Speaker 1Hi, I'm Steve Clapham, and welcome to the Behind the Balance Sheet podcast, where we meet leading investors and commentators and educate ourselves about the world of investing and the world. Our mission is to remove some of the mystique around investing and improve our understanding of successful investors' strategies and tactics. Behind the Balance Sheet is a research and training consultancy that worlds top investors from U.S. firms like Fidelity in Wellington to Bailey Gifford in Schroeders in the U.K. rely on us to train their staff. Check out our forensic analysis and valuation courses for institutional investment teams on our website. And we're not just for teams. Thousands of individuals have taken our Analyst Academy and other online investment training courses. You can find them on the website, where you can also sign up for our free weekly newsletter. Check out our homepage, BehindTheBalanceSheet.com. In the investment business, everyone has access to the same data. Gaining an information edge comes from who can get through it faster and who spots what change before the rest of the market does. AlphaSense just launched Super Analyst, an always-on agent that works like a power user inside the platform on your behalf. It draws on over 500 million premium sources, from company filings and earnings calls to broker research and expert transcripts. And it keeps working after you've closed your laptop. Think of it as an extension of your team. It remembers your projects, follows your process and your house templates, and every output links back to its source. So when a PM or an investment committee asks where a conclusion came from, you can show them. So hand it the work that usually eats you a week, like getting up to speed in a new sector, preparing for earnings across your book, or building the first draft of a research report or IC memo. Start your free trial. At AlphaSense.com/btbs. That's Alpha-Sense.com/btbs for Behind the Balance Sheet. Today, I'm joined by my friend, Benedict Evans, one of the most insightful tech analysts around, longtime publisher of a popular weekly newsletter, and former strategy director of Anderson Horowitz. I've learned a lot from Benedict over the years. I think his particular skill is the ability to both understand the tech side of the industry, and the financial ramifications of new developments. We discussed the publication of the prospectus for the Anthropic IPO, which is our first real insight into the extraordinary economics of the large language models driving artificial intelligence. There are some really fascinating questions here. Can Anthropic and OpenAI develop sustainable competitive advantages? What happens when the cost of producing intelligence continues to fall very rapidly? And will the hyperscalers earn acceptable returns on the hundreds of billions of dollars they're investing in AI infrastructure right now? We also discussed the opportunities for specialist software providers, the implications for the big tech companies, and why the winners of the AI revolution may not be the companies investors currently expect. Benedict draws some fascinating parallels with previous tech revolutions, particularly the internet and mobile telecom. This is a conversation about the difference between technological success and investment performance, and I think it provides a really useful framework for thinking about investing in AI. I hope you'll really enjoy my conversation with Benedict. So, Benedict, listen, it's been ages since we've sat down at a proper discussion. We sat down in New York about six months ago, I suppose. Welcome to the podcast. I've been dying to interview you, and we're going to talk about AI and Anthropic. And how long did it take you to read the IPO, or you used your favorite AI to read it for
Speaker 2you? Well, I know this is a joke, and there's like, you know, 11 million pages about how this is going to kill us all as well. It kind of struck me that this is sort of a Rorschach plot, in that if what you're interested in is debt, or you're interested in circular revenue, or you're interested in, you know, you're interested in operating details, whatever you want to know, you can sort of find stuff in there. So, you know, if you're an accelerationist, you can say, well, you know, the revenue has grown 10x in a year and, or more, and a growing portion of the research is being done by Claude. If you are one of the diminishing band of people that thinks this is all kind of nonsense, well, there's like $80 billion loss, if you don't bother to work out what that loss actually was, which is basically pure accounting fiction. Yeah. So, I think the challenge I have is what it doesn't tell us. And firstly, it doesn't tell us depths of usage and spread of usage. It tells us revenue and customers, but it doesn't really say how well this is working outside of a narrow portion of people, which is partly, and Swapit can't see that because it's API usage. But it's also how much predictive value does this have? So, we've had this complete transformation in this industry in the last 12 months in that software, using AI for software development, suddenly got explosive product market fit, which has driven this enormous growth in revenue and means that the company is now more or less breaking even on a kind of cash flow basis. But we don't know what that's going to look like in the future and the sort of, if you go and look at the P&L. Yeah. Yeah. It's funny. I was talking about this with a friend yesterday. The fixed cost looks like marginal cost and the marginal cost looks fixed in that the cost of inference, which is the cost of revenue, has basically stayed flat as the revenue has increased NX because there's so much efficiency gain that's going on in the operations of this stuff. On the other hand, the training cost, which is your R&D, which should be a fixed cost, has gone up in line with revenue. So, if you didn't label the chart, you would think they were the other way around. Yeah. Yeah. And so there's two things. And we don't know, the problem is we don't know what's going to happen to either of those things. The easier thing to talk about is the pricing of the inference. So you're basically selling tokens, which is rather like selling bits in like 1998. Right now, there is a once in a generation shortage because the software industry wants to buy far more of these things than you can get. And so you can, if you weren't making a profit now, your CFO isn't doing their job because you can price at the customer's ROI. But on the one side of the table, the combination of more CapEx, more model efficiency, more inference efficiency, more chips means that the cost is dropping sort of 10x a year or more. And on the other side, agentic means one, two orders of magnitude, maybe more token demand. And the usage is still broad, but really very shallow in that. Like the really kind of intense usage is very narrowly concentrated. So I think ramps data and ramps data skews to startups and early adopting technology companies said that like the top 1% of their customer base is spending 15, 10, 15 grand a month per employee, but the average is $15. Yeah. So you've got this very small, what I was going to say is like, on the one hand, like you've got this very, very small portion of people who are using this a lot, but you don't have a billion people using this a lot like in Instagram. So you know the demand is going to do this, you know the supply is going to do this. You don't know what the balance will be and so you don't know what the pricing is going to look like. You also don't know how the research is, how the training is going to evolve. You don't know if the scaling will continue. You don't know if they're going to continue having to spend more and more money to get a better and better model. And you don't know whether they're going to be able to differentiate and have some kind of competitive advantage over model companies. So you can say today, this is what they're spending. But you don't know what any of those are going to settle out at in 2, 3, 5 years' time.
Speaker 1No, sure. I mean, obviously we're, and spoiler alert, we're not going to solve the world's problems and the valuation of Anthropic in the next half hour, 45 minutes, because there's a lot of unanswered questions. But let's try and figure out what we're looking for in terms of a roadmap, because I think that would be the really helpful thing to listeners. And the thing that, you know, people always start with is the hyperscalers are spending all this money. What does that mean in terms of how much does Anthropic and OpenAI need to sell in order to justify that spending? And I don't know if you have a perspective on this. I've seen figures of $3 trillion by 2030 or $2 trillion by 2030, but whatever they spent, they presumably want to earn in return. And that means quite a lot of revenue. So half a billion dollars. There's three quarters of a billion dollars of revenue. Does that make sense? And is there any hope that by 2030, Anthropic and OpenAI will be generating that sort of level of turnover?
