Speaker 1and he said that he thought there was a shift from information asymmetry to imagination asymmetry and what he meant by that was you have to have the imagination to know what to ask welcome to the angle
Speaker 2from tiro price a podcast for curious investors just a reminder that outside of the us and australia this podcast is for investment professionals only thank you for joining me today i'm justin thompson head of the tiro price investment institute my guest today is technology analyst benedict evans benedict has seen various technology waves at close hand in a 25-year career on the sell side as a strategist in the telco industry and as a partner in venture capital as we talk about today's ai investment boom we are going to explore why this matters for investors and why this matters now benedict welcome thank you um let's start with you what motivates benedict evans uh
Speaker 1boredom and curiosity i suppose um i think i mean it's a general answer i did my degree in history which is read a lot and try and work out what you can say that's interesting and different and that's what analysis is in technology the point that you understand something is generally the point you should be looking for a different question and so you kind of have to keep moving and asking well what's the next thing um what's the next puzzle and so it's kind of nice to have some new puzzles to talk
Speaker 2about interesting because part of our branding is all about curiosity and ask better questions get better answers um but i've also noticed in that that you have a low boredom threshold so i hope i don't bore you um there is a saying this is totally different just like the last time so i know every technology wave has its differences but what period in history um you can allude to recent waves of technology or you can go back to railways canals yeah is the closest analog to where we are today
Speaker 1um i think from a technology rather than a financial point of view ai today looks a lot like the late 90s in that this is very exciting and it's clear it's going to change everything and you have people getting into a sort of state of euphoria um that the entire world is looked going to look completely different in like three weeks and you have 25 year olds who never really done anything except work and then they're like oh my god i'm not going to be able to do this in computer science and work for a tech company who suddenly imagine that they understand the whole of the rest of the world and can know exactly how everything's going to work in places that they've never been and things they've never seen. Within the technology, it's amazing and exciting and it's clearly it's the next big thing. But a lot of it doesn't really work. It just kind of sort of works yet. And it's not clear what the actual strategic building blocks are going to be. So, you know, go back to the mid, late 90s. Mark Zuckerberg is in middle school. Larry and Sergei are students. Jeff Bezos is a bookseller with a bad haircut and a goofy grin and no muscles. And it wasn't clear that AOL and Yahoo were going to disappear off the face of the earth and that there was going to be this thing called social. It wasn't clear how any of it would work. And I think that's kind of where we are now. It's very exciting. It's not clear how any of this is going to work. The obvious way that this is different, I mean, there's probably maybe two ways that this is different. One of them is that it looks like this happens much faster, which I think is slightly deceptive. The cloud is already there. The data is already there. ChatGPT from a consumer perspective is just a website. So, of course, you can get hundreds of millions of people using it straight away. I mean, it's an unfair comparison, but it's kind of a relevant comparison. However, it actually is taking a lot longer than that to get to real usage. So, you've got very wide adoption, but very shallow adoption. Every company has deployed Copilot, but that doesn't result in wide-scale automation or changes in how anybody does things. So, it looks much faster superficially than it really is. I think the other fundamental way that this is different from the dot-com bubble, or indeed other things like the growth of mobile -- and remember, mobile didn't come with a bubble. It came in 2000, and then mobile actually happened in 2010. But the other way that this is different from PCs, or the internet, or mobile, is that with those, we didn't know exactly how they were going to work, but we knew what the physical limits were. Whereas with this, we don't actually understand the science of AI, really. We don't know why these models produce such good results. So, there's a sort of fundamental uncertainty in how this will evolve that wasn't the case with previous platform shifts.
Speaker 2I think that moves us on nicely to talk about the technology. And I don't know whether you're re-quoting someone, but you quoted that a frontier LLM is the most expensive and fastest depreciating asset ever created, which would call into question the economics. I actually think that you think it's more nuanced than that, but let's talk a little bit about the economics of the model here.
Speaker 1Yeah. So, with other of these, is it the same or is it different? And there's a certain kind of person that loves to talk about fiber here, where we had obviously, we had this huge over deployment of fiber in the dot-com bubble. It's not like that in the sense of that was built out way ahead of demand, whereas this is being built out way behind demand, which is really kind of important to understand.
