more interesting is to kind of think about a broader taxonomy of software, which is that you have the big-iron enterprise ERPs and what they call systems of record. So you need to store the same data in the same way for 50,000 people and a billion transactions a year.

No, you can't have David Brent from The Office making his own. But then you have tasks and workflows where that single system, call it SAP, is too inflexible. And so you break it out, unbundle it into something dedicated and vertical, you get vertical tools. And so this is why the typical big U.S. company today has, depending on your numbers, like 400 to 500 vertical SaaS apps.

Yeah. Frame.io is unbundling Google Sheets and Dropbox. Like, theoretically, that's what people were doing. They were managing video workflows in Google Sheets and Dropbox and email. And it says, no, we're going to turn this into a dedicated tool. And I don't know, theoretically you could have done that in Oracle, but, like, no. And so you've got the vertical tool that's specialized, and you've got the horizontal general-purpose tool, and then you've got this middle case of using Google Sheets and email or using Excel or exporting a CSV or something where you could almost kind of call it improvised software, where it's not a process that's being done over and over again in the same way every time, and has to be recorded and tracked and has to have compliance and everything else.

And it's not kind of enough of it or big enough, or no one's realized it's big enough to make a dedicated tool. And so you're doing it in Excel or Tableau or Google Sheets or email or something. And now on the one side, AI coding means coding is way cheaper and easier, and also means that there's a whole bunch of stuff that you couldn't do in software that now you can. And so there will be way more software, and that will pick up many more of those use cases either that weren't automated before, either because they were too small, or because you actually couldn't automate that thing with software before.

And now with AI, you can.

Well, this was going to be the other half of it, is now you expand this category of improvised software. And so where something, it might have been that you do it, get the CSV, and do it in Excel, or maybe a Perl script if you're that kind of person. Now maybe you'll ask ChatGPT to do it for you. Analyze this for me, do this for me. And often it's like stuff that you couldn't have done in Excel before. It's like, look at all those PDFs, look at all those PowerPoints, and then look at this PowerPoint.

Is there something in all of those PowerPoints that we haven't said in this one? Which is the kind of thing that's got these sort of fuzzy answers, that it's kind of right, that works very well with an LLM, except it wouldn't work if there were PDFs because it can't read PDFs. So you have to tell people it can't read PDFs. So that sense of, like, this sort of middle ground, and maybe my taxonomy is wrong, but I think that's more interesting as a way of looking than just saying, oh, everyone will just vibe-code their own thing.

But they may ask Anthropic or Gemini or ChatGPT, can you do this thing for me? In a way that's sort of analogous to they might have done it in Excel.

Way more software.

Again, unpick that, because on one level, okay, if it's much cheaper and easier to write software, there'll be more software. Who will be doing that is a combination of who understands how software works and who has really thought carefully and understood the problem. io was not made by some guy or some woman working at a video production suite somewhere in Soho, because that's not how they think. They're not software people. They're also not product people. There's a different skill to actually work out what should the code be doing, and what is the problem, and how would you work this out.

So there will be way more software, both stuff that doesn't need AI except to create it and stuff that needs to use AI to do the new thing. Some incumbent software will go away, like big expert systems where an LLM can do that better. It's easy, just... But there's also, this seems to me like an analogy of SaaS, or a repeat of SaaS, in that SaaS meant that you had an order of magnitude, probably two orders of magnitude, more software. And some incumbents got completely screwed by this.

And loads of tasks that you couldn't have automated before got automated because SaaS made it much cheaper and much easier to go to market, and easier to unlock those problems. And so, as I said, when we went from mainframes, like, the big company had, what, like five pieces of software? And with on-prem, you've got dozens of pieces of software. And with cloud, you have hundreds of pieces of software. And so you should just sort of presume that with this, you'll have way more software.

You'll also have way more stuff being automated where you don't need a dedicated tool, where you don't need to really shut your eyes and work out every step of the workflow and the process and how this should work and what you should do with it. It is just like a one-off improvised thing where you'll get the model to do it. So you'll have way more stuff being done in software. I mean, this is, people last year, suddenly everyone discovered, looked up the Jevons paradox in Wikipedia.