Speaker 2So I think there may be two things within this. The first is, this is a slightly iterative calculation in that you don't really know what the run rate capex is. There's a sort of presumption that we've got this huge hump of capex now, and then it'll kind of recede somewhat. And the revenue will catch up. But is the run rate capex half a trillion or a trillion? I mean, it kind of depends how much demand there is. Of course. So that's one answer. Secondly. You know, I started my career in 99, and then I became a telecoms analyst in 2000. And at that point, the mobile operators in New York spent about 110, 120 billion euros on 3G licenses. And many years later, it turned out that all the stuff they thought people would do did in fact happen. We do all of this. But they didn't make any money from it. You know, I haven't run the calculation, but I'm pretty sure that most of those operators did not make a return on that investment, even though it turned out mobile internet was kind of a big deal. So it's entirely possible that you get an enormous amount of adoption for this stuff at quite low prices, that you get a price and collapse. And so you don't produce a return on this investment. It's also possible you get an enormous amount of new revenue, but the infrastructure itself is a low margin. It isn't. There's low margin commodity infrastructure with no pricing power. And so the value goes to other people, which is kind of what happened in mobile. Like Uber and Amazon and Instagram and everybody else make money on mobile, and that money doesn't go to the mobile operators.
Speaker 1And that would be possible. It's possible that both the model providers and the infrastructure providers, the data center providers, neither of them make any money because they're both in commodity businesses.
Speaker 2Exactly. And, you know, the wrinkle at the moment. The wrinkle at the moment is that there's such an enormous imbalance of supply and demand that you have a more or less undifferentiated commodity that can be sold at high prices, which is obviously what happens if you're an oil company. Oil is oil. Well, obviously, oil isn't exactly oil. But, you know, one barrel of West Texas is the same as another barrel of West Texas. But if there's more demand than supply, then you can price at a very high margin. That doesn't last. So, you know, it's possible that supply remains behind demand for quite a long time. And so you could have quite high margins with zero differentiation. So that's kind of, you know, another moving part in the question. I think there's a sort of an entirely different conversation here, which is what is the AI product? We're sort of at the stage now where you gave everyone a model. And for some people, that was life changing, which is kind of like giving everybody a spreadsheet in the mid 80s or giving everybody a web browser in the mid 90s. Like spreadsheets were kind of life changing for 5 or 10 percent of people in the mid 80s, early 80s. I remember. But that wasn't how you rebuilt your business around computers and change your financial operations. That was SAP and Oracle for the sake of argument. It wasn't like that you completely changed how you manage your internal financial operations by just giving all of the people who were doing paper and running mainframe spreadsheets. That wasn't what happened. Same thing with the Internet. You gave everybody in the company a web browser. That wasn't how you built an online business. That was something else. The same thing now. You gave everybody Copilot or Claude or ChatGPT. That isn't how you work out how AI transforms your business. But does that happen with things that are fairly close to Claude and ChatGPT and Gemini and everything else? Or is there a whole new world of enterprise software companies and consumer software companies that use models as low margin commodity infrastructure somewhere inside, kind of like cloud is? I mean, cloud is a good business for some people, but it does not set the agenda or control the entire Internet. And so there's an outcome in. Which a high portion of what you do is something that's more or less Claude or more or less ChatGPT. And they capture the value from that. And there's another outcome in which, no, it's Harvey, but lawyers by Harvey or something else. And that the typical big company today in America has four or five hundred SaaS apps. And in 10 years time, they will have a thousand apps, half of which are running on AI models. But. They're not paying Claude and they're not paying Anthropic. They're paying a thousand new software companies. And so the value goes further up the stack. That's how it's always worked in the past. Yeah. And we don't kind of know how this is going to evolve. This is part of the challenge. If you're kind of super accelerationist, you would say, look, the models are just going to get so much better that they'll be able to do the whole thing themselves. Or at any rate, the CIO will be able to get them to do the whole thing himself or herself. And so most of the value. All of that reconfiguration of how the enterprise works will be captured by the model companies themselves. There's another view, as I said, which is kind of no, I don't think that's how any of this ever works. And that this will be infrastructure that is where most of the value gets captured by the people further up. Part of the challenge here is. Unlike every other platform shift, we don't have a good understanding of how the technology works, like we don't have a good first principles understanding of why this works so well. So we knew with the Internet. We didn't know that Google was going to get created, but we knew that PCs were expensive and telcos weren't going to give everyone fiber broadband next week. And like we knew the basic parameters of what was going on, whereas with this, we don't know what the science is going to produce next. And so we don't know how good the models will get on what dimension or what cost. And so we don't kind of know what's going to happen back to the PNL or the Anthropic S1. We don't really know what that R and D line is going to look like because there's like a worst case scenario, which is the whole thing. Kind of flams out next summer and the R and D cost drops off the PNL. But so does the revenue growth.
Speaker 1Well, I'm going to come back to the points you were raising about who captures the value because I think that's very important. But just before we get there, I just want to just finish off on the hyperscalers because I wrote, I don't know, three years ago about the fact that the these companies were extending the lives of their servers. And so they. They started off depreciating their servers over three years and ended up after several changes, depreciating them over five or six years, depending on the company. And I pointed out that I didn't know how long a chip lasted, but I knew which direction they were going and what they were doing was they're boosting their earnings growth. And so, you know, this is I don't know what the right answer is, but what I do know is that they're accelerating their earnings growth and making their companies look more valuable. And that that's quite dangerous from the perspective of an investor. But it's now become like a serious issue. I mean, when I was talking about this, nobody was worried about servers spending that much. Now, Michael Burry has been, you know, shed the spotlight on this. Just give us your sense of if the chips last five years, what happens in year six? They just use them for inference or reduce demand, you know, not as high pressure kind of applications. They're still useful or do they need to just go back and put a whole new set of chips in?