Speaker 2Did we know that ex ante?
Speaker 1Yes. Well, you can argue about this. So, there were these claims about how fast the demand was growing that turned out not to be really true. But clearly, today, we have massively more demand than we have supply. I think the challenge here is, the advantage of the fiber comparison is, you've got so many unknowns in trying to forecast this, so you can kind of try and model out, okay, this is how much data center is going to get built in the next couple of years, how much is going to get lit up with power. What chips are going to go into them? What the token per watt is for doing that? And therefore, you can pick a number of how much kind of underlying fundamental capacity is going to expand. And you can pick a number, 50x, 100x, 200x, whatever the number is. Then you know that we've only really got product market fit for one thing, which is software development. Then you have a few other things around the edges of that, but are much smaller. We do not have broad consumer product market fit. Yeah. We don't have a thing like Instagram that hundreds of millions of people are using a lot every day to do with LLMs. So we've got this explosive demand, which looks like it's at $50 or $100 billion run rate revenue just for the two model labs. So it's not unreasonable to say, well, yeah, you could have way, way, way more demand quite quickly if you had other things start working. So you don't really know how much more capacity we're going to get. You don't really know how much more the demand is going to grow. It's plausible that you could have demand ahead of supply for quite a long time in investment terms for like multiple years going out, outside of the framework of capital markets or the business dynamics. Forgive
Speaker 2me, but I'm going to pin you against a wall. That wasn't the metaphor I first thought of. I. So we talk a lot about scaling laws. By that in frontier models that the capability and performance of the model rises in proportion to the investment in the compute. Is there any physical or indeed practical limitation to those scaling laws?
Speaker 1I can give you a really, really short answer. We don't know.
Speaker 2But I'm pinning you against the wall.
Speaker 1We don't know. I mean, this is the thing. So you can hypothesize some practical limits. So you can say, well, at a certain point, we run out of training data. And that's kind of a jagged statement in that there are some fields where training data, there's much more training data. It's much easier to get where synthetic data works. There are other fields where it maybe doesn't. So it's much easier to do synthetic data in coding than in writing, for example, or in video, say, than in writing. Some of the approach in the last couple of years has been shifting from just scaling just more training data to having the model think longer, which is reasoning, and then having lots of models, which is a sort of agent swarm thing that's happened in the last six months. So there are things that you can do that can expand the capability without more-- without scaling law. So I mean, for the sake of argument, if you were to say, well, presuming the model had stopped in January, you could have still taken the model progress that had stopped in January and say, yes, but if we throw you 10,000 of these things-- Yes. --we're still going to get a much better result. So there's a bunch of different engineering and deployment ways that you can get better results and use more compute, even if the underlying scaling doesn't take care of it. keep going keep carrying on and there's a bunch of ways in which you can say well it looks like in that field maybe that isn't going to continue but the underlying issue is that because you don't know how this works at a theoretical level all of these conversations become hunts for metaphors which is why the doomers start talking about you know imagine if there was a meteorite or imagine if this was a nuclear weapon because you don't know and the metaphor i've always kind of liked is it's sort of like you're building the apollo program when you don't have a theory of gravity and you don't actually have a theory of the chemistry of rockets either like well but you made it twice as big and it went three times higher okay so what if we make it twice as big again
Speaker 2um one of the great things about tiro price is we have a genuinely global research platform um so we're able to draw comparisons between what's going on in china and what's going on in the u.s i mean there's you know there's an awful lot to talk about in europe unfortunately but i'm interested to know what your interpretation of how this is all going to work do you think all these models are going to be fungible or will the chinese and the u.s ecosystem evolve separately um i think everything
Speaker 1is still wide open and everything is to play for you don't want to depend on one company you don't want to depend on dario or sam deciding what you're going to be able to do so every company i talk to is doing multiple sourcing i think every product and use case is making a decision around but do i need a frontier model and if you are doing um a particular point solution for fraud detection inside an e-commerce apparel company you're going to choose the right model for that and you're going to look at the cost and the reward and that might be open it might be closed it might be last year's anthropic model running on a neocloud it might be a chinese model it might be a chinese model it might be a new cloud it might be a chinese model it might be both i think you have a lot of fragmentation along along that price performance curve um and then a conscious decision that we're not going to depend on one um second answer is is a pure geopolitics question which is um on the one hand america is determined on decoupling and on the other hand it looks like china is determined on decoupling and china doesn't want to have that dependency on american technology either they don't want to to risk being cut off um the lack of access to chips at the same time scale that the u.s labs have and a much more kind of pragmatic less winner takes all less um driven to make agi kind of culture means that the chinese labs are not pushing for absolutely the best possible lab result every day instead they're pushing for you could argue what they're actually optimizing for is let's be a month behind the frontier and a tenth of the cost but clearly we have a kind of a decoupling dynamic there that points to the fact that the u.s labs are not moving towards multiple models multiple choices for quite a
