Taking Stock - After The Bell Podcast: AI: Boom, Bottlenecks and the Next Big Bet

Sounds about right. Predates me.

Uh, 20...

Formative years, long time ago.

Well, I think we're all sort of the sum total of our experiences, aren't we? We all carry, certainly investors carry tons of scar tissue. Um, and I think, uh, you know, it was an amazing time to start out in the industry. Um, my colleague at Polar, Nick Evans, uh, started out at a similar time, could be a little bit later. Um, and I think, well, I guess I've been very lucky in that the AI cycle, we'll come back to obviously, um, is probably the second kind of critical moment or key moment in technology, maybe human history, the first one in at least my life, uh, being the internet. So, I mean, there's some very, very strong parallels between what we're seeing today and what we were, um, excited about in the mid-90s and the building of the commercial internet before the years of the dotcom meltup. Um, what have I learned? I mean, I guess, you know, you tend to do most of your learning, certainly as a younger person in this industry, um, during difficult periods, um, and I think...

Um, that probably has left quite a lot of, um, scar tissue actually, um, that period of sort of overexuberance and then how it ultimately played out, how much lower, um, you know, where a value investor thinks there's value being quite different to where a growth investor thinks there's value, watching actual companies come public and disappearing within 18 months is quite an experience. And so again, I wouldn't really wish it on anybody, but, um, but having it bookended, um, a very exciting period prior to that, um, and, you know, that doesn't kill you, makes you stronger. And I hope that, um, that period has made me a better manager. It certainly has helped us understand some of the parallels this time round. I would say on balance, actually, the other thing it has done is I've ended up probably carrying more liquidity in the portfolio over the 20 years that I've managed it, probably shouldn't.

Um, and maybe there's, you know, maybe that's the actual, you know, what scar tissue actually has looked like. So, you know, we obviously have the ability to gear, we've used gearing on occasions, um, but I think that perhaps that 00 to 03, 06 period was, you know, maybe left a little bit more scarring than I would have liked. And then we went through it again, didn't we? I mean, obviously, as I said, I started in '05, '06, and then it wasn't long before we were into financial crisis, getting hit over the head, Europe, you know, we didn't really come out of that until 2012, 2013. And even then, you know, I do remember being, it's all fake, it's all, we're all going to collapse again. This is awful, you know, but I think you're right. You know, you never regret having dry powder and liquidity available when the proverbial's hitting the fan because that gives you optionality.

100%. And also, I think you just, you know, there's so many great things to say about the tech sector, but, you know, on the way up, one of those great things is just operating leverage and just these are wonderful businesses with incredible, you know, incremental margins. And unfortunately, that works both ways. Um, and so again, I was never really an investor in companies that required capital. I think, you know, listed markets do a very poor job of pricing assets like that. We just don't do binary. Haven't done for my entire career. But, um, but, you know, in that period, had you been invested in companies that required capital, that they went to zero, um, companies that were, you know, didn't have balance sheets and so on. So I think that has definitely left its mark. And, you know, again, today, for example, we don't invest in companies that we believe are going to require more than two years to get to cash or break even. So there's an actual example of what an earlier period, um, can do to help you refine the way that you invest. And of course, that means that you can miss stocks, but I think it's, you know, again, important lessons from price.

Yeah. I mean, I would liken them to Yahoo coming public in '96 or eBay coming public in '98. And I think that it's very easy. There's so much bearish narrative around, say, the AI cycle because it's happened so fast, so furiously and has left a lot of people sort of wrong-footed.

So, I think that again, large IPOs obviously going to sort of suck liquidity out of markets perhaps over a short period of time or compete for assets in a way that they wouldn't be if they weren't there. But that doesn't mean that it's the end of a cycle. These are companies that have been private, not the large language models obviously or the frontier labs, but certainly SpaceX has been, you know, a meaningful business for a very long time in the private domain and there are lots of others out there that you know about that have stayed private for longer. The companies that came late in the late 90s were companies that had literally been incorporated five and a half years before IPO. There's a very different cohort of companies than the ones we've seen so far.