Speaker 2So I think there's probably two or four bits to put things to threads to put on here. First of all, I mean, as you point out, like five years ago, these companies probably spent one hundred billion dollars on CapEx this year. Like Meta, Google, Amazon and Microsoft combined plan to spend about seven hundred and fifty billion this year. Five years ago was one hundred billion. So this is enormous search in CapEx. And that means that three of those. Four, like the other than the one Microsoft, so Meta, Amazon and Google, sorry, Meta, Microsoft and Google are all going to spend over 50 percent of revenue on CapEx this year. And so one of the big structural changes is that the CapEx is outgrowing the free cash of these companies. And so they've had to go to the capital markets, which is why suddenly this becomes people start talking to credit analysts. Well, how much credit is there in the world? Like you didn't have to pay attention to this if you're a tech analyst before. What does IG and HY stand for, please? Like those kind of questions. Suddenly, like you have to ask. I think secondly, you've got this continuous acceleration in efficiency of running the models. And so between more CapEx, but also better chips and better models, you've got like you've got the capacity, raw capacity people are measuring in gigawatts or megawatts. Like how many gigawatts of data center have you got? But then there's how many tokens can you get per gigawatt? Because the chips get better and the models get better. And then there's how good a result can you get from token? And so there's sort of a lot of moving parts in the return on all of that investment. Then you have, and of course, then what can you charge ultimately for the results? Are you selling tokens? Are you selling results? What are you selling? What can you charge for that? Do you have differentiation? So that's one piece. The second is. At what point do you decide? Decide? It's a better. It's a better overall return to throw those chips out and put in new ones that get better results per watt. And obviously, that's a spreadsheet calculation and a capacity and a timing infrastructure. That's for infrastructure people to kind of work out. Like, does it make sense for us to rip that stuff out or to leave the stuff that's in there and put new stuff in a new data center and a new shell? And obviously, part of the calculation there is your ability to get access to power. Because if you could have as much power as you're willing to pay for, then it would be easy. You'd be more likely to leave the old one. If they were producing an economic return on their own right, in their own right. Whereas if you're power constrained, then it makes sense to put in new chips that get a higher return per watt. And so that's even without asking, does the thing actually work? Then what's the working life of a chip? Does it actually start kind of packing in? Does the failure rate start rising? I mean, I think the answer is you have a failure rate and then there's a cut. And as you know, over time, the failure rate. It starts getting worse. I don't know what that looks like. I'm not a data center analyst. I don't. think anyone has a really really good sense of what that looks like until you've just run these things maxed out for x amount of time and then there are also things you can do to mitigate that and what does the cooling look like and so on so you get to give a quick get quickly get into a lot of quite nitty-gritty technical questions on you know what is the um what's the actual um mechanical life of these things as opposed to the useful economic the the life at which you're still making an economic economic return on these things i think and that also of course depends on whether you're using them for inference or for training and what's the loading on your system and how fungible are they between inference and training um all of which is a long way of saying i don't know go ask your data center analyst but there's also there's also i suppose there's a are we focusing on eps or are we focusing on cash flow because obviously amazon has always said you know ignore the eps look at free cash flow and if your objective is to manage capex and you're not going to be able to manage capex and you're not going to be able to manage to sales and to go out and raise debt to pay for that um is that an eps question or a capex or cash flow question um i mean one of the things that's occurred to me with these companies raising external capital is obviously the question is do you want to if you believe that you're going to grow a huge amount then you would by default you would go with debt rather than equity because you're going to grow into the equity debt without giving up the equity or you do a convert maybe um as a credit investor it kind of doesn't matter if google gets zero return from all of these data centers as long as the ad revenue is still there because they've got no debt and you know massive cash flow so if the data centers all empty and end up as space eaters as a credit investor you kind of don't care because the ad revenue will still be there um as an equity investor obviously well that was all of your money that's just gone into a hole in the ground um but there's a sort
Speaker 1of but if you look at the big picture the big picture is that we're having trouble building enough data centers so in five years time having a data center that's got power and all the rest of it water whatever that's going to be the constraint right so it's quite it makes it more likely that you're going to bin the old chips and replace them with new ones because you can't just go and build another data center yeah which would normally be what you would do if you were
Speaker 2do right yeah exactly then of course the question becomes well as i was saying well what is the supply what is the demand for tokens they're not looking like and you know this is the comparison people like to do is with with fiber 25 years ago which isn't quite right because fiber was built out massively ahead of demand whereas this is being built out behind demand but what also happened with fiber is that it got massively disaggregated and fragmented and then everything at every layer in the stack was built out and then it was built out and then it was built out and then it was built out and then it was built out and then it was built out and then it was being sold at marginal cost and that is very certainly what could end up happening
Speaker 1here expert calls have always been one of the most powerful ways to build conviction but today investors are asked to cover more companies move faster and do it with leaner teams with office sensors ai led expert calls their tk's call service team sources experts based on your research criteria and lets the ai interviewer get to work the magic is in that ai interviewer purpose built and knowledge base informed to conduct high quality context rich conversations on your behalf acting as a trusted extension of your team then they take it one step further your call transcripts flow natively into your office sense experience and become queryable searchable and comparable so your primary insights plug directly into earnings prep diligence work streams and pitch books with zero tool switching and with office senses expert call services the ai led expert calls are just one option because of the high quality context rich conversations on your behalf acting as a trusted interviewer because they know the importance of a hybrid expert research approach ai for coverage and efficiency humans for complexity and conviction it's the institutional edge that scales research without scaling headcount for hedge funds that means validating thesis assumptions across dozens of experts before earnings instead of just a handful for private equity it means faster pre-ioi scans and deeper commercial diligence for instance if you're a private equity investor investment banks and asset managers it means pulling real operator perspective straight into models and sector positioning without disconnected tools or manual handoffs although it lives inside the alpha sense platform trusted by 75 of the world's top hedge funds alongside filings broker research news and more than 240 000 expert call transcripts turning raw conversations into comparable auditable insight take advantage of alpha senses ai led expert calls now the first to see wins the rest follow learn more at alpha hyphen sense.com forward slash btbs so let me just turn to the this idea of the commodities um are the models commodities because what i don't really understand is why would you i mean i know from my personal perspective i don't really understand why would you i mean i know from my personal perspective i don't really understand why would you i mean i know from my personal circumstances that i like to use chat gpt when i'm wanting help with editing my newsletter and finding a title and i don't find claude so good at it and i find claude better at producing spreadsheets than chat gpt um presumably there are some some differences but at the end of the day is is there a really distinct business with a moat as we recognize it in in a model in an l.l.