Speaker 2long time um okay i want to move this on to the capital cycle because i don't i you know i think this is in terms of market positioning and asset allocation this is the topic that we cannot ignore um what fascinates me about this capital cycle is how commonplace trillions has become whether it's trillions of capital employed whether it's pre mark pre ipo valuations whether it's the number of trillion dollar market cap companies we have i mean trillions has become common you've got a very visual metaphor that i really like which you've described the current cycle is like a pig in a python i.e we we are in the process of swallowing this enormous amount of capital but actually digesting and deploying it is is going to be a little bit more difficult is it possible to deploy capital at this
Speaker 1scale um this quickly clearly it's difficult um the tech industry i mean from memory i think the big four platform companies were spending like 100 billion dollars a year on capex as we came through the pandemic this year their guidance is something over 700 billion dollars this year and that probably numbers kind of crept up in the course of the year and if you add oracle and neo clouds in china you get close to a trillion dollars and then most people think that's numbers going to go up next year um so you have this enormous surge in the semis industry and a pricing crunch in the semis industry so the revenue has gone up a lot more than the volume in semis the semis industry has always been cyclical so they were always super nervous about increasing their capacity and it's really only this year i think that they've really started increasing capex to fill this um and then you have challenges of every bit of the supply chain around data centers down to like obscure power electrical equipment that you know none of us had ever heard of and there are all sorts of companies suddenly spiking because they're getting you know the business has suddenly doubled to build all of this stuff down to things like gas turbines where um you know from memory the whole industry does like 75 to 100 units a year that this of these kinds of things and suddenly they're being asked to double production and they're like yeah it doesn't really work like that um and then of course you've got the whole political issue particularly in the u.s which some of which is actually about data centers some of which i think is not actually about data centers it's about other things it's border political issues um so the short answer is yeah yeah it's hard to deploy this stuff and of course power more than anything else it's hard to actually get enough power into these things this quickly um that gets me to the conversation we're having earlier that like you know it's easy to say shortages are followed by gluts um it's important to remember how narrow the adoption is and how much more it could go up quite
Speaker 2quickly to my mind there are three ways this capital cycle could come to an end one is nash equilibrium the ending the nash equilibrium where everybody today their optimal strategy is to invest as fast as they can and it's in nobody's interest to to break that prison's dilemma prison's dilemma what what of which is all part of game theory to it could be regulation whether it's on the bill physical build of data sense or on the regulation about or it could be capital markets just capital markets go on strike what you mean in in investment in markets we're dealing in probabilities i would
Speaker 1downweight the middle one of those just given the nature of u.s politics and i'm not a politics analyst by any stretch but it's it's hard for me to see um consensus shifting like the u.s political system is unlikely to do that quickly um the so the first is like does something happen that we phrase this a different way does something happen internal to the tech industry the results in the investment slowing down and secondly does something happen external to the tech industry um what's happened in the last certainly this year arguably like course of last year is that the capital outgrew the cash flows of the hyperscalers so certainly in 2024 this was coming out of cash flow because these were very cash generative companies um it's now reached a point that i think meta and um microsoft and um alphabet are all at or above 50 percent cap ex to sales this year and will probably hire next year amazon a slightly different story because it's a much lower margin business so they've outgrown it and so now they are dependent on you um now now they are dependent on the capacity of the cap unwillingness of the capital markets to continue providing the capital markets and to continue providing the capital markets to continue providing that in all the ways that that means and that's debt and private equity and public equities and all sorts of other stuff um so that's one piece um and i'm not going to have an opinion on like what's the capacity of the capital markets to absorb this that's for other people to talk about i