Yeah, I think it's a very good synopsis of why we are where we are. And I think, um, you made the point about how well tech has done over time. And I think that ultimately we've... you have to make a good few decisions in your life, and one of mine was obviously becoming a tech specialist more than 20, 25 years ago. Um, but, you know, why has tech performed so well? Because it's grown its share of GDP, because it's delivered earnings well in excess of the market. And, you know, the things that we used to spend time doing and profit pools that have been basically, you know, redistributed, as a kind word, um, away from newspapers and, I don't know, high street shops and into the profit pools of tech companies. And, you know, if you look at the returns in that recent period that you showed earlier, I mean, most of the earnings growth in global equity post the financial crisis has been driven by tech. Yeah, I think that's the elephant in the room. And, you know, um, that doesn't have to continue indefinitely. Of course there can be things. In fact, one of the themes that we're excited about, maybe, you know, referenced resources earlier, but seems to us that one of the kind of key takeaways from what AI promises, um, and we're already seeing beginnings of, I think, is this idea of digital abundance. And that juxtaposes very interestingly against physical shortage. Lots of people are making that call, but I think, you know, most people haven't lived in a world where there's been physical shortages of anything. I mean, we have during the pandemic, of course, but I mean, in investment sense, we've lived against the disinflationary backdrop. Services, you know, service-side inflation has always been outstripping goods inflation for a very long time. Um, and again, we wonder if that might be coming to a sort of a bit of a crossroads with obviously... [clears throat] this idea of abundant cognition and AI, um, could have a similar impact to what happened when the Chinese entered sort of global manufacturing, what you saw with goods deflation, disinflation, deflation. So I, you know, at the same time, we have to physically build assets against, uh, against AI demand.

Well, first and foremost, I mean, this is a technology company, a technology trust. So there's only so far that we can strengthen our mandate or at least try to move the borders of where it is and isn't acceptable to invest. And, you know, a few years ago, um, in fact, more than a few years ago now, our AI team, so we have an AI fund within our team, which is run by someone else. I'm a named manager on it, but it's a global equity portfolio. And so its remit is obviously to search through anywhere of interest where AI is having an impact. And, um, it identified the managers in, we identified the power sort of complex as being a potential bottleneck given how power-intensive AI is relative to, you know, old-fashioned serial compute. And that led us to companies like Gino, you know, Siemens Energy or whatever it might be. We've got lots and lots of exposure in power. So that's an example of something that's physical. It can't be, you know, created with code. Um, but it's, you know, it's not a metals business and we're not in the mining world either. But, yeah, I think that, um, if you looked at the gross margin profile of the portfolio as a proxy for this idea of rotating away from software or the sort of this idea that, you know, from all of my career really, apart from in that early period, um, hardware has been commoditized. It's the workhorse. It's the bit that everybody else sits on top of and builds wonderful business and uses as commodity hardware. You know, that's maybe not been true for Apple's iPhone business, but it has been true for most parts of the world that we touch. Hardware has been commoditized. All the value has been in software. And that seems to be just flipping on its head here, and again we can come back to it. But I think in the end, you know, software is, um, software is built on code. Code is being challenged by AI. Again, we can come back to it. So, yeah, to the question, you know, areas like power, um, transmission, cabling, um, those are companies that have meaningfully lower gross margin. So it used to be a 70s-type gross margin profile portfolio and now we're probably at the very low end of the 50s. So we brought our gross margin profile down a lot to sort of make sure that we're not on the wrong side of that idea of abundance and scarcity.

So, I mean, 25 years into the industry and thousands of company meetings later, I think it's fair to say that no company will ever explain what it is they're doing or at least why they're doing what they're doing. Um, and I think at face value, um, the buildouts of, you know, the hyperscaler capex is, you know, really designed for two things. It's designed to, in most cases or many cases, build frontier models, i.e., the next model that's going to move the whole AI story forward. Um, again, we come back to the sort of model progress and scaling laws and what have you, if you're interested. Um, the other reason, obviously, is to do the inference, the kind of actual, um, the work that you carry, the AI workloads. And, you know, I think in prior years, um, in fact, we describe this in, um, lots of our documentation, um, read online, um, but we describe it as sort of, we feel like we're in a sort of phase two of the AI story. Phase one was very much the build it and they will come period where there really was an absence of revenues and all of the build was associated with building models.