Speaker 2them so the simple answer is today not really um if you and it depended and then it depends what you're doing so the aggregate benchmarks the headline benchmarks anthropic open ai and now google and meta are all pretty close and chinese models are not very far behind and spacex is not very far behind and for the sake of argument we'll get back onto the top soon when you get into specific use cases then you can start seeing more variance because you're disaggregating those benchmarks when you are pushing right at the frontier and you're doing something super hard that the model is only just able to do something out then you can see differences but then the question is well what are your use cases so you know sort of pick kind of you know because you know the interesting benchmark curve is actually to look at price versus performance where there's a very very very sharp curve so you can have absolutely the latest full fat model that's very expensive you can have the model from six months ago that's sort of 90 on the benchmarks and 10 of the price i'm making these numbers up but that's kind of the story so the curve does that and at the extreme like voice recognition is done and it runs on your phone so as a company whether you're a startup building something for consumers or whether you're building something for enterprise what is it that you're trying to do is your objective to do um summarization of the call what's being said in the call center that's not going to need the frontier model that won't appear for 10 years maybe next year's frontier model will do that a bit better but clearly as you sort of actually separate out if you try and say well we've got 500 sas apps what would an ai be what what model would be able to do that um which obviously is never as simple as that but like there's clearly there's going to be some use cases that work fine on the open source model from two years ago or that work fine on today's model but only have an roi when the price has come down by 90 in six months um and then there's other things at the other extreme where things really really will benefit from absolutely the latest model um i mean this is the um you know this is the thing we should have mentioned earlier talking about the anthropic pnl 80 of revenue comes from models built in the last six months and that's partly because people only want the model built in the last six months and it's also because the model from a year ago is such a commodity and it's super super cheap and maybe it runs on prem and so you can't make any money from it and there's a lot of use cases in that category there's a lot of things companies are doing with those models so the other extreme would be you know um go look at all of our customer emails and every recording with a customer call and all of our usage telemetry and do a competitive review and tell me why our churn is changing and that might or might not work now but you can like that will get better answers with a better model for quite some time um but this is the kind of the question is do you need the newest fastest fullest fat model and for that there's pricing power and maybe some differentiation like i like this the result i get from this one more than the result i get from that one um for this use case or that use case but as soon as you start doing going down that curve then they all blow into one and you know if you are building you know a tool to do that summarization of you know what the trends in customer calls um then you're going to be looking at your price per token you know which model running on which cloud and you're going to be swapping them in and out and the enterprise is going to be asking itself just as it asked for cloud do we build this ourselves or for sass do we build this ourselves or do we buy it and the answer as it was with sass was kind of depends we're not going to build a own email server. We're not going to build our own cyber intrusion notation systems. We may well build our own thing to manage that thing inside our payment flows.
Speaker 1The answer to my question is that there's a relatively small number of use cases, probably, that require the most expensive, most up-to-date, most profitable model.
Speaker 2Well, I'm not sure if it's small, but I think the point is there is a distribution of many use cases, many models, many prices. And there are some use cases where you absolutely want the newest, full-fat, most expensive model and where you also may be able to see differences between them. And as soon as you start moving back down the curve of those use cases and those prices and those models, the more they become interchangeable, which is part of the reason why everything they built before April is only 20% of revenue.
Speaker 1So, I mean, just talk us through. I mean, what would be? An example of a customer that needed to have that latest type of model. So they've got 1,900 customers that are spending more than a million dollars. Do you know, do you have a feeling for what sort of companies those are? I mean, presumably some of them are.
Speaker 2I think it's concentrated. I think we all understand that this is very, very highly concentrated in software development. But I think that's, you know, that is transitory. And this is one of the questions is how fast does that expand outside of software companies? The way. The way that I've been sort of trying to think about this is that you can think of AI as unlocking two things. Maybe one of them is automation and the other is abstraction. And so automation is faster, cheaper, doing the old thing faster, cheaper, doing things that you needed people for. Maybe that people couldn't do. I listen to every call coming into the call center. Tell me what people are really saying. Tell me if people are more angry. Are people more upset than they used to be? Are they happier with the product? And these are things that you kind of couldn't ask a database and that you didn't really type into Salesforce. And sometimes that's transformative. And this is obviously what you see in cyber now. If you throw 10,000 agents at your software, at your code base and say, try every possible vulnerability in every possible combination, then they'll find stuff that people could never have found before. So that's good. If you do stuff faster and cheaper, that might just be price elasticity, but it might actually change the way the business works and it might unlock new stuff. Or indeed new kinds of competition. But the other side that I think is, and so that some of that is, okay, well then, you know, we will throw 10,000 agents at that. And so we'll use way more tokens. The other side that I think is more intellectually interesting is to say, what are the new questions you can ask? And so there's this old saying in computer science that a computer should never ask you a question if it should be able to work out the answer. And that used to mean things like what printer do you have, but now it might mean you open. Salesforce and it says, what question do you want to ask? And you construct your carefully devised 35-step database query that produces a list of who you're going to call this week. Customers who've said this and done this and we called them at this time and they're using the product this much and this is how much they paid. You know, you construct your complicated database query. But what you actually want to know is who should I call this week? And so you should be able to ask the computer. You should be able to call, say, which customers are going to churn instead of doing a 30-step query. You should be able to say what invoices are due this week and what invoices are overdue and why. Pick your business question as opposed to pulling lots and lots of levers and switches and knobs to get the computer to give you the question. And you should be able to tell the computer what you want to do rather than give it lots and lots and lots of complicated criteria. And you should be able to tell the computer what you want to do rather than give it lots and lots of complicated criteria. That's new software. That's a new software company. And that's a lot of tokens. Especially when you start saying, look, instead of asking this question to Salesforce, it's got all of our app telemetry and it's got all of our recordings of all of our customer calls. And it's got all of our payment data. So ask me who's going to churn. And it should be able to work that out by looking at all of those things. And when you start building things like that. Then you need a bigger, better model. You could imagine a bigger, better model is going to produce better results for this. And it's probably 10 models that are chained together and built in various complex ways. When you start doing that, that is also probably not a copy of Claude running on the CFO's laptop. Any more than the CFO ran inventory for a 10,000 person company from a copy of Excel running on their laptop. Yes, they have Excel, but that's not how that works. Yeah, true. That's something that needs to be built carefully and thoughtfully by a bunch of people. His job is to think about how this stuff works. And you're probably going to pay them and they may not work for you. They may be a whole other company that might be Accenture. It might be Bain. It might be a new startup. It might be SAP. It's not Anthropic. Anthropic isn't going to be able to create all of those by themselves. And so all of that sort of gets to a hypothesis that says, well, no, explain to me why this doesn't work like cloud. Explain to me why this isn't infrastructure, that other people have to go out and invent, work out what that stuff is.