Yeah, funnily enough, I actually kind of did know about this, but more just because I'm sort of interested in industrial history or something. I could vaguely remember hearing about it, but I looked at this again and I thought, but this is just price elasticity. That's all you're really talking about. If you make it cheaper and easier to do something, you might do the same thing for less money, or you might do more for the same amount of money, or you might do more with more money if you have a completely different ROI, which is exactly what you see in financial services.

Spreadsheets did not result in a collapse in the number of people working in finance. Quite the opposite. You have way more people in finance because now it's possible to do all this new stuff that you couldn't have done before. I mean, back when, if it took you a week to do a DCF, how many DCFs did you do? That was it, you were done. If it takes you 10 minutes to do 20 DCFs, then you do way more DCFs.

And so then, of course, it's kind of my point about the jaggedness of adoption versus the jaggedness of what the models do. There's a sort of almost like a cliché that people always talk about, the Marc Andreessen "software is eating the world" thing of Uber and Airbnb. That Uber didn't sell software to taxi companies. Airbnb doesn't sell software to hotels. They change what you mean when you say hotel. Fine. Take a look at the market share.

And so many cities, Uber basically demolished the taxi business and also unlocked huge new demand. So taxis, Uber in New York, I think Uber rides per day is like double or triple what yellow cabs were. And yellow cabs are down by three quarters or something, like ballpark numbers. Look at hotels. Okay, hotels are maybe, hotels are still growing. Maybe they grew a bit slower. Maybe not, maybe, maybe not. And Airbnb was mostly additive. And you dig into that and the answer is, why?

Somebody, I remember, saw somebody on social media saying, "The problem with Benedict is he always says the answer is, it depends." It's like, well done. Thank you for paying attention. Yes, it depends. Well, is software going to completely change hotels and taxis? Yes. How much? Well, those are completely different things.

So, like, my fiancée goes on a business trip to some Midwest American city, and she lands at nine o'clock at night, and she's got a client meeting at eight o'clock the next morning, and she wants a gym, and room service, and a fridge, and a bath. And she's not going to go and stay in an Airbnb. Absolutely zero, negative possibility that she's going to go and stay in an Airbnb. And travel is—business travel is half the travel business, hotel business.

You can proliferate examples as often as you like. Why did the internet have a bigger impact on selling consumer electronics than selling high fashion? Well, it depends. So this is the problem I have with people trying to score professions' AI exposure, because you're kind of directionally right, probably. Yes, it feels intuitively like that profession is more exposed to AI than this profession.

All of that kind of stuff. It's sort of probably directionally right. But it's directionally right in the same way that an analysis you did in 1997 about the internet would have been directionally right. I mean, most of it would probably be more or less true. Five, and that one is 78. It's just ludicrous. But you would not have got Uber from that analysis. You would have said, well, taxi drivers, well, how would the internet change that? Obviously, it won't touch that one at all.

Well, there's even, again, in hindsight, but at the time, newspapers looked at this and thought, well, this is going to be great. I mean, remember the AOL-Time Warner deal? Why did AOL buy Time Warner? Like, a whole bunch of magazines that don't exist anymore. I mean, Warner too, but, like, people didn't really understand what it was that the internet would do to the media business. And then within that, clearly the impact on, like, regional newspapers was completely different to the impact on Disney.

Disney's fine. Regional newspapers disappeared. And so you go right in, and in hindsight you can say, well, what really happened was you unbundled physical assets from the underlying product. And if your defensibility was based on owning a physical asset and that physical asset suddenly stopped mattering, then your whole business model has just exploded. And you could apply that lens to Airbnb and also to Uber. But then, of course, those two turned out very differently. I think the other side to this, I mean, you pull the Uber example through to today, you can do that evaluation and say, well, like, fitness instructors are fine.