I think there's, I mean, again, there's just, there are some, you know, truly brilliant individuals that were there at the time of the transformer at Google and elsewhere, and that they've, um, they've, you know, they've understood what the transformer has meant for the, um, trajectory of AI model progress. You know, AI had been making progress, um, for 60, 70 years, but the transformer has allowed sort of a completely different approach. And this is something that Google, um, a white paper Google basically highlighted. It's just changed the whole trajectory of AI model progress. So yes, tech bros, but also, you know, just, you know, brilliant people at the coalface. And I think that what was very interesting in that early phase one and why we, you know, why we were sort of maximum bullish, describe ourselves as AI maximalists, is that we lent into what those people were telling us. And we were listening to not just Sam Altman's diaries, but also great people at Microsoft and Google, and they were telling us, and, you know, in podcasts or papers, that scaling laws were alive. And what I mean by that is that the idea of more, larger data sets, more parameters, and more, most importantly, more capital, more money, were having demonstrable, in fact, observable and predictable outcomes on model performance. And so that early period, you had to believe that, and you had to believe that, you know, GPT-4 was going to be meaningfully better than GPT-3. And I think as a rule of thumb, it's, you know, people expect that models will continue to evolve at the pace that they have been. But I mean, for 100% model improvement, it's sort of widely considered that you need to spend an order of magnitude more. So, so again, that's been, you've had to make a leap of faith in order to be in that.

Most important. Um, well, I mean, look, we launched an AI fund. It was called an Automation and AI Fund when we first launched it eight years ago with a view that at some point AI was going to be the next, the most important thing that happens is the internet. And that's, and then, and then of course the transformer lands, CHP. We didn't know about the transformer in 2017, but, um, but we obviously knew about CHP when it landed November '22, May '23. Nvidia has that phenomenal blowout.

You know, it's quite funny actually, as a quick segue, I think one of my first days at Polar, um, I mean, I think May 2003, um, Genentech, I think, biotech company, landed a vast, um, and the stock was up, I think, 60% on the open. And it was in my index, and, uh, I was helping Tim Woolley at the time run the US book, and we covered the, we had a zero, we covered the, we covered the position on the open, and the stock never went back to that level ever again. And just moments like that. So that was not dissimilar to how we approached that quarter from Nvidia, which was that we had been super excited about CHP ever since we'd seen it. Um, we had read about it, put in our strategy paper in January of that year, um, and we had started to sort of do more work, and then this quarter around, and that's when we went essentially all in on the AI trade, because it was at that moment that you realized that, and of course, people that were spending their money with Nvidia were the companies that we were most excited.

Well, one thing I should have said, if I may just, I talk a lot and then end up forgetting what I'm supposed to answer the question.

Um, but no, one thing I should have said previously was that they're the reason, the capex, you asked about capex, and the capex was about building models and then ultimately carrying the traffic or the workloads, the inference. Um, and but of course there's a third reason, which is the one that no one will share, which is of course that there must be defensive reasons too why companies embark on capex. And I think that again we were kind of alive to that a long time ago. Um, we meaningfully reduced our Mag 7 exposure, not to all hyperscalers equally, but we have, you know, these, the Mag 7, let's just use that group of companies as a proxy for what I'm about to say. Um, they are the dominant, they dominate the profit pools of the smartphone stroke cloud era, um, social media, right, social media, um, smartphones themselves, obviously, um, streaming content, you know, the Spotifys or Netflixes and so on, um, the software companies, the Microsofts and others. So these companies are dominant of pre-AI profit pools. Um, and many of them, um, are obviously building, or certainly the Microsofts and Metas and what have you, are spending hundreds of billions of dollars, um, for good reasons. But there is obviously for almost everyone, um, spending money today a defensive reflex too, which is of course that in a world where AI, you know, displaces or redistributes profit pools, then Microsoft has 450 million people using productivity software that may become moot in an AI world. Um, Google is 100% of its business comes from advertising. Um, and exactly. So again, I just wanted to make sure that I made that point, which is that companies, of course, don't share their fears. They certainly share their excitement, but they don't share their fears. Um, that's to the question on, um, on spending and revenues. Um, because phase two is all about revenues. Um, and the reason that, um, that, you know, you asked a long time ago now, like, why is it that, you know, tech has done well against the backdrop of Iran, Trump, I don't know, just 97, $100 oil. And I think the answer is really that, um, capex budgets have carried on moving higher. And I think anybody expected there to be upward revisions to '26 numbers in Q1. Um, we've just seen Google come to market with, I mean, a very surprisingly large and just surprising secondary offering. Um, you know, this is a company, I think, with 200 billion or thereabouts of gross cash on its balance sheet coming back for more.