Speaker 1But that's where the real added value will be. And I mean, I mean, my impression is that the models are largely commodity businesses and that the real added value will be in the application layer, the vertical, the different industry verticals and where the domain expertise will be significant. I met the Harvey CEO, Winston Weinberg. He was speaking at London Tech Conference. And he was saying, well, actually, I think, you know, young lawyers have got a great future because they'll be able to do more interesting work. And he says the thing that's really important is having the domain expertise. And that's something I took away. Well, in the words of Mandy Rice-Davies, he would say that, wouldn't he? Well, he would say that. But it was a certain logic. To that, that
Speaker 2I could come up. There's a sort of, yeah, I mean, there's a sort of an ontological accelerationist argument here, which is to say, look, if I define the future of AI as something that's going to be able to grow to the point that it can do anything, then it's going to be able to do anything. And it gets kind of circular. Yeah. I mean, there's a thing I always like. Obviously, you've got a fairly intellectual audience. So there's a young. There's an argument made by Anselm, who's a medieval theologian, who said, step one, definition of God is that God is the greatest thing imaginable. Because if you could imagine something greater than that would be God. That feels reasonable. Step two, something that exists is greater than something that doesn't exist. Because if there was an X that existed, that would be a greater thing than an X that didn't exist. Also seems hard to argue with that. Step three, therefore, God exists. And every other theologian in the room says, but that obviously doesn't work. And Anselm says, OK, prove it. And it takes like eight or nine hundred years of philosophers arguing about this to explain, well, why is that not true? Why does that not work? I think Bertrand Russell said it's actually much more interesting to work out why it doesn't work than the actual proposition itself. It's like it's obviously not true, but try explaining why that doesn't work. No, sure. And you get this talking about AGI, which is like, well, clearly these models keep getting better. And so if they keep. Getting better indefinitely, then they're going to get to the point that they're massively cleverer than us. And if they do, then clearly they would have all the domain knowledge you're talking about. And they would know what lawyers are and they'd be able to just do what the lawyer does. Like, well, yes, but there's two giant ifs in that statement. Like you're just kind of presuming you're defining the problem away. You're defining the future of A.I. as something that can do everything and then saying, well, therefore, it can do everything. And maybe which is the point I said earlier, we don't have a good knowledge of what the physical limits of this are. And so if you just presume there are no physical limits and presume that it can do anything, then it's going to be able to do everything fine. Otherwise, let's kind of plant this in reality and say, yeah, you're going to need to build an actual piece of software for this, not just give everybody fraud. And who's going to do that? And who's going to build 500 or 1000 individual pieces of software with support and integrations and compliance and auditing and go to market and everything else?
Speaker 1But this is kind of. Fundamental to how to think about the future for anthropic and indeed open A.I. And if you believe that Harvey and Legora, they've got a big lead. So it'd be very difficult and pointless in a way for anthropic or open A.I. to try and devote resources to try and catch up with them. If you look at sort of my area, finance, well, there is, you know, models everywhere, software applications everywhere in tiny little. Corners, people, you know, you wouldn't even have thought of. So it seems it seems to me it's almost impossible for the model to to the model company to capture all these verticals because it does require a certain amount of knowledge at the start. And once you once you start, once you build a product and you've got customers and the customers are using your product, it's very difficult to make them shift. So this seems to me the case for. Yeah, I would
Speaker 2comment this one. slightly different angle which is on the one hand yes in principle you should be able to ask the model like work out 200 things that a financial analyst would need and then make them maybe what's more interesting as a cap is where does this actually have that information where is the stuff that it needs publicly accessible where is it available where is that implicit knowledge where is it that you want this done the way most people would probably do it because what these systems do is they give you the answer that most people will probably do they do that the way most people would do that and often that's what you want sometimes that's not at all what you want um what are the kind of conceptual limits on what this can and can't do um can you how much do you need to be in the room how much do you need to be in the room how much is this not something that you can automate how much is the whole point that you don't want it done the same way is can you automate it is is it is a kind of a tricky statement because can you automate it is actually asking two different questions one of them is is it physically possible for a machine to do this thing and the other is do you want it done in the same way every time those are two very different kinds of questions yeah okay and so if you say can i automate making the nda one answer is is the nda any good the other answer is do i want the nda done the way most people would probably do the nda and the answer is sometimes yes like sometimes you just want it done the way most people would do it sometimes no sometimes that's not at all what you want and then you kind of spiral out the conversation becomes fractal like is this is the data available who has the data can and profit get the data what does it take to plug that in who is going to be able to deploy this what does it actually mean you know go back to my you know my example of looking at across every system in the company who does it take to plug that in who is going to be able to deploy this what does it take to plug that in who is going to be able to deploy this what does it take to plug that in who is it that actually makes that happen because it's not the cfo is it the cio hiring accenture who is it that works out that that's the thing that you would want to do is it this is it the cio is it accenture is it bane is it a software company who is it that invents all of those things and imagines all of those things and then actually goes and gets done it's not claude running on the cfo's laptop that's one thing you can say with certainty that's not how any of that happens so how is it that that happens and who builds that um who is it that comes to the company and says hey did you realize that this thing could do this for you and again you can propose well you know you just have like the giant world brain and you say yes i'm going to give you access to all of my stuff and you hire accenture to plug it into all of your stuff and then you say great now what can you do and it comes back and like then you get into like a scene out of the hitchhiker's guide to the galaxy or something um but like in the meantime this is a world of professional
Speaker 1services so i always enjoyed this piece um did evaluation of anthropic which came up with one and a half trillion and he was assuming that it would come up with a big drugs company and i'm looking at that and thinking well i i absolutely accept that ai could be brilliant for drug discovery and who's more likely to be successful is it anthropic who's really good at building a model or is it a drugs company is quite good at developing drugs and i i don't know the answer to that do you know that back in