Okay. Have you seen—so I put my phone in my room and point it at me and turn on the AI with the camera. Why do I need, like—I don't know. Maybe that won't work. Maybe it will. But you can't know those things at that level of granularity. I think that's kind of the problem. What you can do is just kind of point to thought experiments, I suppose, and say, well, here is a way of—here's a test you can apply.

To look at this field and ask yourself, well, does that field have a question? Does that field have a problem?

Well, will a lot of these companies go bust? Yes. I think I may have mentioned last time, there's this classic book on the history of bubbles, and the title is This Time Is Different, which has a double meaning because, A, in a bubble, people say, "No, it's different. It's not a bubble," and they're wrong. But also, when they say it's different, they're kind of right. The dot-com bubble was different from every previous bubble, and what's going on now clearly is different from the dot-com bubble.

We don't have loads of IPOs of consumer companies with no profits. We don't have any IPOs. It's not being driven by retail speculation in public market stocks. It's not being funded by venture capital, apart from anything else. But it can still be a bubble, which is an uninteresting observation. I think, do we have a lot of overinvestment? Yes, of course. Will a bunch of this end up not producing a return? Yes, of course. That's how this stuff works.

Where were the smoking holes in the ground? A bit more difficult to tell. I mean, there are people who, obviously, you can kind of look at the NeoClouds or look at Oracle and say leverage doesn't tend to end well. But hey, I haven't been an equity analyst for 20 years. I don't know. No investment advice here. But just deterministically, you should go back and look at the internet and look at mobile and make a list of all the stuff that didn't work.

All the stuff that was a big deal and really exciting and cool and interesting, all the acronyms and companies, concept ideas that didn't work. Well, of course, a bunch of stuff that people are working on now is going to not work. That's just how the world works. That's how innovation works. There'll be a whole bunch of creative creation, and a lot of it won't end up being the thing. I think Hunter Walk once said that money in Silicon Valley is like one of those little rubber balls, bouncing balls.

Like, you throw it into the room and there'll be a bunch of people who get rich because the ball hit them at the right time. It's like the person who joined WhatsApp two days before Meta bought it. Well done. Nice guy.

You can't call bubbles. It's a joke about the economist who successfully predicted 10 of the last five recessions.

You can't call the timing. If you could, we'd be in a different universe. If this is not a bubble now, it will be. Some of these companies will end up like—and you can kind of point to some of the places where that might happen. And obviously, many people know where people are kind of nervous, but you don't know which and you don't know the timing. What you do know is you've got this kind of hugely consequential technology, and that's what drives all of this alpha and all of this uncertainty.

Well, generally in a bubble, everybody's a rational actor. Almost everyone's a rational actor given their situation. If you're Sam Altman, you've got a commodity technology, you're competing with people who have giant legacy cash flows. You don't have your own infrastructure, don't really have any differentiation, but you've got massive mindshare. So what do you do? Well, you try and swap that for hard assets and you try and talk your way into a self-fulfilling prophecy. You try and turn those lightly engaged 900 million weekly active users into something more tangible.

If you are Larry Ellison, you've got this very cash-generative legacy business that's been in structural decline for 25 years. Most people going through YC have never heard of Oracle, almost literally. No one has invited you to a party since 1998. So what do you do? And here is this thing. You grab onto it with both hands and you burn your way through. The same with NVIDIA. I haven't updated my number here, but I think in Q3 last year, NVIDIA had something over $70 billion of trailing 12-month free cash flow.

So they can't give the money to TSMC fast enough. TSMC won't take it fast enough. And TSMC says, dude, this is a cyclical industry, go home. No, we're not going to triple our capacity this year. And ASML can't ship the stuff fast enough either. So what would you do with that $70, $80, $90 billion? Put it into— or put it into building infrastructure, building market position, building market share, building up your ecosystem.

I'm old enough to remember when this was called vendor financing. Vendor financing is, as long as you're disclosing it and you're not lying about where the money's going from, there's nothing wrong with that in principle. But it's leverage. And leverage is always fine when stuff is going up. And we've got an awful lot of different kinds of leverage going on at the moment, whether that's SPVs or circular revenue or all this kind of stuff.