I'm not going to pass comment on Supermicro. I'm not, [laughter] I'm not, I'm not a holder and have never been a holder, but, um, that one, that's sort of, uh, yeah, interesting. Um, but Google more so because obviously it's a company that has very deep pockets and incredible business. So why is it coming to market? And again, let's not go there yet. But just capex budgets have moved higher. But the real story of 2026 is Anthropic revenues, where the company was on, I think, a $9 billion ARR in January and I think is on 47 billion or something at the end of May. Uh, or is it April? But I mean, we've never seen a company do this before. In fact, when it reached 30 billion of annualized revenue, um, we had past that showed that no company on earth has ever done this before. So that gets to the question of what is it that companies are doing? Because the Anthropic takeoff of revenue, revenue takeoff is a function of enterprise adoption of AI. So why now? Why has it taken, you know, three and a half years for enterprises to adopt something that, you know, lots of people, a billion people or what have you, or 950 million people are using OpenAI, for example. Why has it taken three and a half years? And I think the answer there is models have got good enough. Um, you look at early, all the early pushbacks on models were around hallucinations and just not good enough. And in the end, you know, um, you know, good enough for a consumer is not good enough for an enterprise. Enterprises need repeatable workflows. Um, that is the sort of the backbone of a business. So now we have models that are on almost, on very, on many different, um, benchmarks at human plus levels and obviously improving all the time. So they don't make mistakes, or if they do, they make no more mistakes than humans. Um, and I think critically, um, in this sort of, you know, we had reasoning models about 18 months ago. So the ability for models to not just, um, try and answer a question based on what it's trained on, but being able to break down a question and then go out and find the answers and then recobble it. But then most recently, more probably the most important thing that's happened, uh, in the last 12 months is this sort of new class of models that are not only reasoning but are able to call tools. And this is something that's probably too technical and I probably won't do it justice, but this is the idea that the large language model is beginning to act much more like an orchestra, or sorry, excuse me, like a conductor or an orchestrator, um, than the orchestra itself. The models are now able to understand that you don't have to do the maths inside of what we would describe as token space. It, you can call a Python tool that could do the maths for you while you synthesize, let's say, the information on tax, your tax return. The math of the tax return can be done in a sort of side pocket in a Python tool, and then once you've done the math there, you bring it back into token space. And all of a sudden, what was essentially a probabilistic thing, because in the end, large language models predict the next token. Yes. Um, but if you can call tools that are deterministic, you can tend, a large language model can tend to a deterministic. And of course, what that means is that they can be, they could do, they could, you could ask the same question twice and essentially get the same answer. That's important. And then finally...

Perhaps. I mean, they, these, um, well, the next thing I was going to say was that as the models have also got bigger and stronger, it's no easy parallels to sort of illuminate this, but they are, um, there's been a breakthrough in coding, which of course has been the lead application, um, for Anthropic, um, in terms of revenue generation. But the breakthrough on coding is really a breakthrough of AI work that hasn't required humans in the loop. And if we went back to ChatGPT, um, so in 2023, um, software, software work without a human in the loop, the models could have done something like 35 seconds of human equivalent work. The models today are said to be able to do 14 hours of equivalent human work without humans in the loop.

Yeah. And I think that's really the story of 2026, which that's why revenues have gone vertical, and that's why the tech sector or certain parts of the tech sector have been sort of very strong against a much more difficult backdrop.