Speaker 22049 i don't know the answer to that i don't know the answer to that i don't know the answer to that because back in 2014 he wrote this very um viral piece saying that uber was massively overvalued at the time yeah and his argument for why uber was massively overvalued was that he said well the town is taxis and what's the market size for taxis and what share of that will they get and first of all the town wasn't the existing taxi market it was completely transforming the taxi market and secondly then they did delivery and ads so his estimate was 2024 tam of 183 billion dollars and uber would get 10 of that so 18 billion dollars um actual gross mobility in 2024 was 83 billion not 18 delivery was another 75 and then freight was five on top of that so i'm not saying he's come up the right answer and i'm not like and i'm not doing this as like a gotcha of like therefore i'm not going to do this i'm not going to believe what he's saying now the reason that i'm because like you know predictions are hard especially about the future as yogi baro tells us the reason i point to it is like you can't value things like that you don't know what the new things are going to be sure you can do a very very high level exercise and you can say well what if productivity growth and gdp did this then what would be the aggregate value of that that doesn't capture consumer surplus it doesn't capture what happens if every law firm buys this and they have exactly the same number of associates just doing 10x more stuff in the same hours um it doesn't capture why is it the number of accountants went up every decade in the 20th century even as we automated all accounting work and why are there more people working at pwc today than there were before spreadsheets where is that in your town like what what what you're having with that where is all the value that pwc has to offer to people who are working at pwc today than they were before is all the value that pwc got out of spreadsheets is not in pwc pnl nor is it in the microsoft pnl nor is it in productivity statistics so you know sitting and saying well you know imagine if they can create a drug and the drug would be worth x and that like i yeah you can make that spreadsheet but i don't know what that tells us i mean it's back to what i
Speaker 1said at the beginning i mean have you have you got a range of of ideas as to where anthropic and open ai will be say in five years time well what would it be
Speaker 2there's a view that says that open ai is is the new netscape that first but then the value went got captured elsewhere i mean the challenge is you know maybe this is why i was a bad equity analyst i hate forecasts um um that you know imagine we're back in 1997 and larry and sergio in grad school mark zuckerberg is in junior high or middle school jeff bezos has a terrible haircut and a baggy sweater and a goofy smile and he sells books you knew the internet was going to be this huge thing you didn't really know where world value was going to be captured and you were excited about yahoo and aol and isps like the biggest internet stocks in the dot-com bubble in europe were isps yep um so you don't know what the things are going to be that people do with this and where that's going to be captured and how and all of the stuff that we ended up doing sort of was on the yahoo home page in i don't know 1999 2000 2001 but it ended up not being yahoo and it was all on like the aol home page and the yahoo home page and that and that all happened but not them so is there some mechanic that says you've got to be a yahoo home page and you've got to be a yahoo or everything we've just been kind of talking about are all the reasons why it's going to be really tough for the stuff that gets built with this all to accrue to anthropic and open ai now you can propose no but actually the models themselves will be able to invent all of that stuff and so and claude itself will be able to make those things in a way that sap was not no i mean those 500 apps sass apps four or five hundred sass apps the typical big company in america has they're all basically something you could have done in sap or salesforce or something you could have done in excel at email and they unbundled that they unbundled that workflow out of email out of excel or out of sap and turned it into a company and went and sold it to people they're all basically thin sql wrappers and theoretically um those 500 new things that will get created with ai no actually claude can think of all of those things and build them so that is different so then you say yes theoretically but actually just thinking of that idea and then making the software isn't the hard part it's everything else around having that software existing that's hard and if you propose this as like this sort of that computer that can do anything then fine but otherwise someone's going to have to go and build those things and that means that you keep coming back to saying isn't this infrastructure and i have a comparison i've made is you know i made the comparison with mobile operators which is all the stuff they thought was going to happen did happen but they didn't get any value from it you could also look at cloud where cloud was infrastructure that was the idea that amazon didn't expect to build all the sass apps and they make good money from that but they don't control the whole thing they don't capture the value of the whole software industry it could also look like chips where um where there's no winner takes all effect per se or certainly no network effect in chips but you could also look at cloud where there's no winner takes all effect per se or certainly no network effect per se or certainly no network effect per se or certainly no network effect but over the last couple of decades it got with each generation it got more expensive and more difficult and i think the rocks law says that the cost of a cutting edge semiconductor fabrication plant doubles every four years i think that's the number crucially it's it's it's slower than moore's law otherwise we'd be in trouble um but um so what happened is he went from dozens of fabric of companies at the cutting edge she just won in gsmc and so um it may be that that's what models look like that it turns out that with each generation they get harder and more expensive and they're not going to be able to do that because they're not going to be able to do and then you only have one or two companies doing it. And they make good money. charge whatever price they want, or they can charge at ROI. May look like mobile operators, which commodity infrastructure, may look like something else. But as I sort of said a couple of times, we can't answer any of those questions by reading the S-1.
Speaker 1This episode is brought to you by Oxford Data Plan. Every quarter, by the time a company reports, the numbers are old news to anyone who's paying attention. Oxford Data Plan track more than 800 KPIs across over 260 listed companies, built from a variety of exclusive and proprietary data sources like receipt data, card transactions, and an advertising agency panel. Key revenue drivers like ad spend, order counts, AOV, and more. Updated every day, 12 hours after the close. It won't tell you what company's worth, but it will tell you where the quarter is landing ahead of your model or behind it, before the print, while there's still time to act. I've been using it myself, and I've been impressed with the depth and breadth of the data. It's been built by people who used to run money, for people who still do. Find out more at bit.ly forward slash BTBS for behind the balance sheet, ODP for Oxford Data Plan, and the number one. That's bit.ly forward slash BTBS ODP1. Baseline is an equity research tool and screening system, which scores 3,600 companies on over 30 qualitative fields. Pricing power, barriers to switching, management quality, AI exposure, and more. It allows you to screen on qualitative traits, low pricing power, and high AI risk. Founder-led names with accounting red flags. You can't do this on conventional screens, and each score has a detailed explanation of the reasoning behind it. If you ask, does ASML have pricing power? Baseline says four out of five because of X, Y, and Z. This saves a lot of analyst time when reviewing a new company or an IPO. Private investors pay a low monthly cost. Funds get additional features and the ability to customize the system to suit their specific needs, including additional coverage, fields, and other tools. To learn more about Baseline, visit bit.ly forward slash BTBS for behind the balance sheet, BASE, and the number one. So that's BTBS base one. So what signals should we be watching for that will give us some clues as to which path we're going down?