It's all leverage. That always works until it stops working. And then when it stops working, then you've got a problem.

One of the really basic ways this is different from every other platform shift—I spend a lot of time saying, well, this is kind of how platform shifts work, and this is what happened the last five times. The one way this is unquestionably different is that with all the other platform shifts, we knew what the physical limits of the science were. We knew how it worked. We knew what could happen next month. So you didn't know what the iPhone 3 or 4 would be, but you knew it wouldn't fly and it wouldn't have a one-year battery life.

You didn't really know how the internet would evolve in 1997, but you knew that telcos wouldn't give everybody in the world fiber internet next week. Whereas with LLMs, we kind of don't know the physical limits of how this could evolve. And so it may be that next week we have a paper that means that you can get more or less the same results for 1% of compute. Maybe that's a silly statement, but it might be 10% of the compute, and we don't know that.

Whereas you did absolutely know no one was going to publish a paper that said you could get the same compute with 1% of the transistors on a chip. But we don't know those physical limits. And so we don't know the parameters of what could and couldn't happen to cause a pricing collapse. What's happened mostly so far is the other way. We keep inventing ways of using 10x more tokens. So you get reasoning models and then you get agents.

It's like, great, we're using 100 times more tokens now. The funny thing is, predicting token usage, to me, it's like being in the late '90s and looking at bandwidth consumption. It's the metrics that you've got. But imagine it's like 2003 and we're saying YouTube bandwidth use is doubling every month. Okay, well, that sounds good, I guess, but there's like five or six different multipliers that are producing that, and you don't really know what any of them are. The same thing now with AI CapEx.

So the usage is going up, and there's multipliers that drive the usage up, which is more people using it, people using it more, people doing reasoning, people doing video and imaging, people using agents, corporations using this, people using it for coding, people using it running all day. So what do you mean by tokens are going up? And then you've got the efficiency gain. So Satya Nadella said that inference cost halves every three months.

Great. So that's pushing it down. But then you're chasing the next model, and the next model is always bigger and more expensive. And your model that you have is going to be irrelevant in six months if the next model comes. So how long are you chasing the frontier? And meanwhile, this year—and we sort of said this at the beginning—but this year Meta says it will spend over 50% of revenue on CapEx, not 50% of profits or cash, 50% of revenue.

So Google and Microsoft are not far behind, but they can't double that again. They can't spend 100% of revenue on CapEx next year, and they've already gone out and started borrowing money. And then there's a whole bunch of stuff in leases. And so the actual CapEx number is double-digit percentage numbers higher than that if you look at the leases and all of that kind of stuff. So there's a sort of financial gravity point at a certain point, which is they can't keep increasing at that rate.

It can't be $2 trillion a year next year. Maybe that's the only financial limit, actual hard limit, we have. And so you've kind of got this envelope of how long can you keep building this stuff out? Is it just that we're building the factory and you're going to get to a point—it's back to my point about oligopolies we talked about a while ago—are we going to get to an end state where there are, say, four companies that each spend $250 billion, $500 billion a year each on building and maintaining this stuff?

What's the lifespan of the chips? Is it going to look like this or is it going to do like that? Are we going to get to a point where the industry collectively spends $1 to $2 trillion a year on infrastructure every year and basically makes marginal cost on that? Then what happens on top? And then you get to these kind of very vague TAM questions where Sam Altman and Jensen Huang say, like, the TAM is global GDP. And no, no, no, it's more than global GDP.

Global GDP is what, $70, $80 trillion, something like that? So we're going to double GDP, and 5% of that will go to Sam, and the other 5% will go to Jensen, and some of it will go to ASML. Great, thank you.

No. I mean, a sort of a micro observation is that it's back to the GDP eval thing of, like, the New York finance professor who tried to estimate the TAM for Uber by saying, he said he'd value Uber because this is the TAM for taxis. And so that's Uber's TAM, which is, no, that's not Uber's TAM. Uber's TAM is something completely different. You get to something like advertising, the TAM for digital advertising has to include retail rents, and it has to include shipping.