think the other side of this is um do we get a paper tomorrow that says hey guess what i get 5x results for a tenth of the compute seems unlikely but given we don't have a good theoretical model i'm not going to get 5x results for a tenth of the compute i'm going to it's possible what's happened one of the things that's happened this year is a sort of agentic is a um agentic and be a sort of agentic swarm if i can make that distinction which we talked about earlier in that what's happened with like maths and what's happened with cyber is you just take 10 000 of these things and throw them at the problem and so you have an ability to use way way way more compute and get better results it's kind of back to our scaling point maybe the models don't get better but now you throw 500 models or 5 000 models or 50 000 of the same model at that problem and that gets you better results um so there's all sorts of questions around this of like well what happens to the demand and the reason i say this is it's perfectly possible that we get to a point that actually we can absorb way more demand without needing as much capacity as we thought i mean one of the things we haven't talked about here is that the marginal cost problem shapes consumer adoption in that what we're sort of used to from the internet is that you could make it free go viral get a bunch get tens of millions of users and then work out the revenue because the cost was kind of could be absorbed within the funding structure of traditional venture capital that whole playbook of make it free because mod marginal cost is so low and then work out the revenue later doesn't work with AI in the same way. Because marginal costs are very high. Because marginal cost is high. And so this is what you see at OpenAI is, and this is part of the thing that shapes public perception of this stuff, which is that if you're not paying, you're not seeing today's technology, you're seeing the technology from a while ago because they can't afford to give you unlimited access to today's technology for free. And so what is the number like? Something like 5% of chat GPT users are paying. So most people are using a relatively limited model, not very much. But this is kind of the thing I keep coming back to is we don't have mass consumer use to drive demand, to really, really drive demand for tokens. Because it's almost kind of circular. The industry couldn't afford to provide that. For free. But whatever you want to say about it, Clay, right now, we don't have something where hundreds of millions of consumers are using this a lot.
Speaker 2Okay. Well, in which case, let's move this on from straight economics to I would like to get a window into Benedict Evans' mind as a dreamer. What are you most excited about here? A, B, you know, we focus. There's a lot of attention on the supply side and less on the demand side. And, you know, what are the consumer applications or the services that you think are going to be the ultimate drivers of demand here?
Speaker 1So, and we talked about a little bit about this before. I think maybe sort of two answers to this question. One of them is I was a. Working at a mobile operator in the mid-2000s. And there was kind of a joke that you would go to a conference and people would say, oh, you'll be walking past a Starbucks and your phone will know and you'll get sent a coupon by location services and you'll get sent a coupon for a free coffee. And this became a joke. And I actually got some research reports from that period that had like forecasts of how much money mobile operators would make charging you every time you did a location lookup. And meanwhile, nobody noticed taxis. And we had everybody, everybody had a phone with a camera. And it didn't really occur to anyone that that was a video camera as well. And it didn't occur to anyone that people would post those pictures publicly because social media kind of wasn't a thing. And so, like, you don't know what these cases are going to be. I mean, this is a kind of tech version of the lump of labor fallacy. Like, you don't know what the new things will be. Human needs are infinite. So, people will create a whole bunch of new stuff, doing things that you couldn't do before LLMs, just as we had with mobile, just as we had with mobile. And so, I think that's a really good point. And I think that's a really good point. The internet, smartphones, that's one answer. Unhelpful answer, but like, it's a deterministic observation. There'll be a whole bunch of new stuff that looks dumb, that we all love, that will be created with this technology. Second answer is the things that I'm trying to think about are, where does an LLM, what are problems that we can see now that we think an LLM can solve for a consumer? And I have a sort of a, like, maybe four bullet points from, firstly, like, automation of boring stuff, difficult stuff, stuff you can't be bothered with, which is our conversation about bank accounts. Secondly, stuff that is inertia, bank accounts, I'm on the wrong health insurance plan. I'm on the wrong, I should change this, I should change that, but I can't be bothered. Is that done by the LL agent? Is that done by a dedicated product? Is it somewhere else? How does that work? Thirdly, like, friction, where it's like, friction and complexity, like, get me onto the right credit card. And I think that's kind of different from inertia. I think there's automation, and you could argue about this, but there's, like, in automation and inertia and friction. Purchases, and I suppose what I'm getting at is, like, the things where you should do that but don't, and the things where it's not clear what you should do. I think that's a useful distinction. And so maybe you'll be on a different credit card if you had an agent that could help you work out the right one to be