Speaker 2So, I mean, I think this is the thing I wrote in my newsletter last night is, um, we have a lot of volatility because all of the numbers are very elevated and all the ratios are very elevated, which makes people nervous. And so people jump on anything that looks like signal. Um, so, um, and when you would know you're really in a bubble or correction, when stuff gets priced in twice, like, well, we didn't, we had that story a week ago. Why again because because we are like deal with it um but you know obviously we saw that with wobbles over google search revenue like maybe this is the moment that everything flips you see it in any every one of those stories suddenly gets leverage applied to it or rather people think maybe this is the moment um and the problem is on the way up you know it goes tick tick tick tick tick and on the way down it goes tick tick tick tick tick there's like i don't what's the equivalent of falling catching a falling knife on the way up right there should be a phrase that means the same point on the way up yeah but you know it's the problem do you want to do new at what point do you want to do you want to step off i don't there's not like some clear answer to that question um clearly we have explosive product market fit for software development right now and then you have cycles you know everyone's saying wait a minute but we're spending way too much on this stuff so we need to moderate our spending a bit but that doesn't change the underlying story we don't have consumer market product market fit at scale for something with real intensity we have lots of people using this a little bit we have a billion billion and a half people who can't who are using this but can't think of anything to do with it today maybe that's agents um um mark zuckerberg on his quarterly call said that muse has product market fit well it's got 10 million users um he says 25 percent of them are using them on any given day which note is not daily active users it's slightly different and so when people don't give you a number or give you a number that looks like a number but isn't that's always always signal um he also said he can't scale it until costs have gone down by 10x which will only take about a year at current rates um so it's not quite product market fit if you can't actually scale it um if we had a thing that looked like instagram we had a thing that where consumers will i know i really really want this all the time then that would be another shift maybe that will happen next year later this maybe that will happen in the next 12 months right now the
Speaker 1closest right i mean
Speaker 2well it is but it is and it isn't because it is actually yes great but it's actually 10 million people 10 million after a month okay give me some charts show me this is really really working and not it's just working for everyone who gets to walk as his podcast um tell me it's working outside the valley um tell me it's working at scale and sustainably and really working uh maybe like at the moment could be we'll see um certainly it's working in a way that open floor didn't work maybe um i mean all of this is to say like you know like the problem is it's like the internet yes clearly this is working clearly this is going to be a huge deal clearly this is going to change everything but where's the value capture going to be who's who's it going to be and we don't know what the usage patterns are going to look like what the apps are what people are going to do with it just you know on muse is the right part of the point of muse is that part of the consumer problem with ai is a blank screen problem you don't really know what you can ask or when to ask it or how and how to form a pattern around that maybe me maybe agents solve that because they expand what you can ask because you can ask it to do a thing for you rather than just answering a question that's good but it's still a blank screen problem is the solution that you make it passive so it watches everything then it can pop up and it can pop up and it can say hey i've noticed your bank account you should move on to this account you should be sweeping your money and you know you should be doing x and y i've noticed you're on the wrong credit card is it that it can watch you and learn and see and so you can pull the use cases through like that if that's your thesis that's medican has some of that google has some of that amazon has a bit of that apple has a bit of that but doesn't kind of doesn't want to use it open ai and anthropic don't have any of that so if your thesis is consumer agents that are watching and seeing all of your behavior will drive a new wave of adoption well open ai doesn't see your email it doesn't see your messaging or your purchasing or your social media doesn't see anything except what you do in chat gpt so which is one reason maybe why they want a device so if that's your thesis that's a thesis for google and meta it's not a thesis for open air and anthropic or it's a thesis where they come at a disadvantage um again we're back in 1997 is it the isps or is it going to be yahoo or is it nescape
Speaker 1these are unanswerable questions have you been surprised that um google search has held up as well as it has and i mean obviously last week there was a bit of a wobble but i mean the growth rate it might not have met any expectations still a pretty good growth rate if you thought back to the chat two years ago that said nobody was going to be searching it was all going to be done via chat gpt i think
Speaker 2so some of this is habit some of this is bubble thinking like oh i'm not going to doing this anymore everyone must be doing this some of it is um what are you trying to achieve if if you're only doing you know using you don't change the world by doing the old thing with the new technology using chat gpt as a search engine is not very interesting and it's not really any better and google can do that too and they can probably do it better yeah what you have to do is get to um not um what's a good credit card but look at my spending and my behavior and work out what credit card i should have it's not here's a coat where can i buy it it's look at my instagram and suggest a new winter coat for me that would change my look but not too much it's actually doing something different it's not doing the old thing but slightly more or slightly better with the new thing and um maybe that's agents maybe that's dots maybe that's muse maybe it's something else is that properly done horizontally or vertically is it properly done within a finance app or is it properly done with an application that looks everything and has my bank my bank account and my purchasing and my travel and my camera roll and everything else who is it that's and who is it that's best placed to work out those use cases persuade the user to do those use cases surface and suggest that who has the visibility on you um you know i used to describe this as like the blind man the problem is a blind man feeling the elephant yeah google and meta and amazon each feel different bits of the elephant and think it's a different thing apple theoretically sees a whole other bunch of it doesn't do anything um open ai and antwerp can't either
Speaker 1with this and um do you have any sort of strong feelings about the the sort of relative winners between the existing big tech companies i mean do you feel that you know the advantages that facebook is getting or meta is getting in in advertising from using the ai is that kind of like a sustainable competitive advantage
Speaker 2well so clearly they you know they've all got like double digit percentage points gains in their ad revenue simply from optimizing part is partly optimizing it's more than optimizing it is going from x therefore y correlation you bought x other people bought who bought x bought y therefore but you might like y to knowing what those products are and the reasons why the people make those purchases um and so that still is still a lot of of of tailwind for google and meta on that and for amazon on that um and then there is a question of what do they have and what are they meta is a company that has always served user behavior what it does is it looks and sees and follows and builds and tracks and pivots its protein it's continually changing based on what people are and what they want and so that's the bet for meta for ai that it's a bet on a company that always works it out and always follows people and works it out for google like you know if this is search this is search google does search um apple is sort of interesting here in the sense that like whatever happens here i'm going to want a device that i have with me that does this who is good at making beautiful devices with a strong privacy story great battery life a great camera and really strong custom silicon sounds like apple to me um so they're sort of well positioned by not playing um or playing choosing very carefully which bits they play in um amazon you know the core of amazon is a network of federal logistics and delivery um the agent the other side of this of course is the other side of the agent story which is why amazon decided to block muse there's two parts to that one of them is a purely tactical operational question which is why all the banks are blocking it which is like where's the liability and the security like we spent 20 years blocking bots because they're not the user we don't know who that is okay well now we're going to let like any random connection from muse connect well we don't know if that's the user we don't know if that's you maybe there's someone sold you someone stole your password they make a transaction where's the liability where's the transaction where's the so there's a bunch of like operational questions around it then there's a strategic question which is nobody wants to be somebody else's dumb
Speaker 1api call true no