And it basically has to include retailer margin, which is 50% of retail. So the TAM for advertising, for digital advertising, is not advertising, it's something else. And yes, GDP growth is GDP growth. You would have to be quite wild-eyed to think that we're going to go from whatever it is, low single-digit GDP growth to double-digit GDP growth. That would be a different conversation. The human labor thing, I mean, look, you go right back to academic discussions of the Industrial Revolution.

Automation happens in places where labor's expensive. Because if labor's cheap, why bother? I was reading a book a couple of weeks ago about Rome, about somebody in the south of Italy in the '30s. It's like there's one car in the whole province. And it's like, well, why didn't you get a car to do that? Well, because people are free. You just tell a person, you tell a peasant, and so they're still peasants.

You tell a peasant to walk eight hours to that town and get something and walk eight hours back, and they do it, and you don't even have to pay them. So why would you buy a car? So that's like the very basic first-year economics student answer to this, is that automation has to have an ROI relative to a person. Then you get back to the price elasticity point. And to me, again, I have a slide with three bullet points. A friend of mine said, "Benedict, you talk in slides," which is probably true.

I think in slides. And so step one is, okay, if you make it cheaper and easier to do a given thing, do you do the same thing for less money, or do you do more for more money or more people? Step two is, was having a building full of people doing that your moat? So the example people kick around here is health insurance. I think somebody said the marginal cost of arguing is now zero. So if your business is based on making it really painful and difficult to do what you do, and suddenly it's not painful and difficult, was that your barrier to entry?

But then, to me, the interesting sort of third step is there's a whole class of stuff that you just couldn't do at all. Go back to talking about the Jevons paradox and steam engines. Imagine if you want to make an express train from London to Scotland in 1800. Never mind that you didn't have steel and iron, couldn't mass-produce iron, and you couldn't have laid the rails. But presume you could have built the whole thing, but you didn't have a steam engine.

You couldn't buy 10,000 horses and put them on the front and have the horses pull your train to Scotland. It didn't matter. It doesn't matter how many horses you have. You just can't do that. But there's whole businesses now where you could not do that with people, no matter how many people you had. You couldn't run a quant fund if you had 10,000 people instead of a computer, even if you could pay the 10,000 people. You just couldn't do it.

And years ago, I think somebody, I can't remember who it was, said there's a huge difference between when something gets very cheap and when something gets actually free, which is what happened with internet bandwidth at a certain point and happened with mobile bandwidth. And again, it's like, are you doing the old thing with a new thing, or are you doing something different with a new thing? Which is sort of a discursive way of talking about your question. But we're going to have people do stuff with AI that doesn't replace a person.

It does a thing that would've needed a million people, which is what happens with all automation. That's what happened with steam engines. It's what happened with electricity, what happened with aircraft and steel and everything else. You don't do stuff that says it's as though you had 100 horses or 1,000 horses. You do a thing you couldn't have done with horses, or no matter how many horses you had. And so, yes, well, but that's also—and I'm not an academic economist—but there's a famous quote that you can see the information revolution everywhere except in the productivity statistics.

And so then you get these arguments about, well, what's in the productivity statistics? And switching from buying 40 devices that cost thousands of dollars to buying this is a decline in GDP. There's a measurement problem and a metrics problem in all of those things. But to the core of this, are you going to replace McKinsey with AI? Well, that's like saying, are you going to replace KPMG or PwC with spreadsheets? Kind of not the right way of thinking about it, because on the one hand, you get that price elasticity.

On the other hand, if you ask that question, you kind of just told me you don't actually really know anything about what McKinsey does. I made the mistake of replying to somebody on LinkedIn who was talking about this, and they said, "But look, publicly listed consultancies have already crashed." But they're all partnerships. Look, Booz Allen has crashed. I was like, okay, do you need me to explain why Booz Allen's share price has collapsed since Trump was sworn in? Because it's not to do with AI.