on. And then the other extreme is my coach example. It's, there is so much product. There are so many things you might like, how would you know? And there, we now have systems that might be a much better way of showing you the things that you might like. You know, you didn't know that you would be interested in that book. You didn't know that you would really like that chair, that lamp, that cushion, that coat, that scarf, that pair of shoes. Now, that systems will show this to you. And the conversation in Cannes Lion every year for the last three years has been, you know, I asked ChatGPT, what X, Y, Z should I buy? What will it tell me? And the guys at Google, like, well, sorry, the people at Google put their head in their hands and say, it's just SEO. It's the same answer. Like, what, you know, that's an SEO conversation. Much more interesting to me is, my previous point is, it might suggest different kinds of things, like the shape of the distribution curve, the distribution of what gets bought might look different. The curve might get steeper or it might get shallower, might get flatter in different product areas. And so, just as we had, like, the wave of D2C on the internet, consumer internet, we might have, like, different kinds of things that are now possible to sell, possible to go to market, possible to get people to buy, because agents can walk you through it, or AI systems can surface those things and suggest those things to you.
Speaker 2I think we've got a fantastic window into your imagination of possibilities here. But here's me going to push you up against the wall again. So, you're back as a partner at a venture capital firm. And I want to know what's going to be in your picnic. Where are you going to lay the bets? Where do you think the value is going to accrue in
Speaker 1all this? So, before I go to that, I made a note to make this point, which is used word imagination. I had this conversation with the CEO of a software company on Monday. And he said that he thought there was a shift from information asymmetry to imagination. Imagination asymmetry. And what he meant by that was you have to have the imagination to know what to ask. And that's a problem for a model lab, because you're asking the users to have the imagination to think of what to ask. And it's an advantage for an entrepreneur, because the entrepreneur is having the imagination to think, oh, this is a thing you could solve with an LLM, and I'm going to build a product around that. And I'm going to show people and I'm going to build the tooling around that. This is a thing that you can do. It's also a, you know, it's a problem for you. Like, OK, the LLM can now analyze all this stuff in a different way. Great. What do I ask for? Other than like, what's giving me what stocks are going to go up?
Speaker 2Which seems a little bit too easy. Well, it's interesting because Dan Gardner, who wrote a very popular book called Superforecasting, the point he made on one of our previous episodes of The Angle is that in a world of abundant answers, which is the world of AI, having the right questions is going to be key. And I think you've made just made the same point in another way. But let's I want to get back to the point. What's going to be
Speaker 1in your VC picnic? So maybe a couple of levers here. One of them is where is AI getting at some fundamental point of leverage in that industry? Is it unlocking something? And you can kind of back test this again. It's against the Internet. Like there were some places where the Internet didn't change very much. And if you'd said, you know, we're going to be an Internet enabled law firm or we're going to be a PC enabled accounting firm, that wasn't I mean, that was maybe an opportunity to gain market share. And if you were the first of your peers to deploy spreadsheets or your execution on that was better than you could take share. But that didn't fundamentally change the nature of law firms or the nature of accounting firms. If you were the first airline to do online booking, you probably gained a bit of share that year. Twenty five years later, who knows who was the first person to join to do online airline book online flight booking? Who cares? So the question is, does the Internet get to some fundamental point of leverage? Does AI get to some fundamental point of leverage that changes how you can do this whole thing and you can pull the whole thing inside out and do it in a completely different way? This is the thing, you know, that and if you go and look at, you know, I mentioned earlier the typical big company today uses four to five hundred SaaS apps and each of those are some entrepreneur realising that there was this problem inside the compliance department at Long Only Funds. There is this thing that HR departments at universities do. Somebody has to think of that and work out this is a problem and this is how we're going to do it and this is how we're going to solve it and this is how we're going to get everybody to use it. And that's 500 SaaS apps. Inside a big company today. And so it's one at a time, you know, is that and this is like the kind of the sort of fundamental venture problem, which is very different from like being a long only fund is it's not what are all the ways that this could go wrong. It's you suspend disbelief and you say, could this work? If it did work, what would that be? Is this a team that can make that work? And every VC now isn't looking at anything that isn't doing AI. But then you have to ask all these questions. And that's industry by industry, team by team, product by product. Have they identified some problem that is really big and important? And could it work? Have they worked out the right approach? Are they going to be able to work out the right approach? Because that's really what you're betting on.