Speaker 2amazon does not want a situation where i say to me amazon did 75 billion dollars of ad revenue in the last 12 months they don't want a situation in which you ask somebody else's agent to make you a purchase and it goes on amazon and makes that purchase for you um they want to show you that they want the opportunity customer ownership the upsell everything else and of course this trade-off is old as organized retail like do i want my own shop or do i want to be in the department store well it depends you know do i want the customer ownership and am i willing to spend to build the beautiful shop on bond street or you know fur book santana ray um or beautiful sandwich you know or beautiful santana or do i want to be in the supermarket and it depends like what brand are you what's the product what's the customer um amazon is walmart amazon wants to own that customer relationship they don't want to be somebody else's dumb api call and all of those old questions get repeated over and over again i have a slide about this that says that all ai questions have one of two answers the answer is either no one knows and we've had a lot of no one knows or how did it work last time right are we going to give this startup complete access to all of our customers and our business without anything in return how did that work the last time how did we answer that question last time should we build our own software or should we pay accenture to build it for us or should we buy it from a vendor how do you do cloud depends so i'm going to finish
Speaker 1now i i mean my sort of thinking about this is you you know in the one hand you've got the sort of ball case you've got a magical product with sort of almost unlimited demand efficiency is improving incredibly fast and demand is going to grow to the skies and the computer just gets cheaper and so we all use more of it and and there's a huge profit gap in terms of how much money you can put into it and so i think it's important to look at the market as a whole where it exactly gets divided up i don't know and the bear case is that you've got so much money being invested in just the sheer scale of investment requires an impossible level of revenue and when the market starts to worry about that this the plug gets pulled the cost of money is going up anyway and because the whole the whole edifice is interconnected there's a a vulnerability to it and i think that this financing issue is really because the sheer scale of the money involved means that financing the cost of the money is a critical component yeah so
Speaker 2there's i mean the the interesting thing there is like the spread of who can get capital and some of that is timing and some of that is who you are i mean i talked about you know obviously google has the cash flow if you're a neo cloud you don't so who has the cash flow who has the access to get to that capital is something that's changed this year i mean it's been sort of implicitly tam a couple of times and i've and i've sort of kind of not answered it because we've been most of our time talking about where who would get the money rather than how much money um and obviously how much money is obviously kind of important here um the the puzzle i have here is um and i've sort of alluded to it in a couple of different ways um so um supposing i go and ask well so let's start by talking about behavioral economics how much money do you have money is left on the table as a result of behavioral economics the fact that you don't bother to do that that it's too hard it's too much friction um it's too painful it's too difficult it's too complex you don't bother to move that money into the savings account because hey it's only 15 pounds a month in interest income you are on the credit card that was the highest score on the comparison site rather than the one that exactly matches your behavior you don't cash in that you know 10 a month discount on spotify and if you cash in all of those that would be like a bits and pieces that are floating around how much does all of that add up to and how would you put that if you have agents and ai's optimizing everything maybe all that money just does this maybe it does that or maybe it goes up and you just rechange the price you know which industries will give all of that up and which industries will rebalance their pricing to hang on to half of it all of it where's that in your time calculation for anthropic um every law firm buys ai okay set aside the question of whether they're buying it directly from anthropic or buying it from harvey um how much do they pay what's the pricing power of that what's the pricing value of that what's the roy does it replace 20 of your associates or not um you know if you've been sitting and saying what's the time for spreadsheets in 1980 what number would you have come up with on the one hand you didn't really know what the economic value it was going to be and the answer is in the end you can't charge more than 50 or 100 quid for it and on the other side you didn't know how many people were going to do this because you'd have built your you'd have done your calculation you said well how many cpas are there in the world and you would not have said one and a half billion people will be using excel because of course what actually happens is there's small people of people number of people who do a lot in excel and everyone kind of has it and uses a little bit um you'd imagine you're doing your time for the internet um where's uber in that where's instacart in that um where are all the things that we actually do with this how much of that value disappears as connected to the internet and how much of that value disappears as connected to the internet consumer surplus how much of it doesn't even show up in productivity statistics how much of it just gets competed away and becomes part of the cost of doing business i don't know i asked this question i was in an event last week with with lloyd plankfein you know i said you know in the the 80s you deployed pcs and spreadsheets and yet the number of associates didn't go down by 90 percent and like yeah and you said a third of goldman's employees now are software engineers um and i said well you know i don't know how many people are going to be using excel where's that in your term for pcs that you calculated in 1980 um and so i think they are half of the answer to this is there'll be a bunch of value that doesn't accrue to anybody it's just efficiency and maybe lower prices or consumer surplus there'll also be a bunch of value that is new spend or spend directed from different places but you don't know who's going to capture that and i've we've argued at some length that it's probably not the model companies but hey you know it could be completely wrong and it could be all go to the model companies sure yeah sure and like i can make a slide that says well this is the value of the global professional services industry this is the value of the global money transfer industry this is the value of the global software industry is a trillion dollars okay the global professional services industry is again make up a number the global marketing industry is a trillion dollars global retail is 30 trillion dollars 25 trillion dollars which percentage of each each of those do you think it's going to be and this is what gets you this kind of absurd thing what you've kind of looked absurd where spacex said our term is like global gdp or whatever it was or half of global gdp which sounded kind of stupid but what they're actually saying is well all of that is money that some people are going to do some stuff with with ai and so pay something for no like what was the term for the what was the term for the internet global gdp yeah but it didn't mean that like all of that went to the isps or yahoo
Speaker 1Of course, the difference this time is that there's a physical dimension. the physical world is being disintermediated. We haven't talked about robots at all.
Speaker 2Is that part of the term philanthropic? Well, kind of, but not really.
Speaker 1Well, I mean, it's such a big and fascinating subject and it's great to get your perspective. We probably need to do this in six months to twelve months time and revisit it because it's so rapidly changing. We'll see if any of these questions have changed. We'll see if some of these questions won't have changed at all. I bumped into a friend of mine, a client who works for one of the big pod shops, and I said, oh, I'm interviewing Benedict on Monday. And I said, what should I ask him? And he said, I don't know, he said, it changes every week. And that, you know, at the sharp end in the stock market, the issues are literally changing week by week. But it's still kind of you to come on. Tell listeners and viewers where
Speaker 2they can find you. Sure. So I do a weekly newsletter that's got 140,000, 150,000 subscribers and there's a free version you can sign up for. And then I write essays and I do a presentation twice a year, which I'm actually about to publish my new presentation. And then people pay me to come and give presentations, give that presentation to them, their investors, their board, their clients and so on. So yes, I'm on airplanes a lot this year because a lot of people want to know what this AI thing is. So yes, Google Benedict Evans. Google Benedict Evans and you'll find me.
Speaker 1So listen, thanks very much. I wish you safe travels and we shall reconvene. Yep. See you soon. Thanks a lot. I hope you enjoyed that conversation with Benedict Evans. What struck me most was the distinction between the potential for the enormous economic value AI could create and the much more difficult question of who will capture that. We could continue to see extraordinary growth in demand for AI with falling costs and new applications transforming entire industries. Yet the companies building the models and infrastructure might still struggle to earn attractive returns. I thought Benedict's comments on the comparisons with the early Internet and mobile telecoms industries were particularly interesting. And of course, his thoughts on the economics of anthropic, useful lives of GPUs and the potential commoditization of AI models were all really insightful. You can find Benedict's work at ben-evans.com and I'll put the details in the show notes. If you enjoyed the episode, please subscribe, share with a friend or colleague, and leave us a review. Thanks for listening. I'll see you next time.
Speaker 3Behind the balance sheet and affiliates and podcast guests may own shares or have an economic interest in securities discussed in this podcast, which is aired for your education and entertainment only. Nothing in this podcast should be construed as investment advice or relied upon for investment decisions. Always do your own research. Thank you for listening.
Speaker 1I hope you enjoyed this episode. It would mean a great deal to me if you could please share this with just one friend and feel free to spread the word by subscribing or even leaving a review on Apple Podcasts or Spotify. Thank you for your support.
PodcastTranscript AI