It's a whole other conversation. I wanted to do digging into this accounting thing. What did automation do to accounting in the past? I decided maybe I should go and look for audit costs. And it turns out that you can get a chart of average audit data, all sorts of data for audit costs, which has basically been flat since the early 2000s, despite everything that's happened in software. Average audit cost hasn't changed. But then I kind of, well, I can dig a bit further, and got into what happened to audit costs in the '70s and '80s.

And what you get is all these 50-page academic reports that mention all sorts of stuff that changed audit costs and what was going on in audit. And then none of them mentioned computers. There's like 700 other things going on that are changing how audit works, none of which are actually computers. And it occurred to me, this sort of applies in every industry. So I was at an event like two years ago listening to Martin Sorrell, who founded WPP, which is one of the big global ad networks.

And he was talking about AI and he said, "No, no, the big impact of AI is not making more advertising assets. The big impact of AI is that actually every big ad agency has got buildings full of people doing really boring crap with spreadsheets and faxes, loading the ads into Facebook, almost literally. One of the huge impacts of AI in advertising is on the operations, not on the stuff that you, if you know nothing about advertising, you would just think, well, it'll make more pictures."

Well, yes, but no. The big thing is all the guys in the back office shuffling paper around. And I think the kind of point applies to every industry: you're sitting on the outside, you can go, "Oh, of course AI will do that." And if you're on the inside, you're like, "No, no, no. Over here, and it's over here, and that's not the hard part."

So I certainly wouldn't say bewildered, at least not any more bewildered than anyone in tech. And if I think, what's the center line? If you think you're in control, you're not going fast enough. Like, if you understand any of this, you're not paying attention. I came across the other day that famous physics professor who says, "I'm going to teach you an eight-week class on quantum today. I don't understand quantum."

At the end of this class, you won't understand quantum either.

This certainly applies to AI tech, but it's like the Schleswig-Holstein question, the joke about that. Oh, the Schleswig-Holstein question is this famously complicated 19th-century politics question. And I think Palmerston said only three people have ever understood that: the German professor who's gone mad, somebody at Bismarck who's dead, and myself, who has forgotten. Anyway, you can cut that out. The point is that no one really knows the answers to, like, does it go to AI and AGI, and what will it do next year?

But everyone's got a bunch of stuff deployed. Everyone deployed Copilot and went, "Oh, okay, that wasn't very successful." It's kind of like giving everybody the internet in 1997 and saying, "There you are, be more productive." Didn't really work. Everyone now has a bunch of pilots. Some of those pilots have made their way into production. Some of them haven't. Some of them work, some of them didn't. Everyone is sort of scratching their head and thinking, "Okay, now what?"

And I think, actually, there's certainly a curve of awareness. So I talk to some industries where my brief before I go up is, they say, "Okay, like I spoke at an insurance conference a couple of weeks ago, and part of the brief was, so everybody here remembers people talking about how drones were going to change insurance. So we need to make it clear this is not another of those things. This is a bit more real."

And so there are people who are kind of at that stage in the curve, which also maps to how hard it is to work out what you would do with a non-deterministic system. Lawyers are probably somewhere in the middle on that. Obviously, this is useful, but also it's really easy to get yourself killed with it. But there's also, like, the head of the curve is more like, "Okay, is that it?" So, like, I spoke to the CMO of a big retailer last summer, and they said, "Okay, we've all got review summarization."

We've got a support chatbot, and then we took that down because it was a disaster. And then maybe we'll put one up that isn't a disaster. And they've got, like, natural language search. So you can say, "What should I buy for a picnic?" This is a classic clichéd use case from Walmart. And they've got review summarization, and they've got a SKU tagging project going on, and the marketing team is using it. And they've got, like, eight different projects automating obscure backend internal things deep inside the company.