Yeah, it's critical because, um, because essentially it's the non-human scaling part of the AI story. Um, and I was actually at the weekend watching a western. My late father was a big fan of westerns. Uh, and I was watching The Big Country, and as they like to tell you in the movie, it's a big country. And when you live in a big country at that time, um, it's, you know, it's a horse, it's a horse and carriage. Um, and essentially information can't flow any faster than an animal. Um, and then of course railways change that, and telegraphs change it, so on later as well. So, um, the idea of, um, information traveling faster than a human or cargo traveling faster than a horse, um, is the idea of non-human scaling. And in the world of cognition, because of course in things like agriculture, we know what that means. We used to have people with scythes and hoes or what have you, and we had combines and all of that stuff in the 19th century, and we had the same with the industrial revolution and factories rather. Um, this is, this, we believe that agentic is sort of the, um, the sort of beginnings of non-human scaling. So by which I mean, um, before agentic AI, the way that we ultimately interfaced with it was via a chatbot. Yeah. We're asking questions in the way that we would have asked Google, and you probably know this, but, you know, before AI, the average number of words in a Google search was around five, and after AI, um, the number of words in a Google search expanded to about 12. I don't know what the latest numbers are, but the token windows for these AI tools can be a million tokens, which essentially means you could drop, I mean, the whole of War and Peace into, uh, into an AI, um, as a prompt and ask questions of the entirety of the book. But, but of course people didn't and didn't know that that's what you could do. And of course, by, in the early days, and of course who would, so for an enterprise, you're trying to, um, so, so we should have said already, what is it that AI does brilliantly? It synthesizes information in a way that humans cannot. Yeah. Um, our excitement around what that means is just things like think about breakthroughs. Think about von Neumann or Edison or look at human history. The polymaths have, you know, are a very large part of the breakthroughs that humans have enjoyed. Um, but of course humans don't, we're not polymaths. We're actually, here we are today, we're experts in a very sort of thin slither of knowledge, and AI is able to bridge that in a way that, I don't know, let's say NASA did to make sure that the US landed on the moon first. And I think we're super excited about this idea that essentially AI can synthesize information. You can take, I don't know, let's say 60 minutes of what would have been human time could be done in, I don't know, one-fifth of that or something, or maybe one-tenth of that. And so that ability to synthesize information means you can start going across modalities and all sorts of stuff, and that's where our excitement comes from. So, um, go back to the point, um, why is it non-human scaling? Because in the end, we were interfacing, we would ask a question, we would get an answer, but of course if you could task AI with a problem, that problem might be, you find me three interesting ideas in the Korean market that are, you know, tradable and might be an interesting way to get access to X, and it can actually deliver that. That has demonstrable value to me as an investment manager. So one of the things that we've been doing, many of the things we've been doing, but two years ago we added somebody to our team dedicated to drive AI through our investment process to what we call supercharging, and that process begins with trying to get all your data in a sort of state of readiness first, right? But then it's all about trying to sort of juggle and, you know, which are the best tools out there, and ultimately, um, sort of feeling your way because the industry is still very new. Um, we're now at a point where we can ask, not ask, we've now built essentially agents that are able to crawl for things that we find interesting. So let's, I don't know, cross-correlation, something that might have been, um, the preserve of someone in the stat arb business or a hedge fund that's trading on, you know, again, we're not traders, we're investors, but wouldn't it be interesting if we could access that part of the market? And now we have agents crawling, looking for interesting correlations that might be something that, or interesting companies that fit the profile of businesses that we like. That's AI substituting for human labor. So to get back to your question from a while ago, who's paying for this? You know, this idea that people are beginning to, you know, seen some high-profile comments, I think, from a few very important companies like Walmart and others saying that they're having to sort of maybe retrench a bit in terms of some of their spending. I think it's worth just recalling that we don't know the numbers exactly, but let's say we're $100 billion of AI revenues. Let's say we're tending towards a trillion dollars of AI capex. Um, it's worth recalling that the global labor bill is $44 trillion and the global opex bill is $60 trillion. And it is just worth pointing out, and again, please, I'm an investor with a sort of a human hat, and as a father of two kids that are going to be looking for work within a few years, all being well. Um, you know, this is, I mean, I'm not saying this lightly, but it's unusual to see companies like Block, PayPal, uh, Git, uh, GitHub recently, um, GitLab, excuse me, recently, um, very large-scale redundancies, um, and they're all calling out AI as the reason behind those redundancies. So I think you're going to have lots of two-way on this where some companies are telling you that they've spent more than they expected. They've used more tokens than they anticipated. I think one company recently said they'd spent their year's budget in the first five months of the year.

I think that what we would expect is for every one of those, I think that the data points of very large-scale redundancies from tech companies that are seeing productivity gains speaks to a sort of medium-term trajectory yet where there may be some cross-currents around usage. I think we're really just starting in terms of enterprise adoption. But I guess when I started my career, like the stuff I started out doing, and let's just for the sake of argument say we had a client meeting next day, or my boss did at the time, he was going to see a client, but we'd manually work out the performance figures based on values now and money in, flow and out. We did it pen and paper-ish, whereas now we have software that do that.