Speaker 2This is intended to be an instructional series with some element of practicality. So we're going to close this. And I would like your view on one skill that investors should learn in the age of AI and one habit they should unlearn.
Speaker 1So there's a sort of super practical point, which is try all the models all the time, throw questions at them. But I think a more useful thing is that imagination asymmetry. What are the completely different questions? What happens if you let go? And this is maybe something we should have spent more time on earlier. But what happens with each wave of technology is that you shift the level of abstraction. And the questions that used to be important stop mattering because the new technology just does those things for you. And instead, there are new questions, new layers of abstraction that become possible that weren't possible before. And so, you know, imagine you're building a using a CRM and classic CRM. You've got this very complicated data schema. You've got hundreds of different field types and you are a great power user when you've constructed this elaborate query that lets you surface all these different things. And now you build an AI native CRM and you build it as a system of record. And maybe what you do is. You just drop every gob of data down on a timeline with a date and a client and no other field type. And like client call, app analytics, meeting notes, any piece of data that you have around that client is just dropped onto a timeline with the client name and a date and nothing else. And on one level, this is insane because now you can't run your 11 field SQL query that says show me all clients where these kind of things are. And you can't run your 11 field SQL query that says show me all clients where these kind of things are. People are senior salesperson spoke to them and they're about to exist and they have this client value and they have this usage data and their contract is up in this period. And then that tells me who we need to make sure we call when we sit down in our Monday morning meeting. But you can. But it means you can also say what are our most important clients who are likely to churn soon. And that's a really. I feel like it's a hard thing to let go of all the knobs and switches that you kind of can't ask that other kind of question. But now you can ask this new kind of question. And that was implicit in a lot of the stuff I was saying about friction and Instagram and so on. And it's the imagination asymmetry of thinking where I could just ask that question. There's a there's a concept in in military tactics of mission based. Tactics which is instead of saying to that lieutenant you need to go to this grid square and you need to march to that bridge and secure that bridge. Instead you say OK the objective of this entire division wide operation is to achieve this and you have these resources and this is the other people go do it and you empower your most junior people to go and you tell them what the mission is. And it's their job to be empowered to go and do that. And so here the point of my sequel analogy is is your objective of head of sales to know the right sequel query. Or is your objective OK. Which bits of our mission aren't working and which bits are and how do we achieve that. And I think that's really hard.
Speaker 2So I'm going to I'm going to read that by saying. You know what you've got to give up is some of the. Metrics some of the comfort metrics that you've been used to working with and really figure out what really matters in all this.
Speaker 1Yeah it's you know what is the stuff that you do every day and you know how to do and you know how to turn the handle and you know the knobs and switches and you know some obscure knobs and switches no one else knows around knows about and you're comfortable that you've seen every single field and that you know you know that you have eyeballed every single meeting. That meets the following criteria and maybe it hasn't occurred to you that maybe there are other things that are important. Where the answer to your yes no question is no. So it's not on the search result but maybe you should have seen that one too.
Speaker 2OK. I'm going to I'm going to say this conversation could be infinite. Our time is finite. At which point I'm going to say Benedict Evans thank you so much. Thank you. And thank you for listening to The Angle. We look forward to hearing from you. We look forward to your company on future episodes. You can find more information about this and other topics on our Web site. Please rate and subscribe wherever you get your podcasts. The Angle. Better questions better insights. Only from T.
Speaker 3Rowe Price. For more information visit us at The Angle.com. Thank you. Thank you.