And Walmart will say, like a lot of people, "We've been doing machine learning since the '80s. We've been doing AI since the '80s. This is more stuff. We've got a huge tech organization." I think that kind of comes back to the framing I've used a lot, which is, I think a lot of industries now are well down the path of, we can use this to do a thing we were already doing, but better. And maybe we can see a few things we can do with this that we couldn't do before.

Or that didn't work very well, and now they'll kind of work. So I was at a logistics conference a while ago, and they were talking about the warehouse person can just kind of point their phone at a box, a bunch of boxes on a pallet, and it will pick up all of the barcodes at once. And you think, like, yeah. And then you realize, okay, that's probably like an hour a day that's saved. So there's a lot of individual point solutions where people are deploying stuff.

I think where it's much harder to get your head around is still, yes, but what are the new things? And what might the fundamental change look like, as opposed to, great, now we've put our catalog in a PDF on our website. And I don't think anybody in tech knows that either.

Everyone is still sort of, which is why you get, I think in the last week, both Sequoia and Khosla did, like, 2x2 suggestions. They're just suggestions of, well, what might the axes be? What might useful ways be to think about this?

Frame.io, which you guys invested in. And when I was at A16Z, I was in the pitch. We didn't invest in it. I can tell you why afterwards. In principle, you could have had that idea at any time in the previous decade. I mean, maybe some of the technical implementation might not have worked. Doing it in the browser might not have worked, but it's basically Google Docs for video. And set aside the difficulty of doing that kind of video in the browser.

And theoretically, you could have done that in 2010. Most SaaS companies are database wrappers, where somebody realized that here is this problem, and here are the people who have it, and here is a way of turning it 90 degrees. And this is your insertion point. This is how you build it and take it to market. And the hard part of that was not writing a bunch of SQL queries and setting up some database tables on AWS. And the hard part of writing software is not writing the code.

It's all the other stuff around: what should the code be doing? And how would we tell people that they should be using it? And what should we charge? And how do we go to market? And which bit of the market should we be selling to? It's all the other stuff. And I don't think that's changed at all. It's just that now, on the one side, it's way quicker and cheaper and easier to build that software. And on the other hand, there's a whole class of stuff that you couldn't do with software before that now you can.

I remember looking at machine learning, and when machine learning was AI, I think I said this last time, there's a progression. It's only AI when it's new. It's like technology. It's not technology after about 10 years. The point was that machine learning comes as image recognition first. That's what starts working with ImageNet. I would show this to a German reinsurance company, and they would say, well, that's great, but we don't have any images.

I mean, maybe damage assessment, but really, we don't. That's not what we do. And it took a while to work out that this was pattern recognition. And once you can conceptualize it as pattern recognition, then it's a lot easier to think, well, what can we turn into pattern recognition? Of course, there was also a class of stuff where people turned it into image recognition, where it wasn't image recognition before. But in principle, you kind of have to think, well, what's the right level of abstraction for understanding what this is?

And the same thing now with generative AI. I don't think we've got a good sense of what it is, which is why people say, well, I used to say AI gives you infinite interns, or is it interns? Is it associates? Is it automation? Is it the next stage of humanity? Is it something? None of those answers are going to be right exactly. But you kind of want them to be useful as ways of thinking about where you might look for those opportunities.

But I don't know, as a builder, what would be an interesting project is what percentage of entrepreneurs went out to look for an opportunity, and what percentage started from seeing the problem. Because there are certainly some entrepreneurs who say, I want to do a startup. I'm going to look for XYZ. And then there are entrepreneurs who say, I believe mobile is the thing, I believe AI is the thing, machine learning, I'm going to look for things I can do with that. And then there are people who are like, for the last 10 years of my life, I've desperately wanted to fix this problem in this industry.

And I suspect that it's the third category that produces the best outcomes. I don't know. It would be interesting. It'd be an interesting analysis. Yes, you can pick up AI and go around hitting everything with it. It's a hammer and everything's a nail. Or you can start from the other end and think, now finally I can solve that problem, or I can turn this into a machine learning problem. And I think probably the right answer is all those produce great companies.

Thank you.