I think that we didn't see anybody lay...

people off because they had smartphones and we didn't see people have mass redundancies because they had just got access to the internet. I think, you know, again, has SaaS software reduced the number of people in an HR department? It absolutely has. But that's, you know, would have been at a departmental level. You know, are there, have there been meaningful productivity gains to software? Absolutely there have. And, you know, one of the... just let's not go in, I could get sidetracked into talking about US productivity and comparing it to various other countries including...

what, this doesn't feel evolutionary. This feels very different to that. This would feel much more like railroads or electricity. You know, again, and in those early days of electricity, I'm sure people were sort of saying, well, I don't see the gains, you know, steam did this. But of course, you can't obviously predict what's about to happen. I think one of the challenges with being, with staying the course on an investment story like AI is that the narrative is always confined to the sort of present or maybe present plus a bit. And I just was reading the history and again I'm giving you my scoop on this because I haven't even published it yet, my annual report, but you're a history graduate.

Well, that's what I did at university certainly, but I've really enjoyed over the years trying to draw on the past to try and illuminate the future. And nothing's exactly the same, but the idea that what's happening to us as a generation is unique beyond the detail seems highly unlikely. And so I, you know, we've lent on lots of different industry parallels to help us before. But one I thought was very interesting was that in 1901 Standard Oil was already the second largest company on earth and, you know, made billionaires. Oil already made people billionaires, but oil at the time was actually just for lighting. It was kerosene and all of the lighter stuff was just basically burned off. And the reason for that was that coal was basically a quarter of the price. So in 1901 there was a find in Texas or somewhere, I think it was called Spindletop, and at that one moment that they found more oil, I think, in that one place than the entirety of the oil industry in North America. The oil price falls to below the cost of water, then normalizes at the cost of coal. But what's really interesting about it is that it took, I think, within the year for ships and railroads to start basically moving from coal to oil. But of course that's not the story at all. The story really about oil, and again this is very simplistic and I haven't read enough on it yet, but I just put it out there, is that it was the change of state from solid to liquid that basically created this portability that didn't exist before. And it also, interestingly, why I think it's interesting, it's obvious, which is it took a lot of humans out the loop. When you're mining a solid, you're digging it out of a mine and then you're moving it physically and you're putting it onto a train and so on. And every part of the process involves humans. And of course, what you need to do, of course, if you're ever going to design the car or the plane or if you're going to create a petrochemical industry, is have something very different. And I think that those things are not viewable or knowable from the boiler room. And I think that that's, you know, again, it's another parallel. I think that also the parallel works because that reservoir in the place in Texas is a little bit like, I think, the well of cognition that we're opening. There's so many things that humans haven't been able to even scratch the surface of feel like they should be coming into our vision or whatever it is over the coming years because or be unlocked in the same way that oil was in 1901.

Look, I mean there are similarities because we're obviously building infrastructure from scratch and, you know, again, if you went back to that early period, it was really driven by the idea that there was a new way to carry network traffic or telecom traffic which was once circuit switched and then became packet switched. And I remember as a youngster, you know, my grandparents were sort of snowbirds in Florida and the parents would wake up in the middle of the night in order to find the cheapest way to phone them because it was super expensive to make international phone calls. And then the internet happened and then the smartphone happened just to, you know, kill off anything that the telecoms thought they were going to be able to capture in terms of that. So I mean there are definite similarities between that period because we're building infrastructure from scratch and today the infrastructure of data centers and the pipes that we need today are maybe fiber optic cables but actually same stock, same companies that survived from that period. So definite similarities, you know, the scale, the pace of the build, some of the hyperbole, you know, the number of tokens are doubling every hundred days is not quite right but it's not that far off. I think it's less than that, but they are the sort of things that we heard during the late 90s. So, you know, history doesn't repeat, it rhymes, that's all true. I think also honestly, I think those periods for me too were, there's a lot of scarring and I think people that were late to the dot-com period will be sort of, I suppose, will remember the bad bits, they won't necessarily remember the early years. I don't, you know, again, look, we're on bubble watch the whole time. And, you know, two-thirds, we think two-thirds of capex today is now going into supporting inference.

Inference is the stuff where there should be revenues against.