The Capital Cycle Podcast - When Two Worlds Collide

Episode Date: August 6, 2024

A discussion on the massive investment boom in artificial intelligence that is currently underway and its environmental consequences. Presented by Edward Chancellor.With Robert Anstey, Portfolio Manager North America.For more information, or to access select articles from Marathon’s Global Investment Review publications which accompany this podcast series, please visit www.thecapitalcycle.co.uk Hosted on Acast. See acast.com/privacy for more information.

Welcome to the Capital Cycle podcast from Marathon Asset Management. My name is Edward Chancellor. I'm a financial historian, journalist and former investment strategist. In these podcasts, I will talk to the authors of Marathon's Global Investment Review about their current investment thoughts. In today's episode, I'll be speaking to Robert Anstey about the consequences of the AI investment boom. I've got with me, Robert Anstey, who is portfolio manager marathon covering US small mid-cap stocks and Canada. And we are going to talk about the hottest subject in the investment world today, namely AI and the capital cycle.

There's a piece, Robert, that came out quite recently from a young fellow. who's just left Open AI called Leo Aschenbrenner, in which he talks very excitedly about the advances in AI being in orders of magnitude. OOMs is the acronyms we need to keep in our head. As we've seen this year, NVIDIA has been bouncing around, and the volatility of the NVIDIA stock, the chipmaker, that produces the GPUs that go into AI's, is five times that of the S&P, which reflects, I think, a certain degree of uncertainty as to the

outlook for AI in general. There is, and this is what we're going to talk about, huge investment surge going on in AI that looks quite similar to some of the great investment searches during earlier technological breakthroughs. So let's talk about your new piece. When two worlds collide, the two worlds being the ethereal world and the real world. And you start with a description of the operation of the human brain. So far away, get your brain synapses. Yeah, I started to think about the analogy of what it means to create artificial intelligence. But just thinking about the miracle that is the human brain,

86 billion neurons, each of them capable of forming thousands of connections, which leads to, you know, over 100 trillion synaptic connections. And all this is done using a simple energy source of glucose. That is then converted for various processes to create something called ATP, which really sort of creates the electrical impulses in the brain. As I started to think about this, I thought, well, actually, if you think about the creation of AI, you might say that semiconductors are the neurons, copper, which I believe you've talked about on the other podcast,

is providing those synaptic connections. And ATP for AI is electricity. And boy, does AI need a lot of electricity. So let's get into this. The AI capital cycle in full flow. Let's talk about, first of all, before we get on the electricity, just how much is being spent. You referred to, I hadn't actually come across this phrase before,

the hyperscalers, shows that I'm living under a stage. Yeah, the hyperscalers are basically four companies, Alphabet, Amazon, Meta and Microsoft. So they're a sort of subset of the magnificent seven, if you like. Cell side estimates have them spending around $200 billion in CAPX this year, and then possibly up to a trillion dollars in CAPX over the next four or five years. And I thought it was interesting that, you know, if you were to look at prior investment booms, I'm old enough to have lived through the tech telecom boom of the late 1990s. That $200 billion cap-ex this year is about the same in nominal dollars as what the telecom

company spent in the year 2000. Yes. And to recap here, in the Capital Count book, the first marathon book, there's quite a lot of discussion of earlier GIS on the investment that was going on in the so-called TMT. and then how that played out to Marathon's advantage by staying out at that at the time. So, yes, they're spending roughly today what the telecoms companies were laying out the fiber optic cable and so on for the internet revolution. And with the prospect, according to people who are AI-pilled, as it's called,

of these trillion-dollar data centers. There's a lot of hyperbole around AI. And to be honest, you know, I've just come back from an investment conference. And every single company was asked about what is your AI strategy. To be honest, as an investor, probably most investors feel like this at these conferences, it just gets a bit dull after a while. Yes. Everyone has to get onto the bandwack.

Yeah. So part of the marathon investment philosophy is that when you get a new technology, you don't know where the revenues are going to be coming from. Yeah, I mean, that's definitely clear with AI. So my thought about this is that this CAPEX from the hypers is defensive. You know, the one constant in technology is change. You know, they know that AI is an existential threat to their business models. You know, if you're Google, you know, basically we've relied on one search engine for the last 20 years, right?

Google. You know, no disrespect to Bing and Microsoft. Now, you can disrespect them because they haven't, they've only got, you know, 10% market share between them. If Google's had 90s, it's had a near monopoly for 20 years. Exactly. So then the question might become, taking this forward logically, how many large language models does the world need? One might argue one, if we get to artificial general intelligence, right, one is the answer. You know, so the latest iteration of chat GPT apparently uses 1.75 trillion parameters. That's a tenfold increase.

over the prior model iteration. It's basically an arms race between companies, and they don't know what the returns are going to be on that KPEX. They simply don't. They can give us lots of exciting anecdotes about what AI is going to do for us. They can say something like if one third of Microsoft customers paid $100 a month, it would justify it. But they don't know whether those customers are going to do so.

And we don't know how many of these models we ultimately need. So when I read that in your piece, it reminded me of the British Railway mania, at the 1840s, where not far from here, three railway lines were built between London and Peterborough when one railway line would have been profitable. Three were disastrously unprofitable. So actually, at the moment, the market is pricing in all these companies being successful, whereas you're arguing that they're actually defensive and duplicative investment.

Correct. That's the key point here, is that by definition, because their spend is defensive strategically. It's also duplicative. By definition, we're buying more GPU chips than ultimately we really need if we just need one model. Maybe we do need more than one model. I don't know. There'll be sort of models at the edge. You wouldn't want to have just one model because that would be the most powerful company in the world. Well, of course, that is exactly what, you know, Sam Altman and Demis Mzalbis, founder of deep mind. Of course, they're pushing towards artificial general intelligence, which is that kind of, you know, single model. But, you know, so it's an arms race not only between companies, but of course between, you know, countries as well.

We know it's an arms race because, of course, the US is denying China access to the chips. Another point you make in this piece, which I think it's very important from a capital cycle perspective, is you're saying that big tech has now become big oil. And we know, just like the miners had a big Cappex boom in the early 2000s running up to 2012, 2013. Big oil also went through a bad patch then, which again, if you remember, Marathon picked up in its GIRs at the time. Google had such a great moat, and they had such relatively low CAPEX. They could throw money away on projects like autonomous cars or whatever, and it didn't really hurt them. So show how their business models fundamentally changed.

Yeah, I think that's one of these sort of maybe underappreciated things that's happening because of this CAPEX gold rush. if you like. In the first place, these tech platforms were admired because they were capital light. Well, now, that's not the case. And so my reference to sort of big oil is a sort of tongue-in-cheek allusion to that. But if you look at the capital intensity for those four hypers, you know, so CAPX as a percent of revenues, on average, that equates to about 14 percent now versus four percent a decade ago. Okay, what about it? And you've got R&D on top of that. R&D on top of that. Because of this gold rush, they don't want to lose.

They're becoming more capital-intensive business. And we don't know, I don't think, they know what the returns are going to be on that, right? So let's get into then the real world, the energy consumption, the huge drain on electricity this threatens. And bear in mind, this is coming at a time when the world is also being required to shift its energy sources away from high. hydrocarbons with massive investment in alternative energy. And a shift to, you know, presumably if everyone goes from internal combustion engine to electric vehicles, then that would be higher demand on electricity. So tell me a bit about the electricity side of the story.

Yes, you know, so this is why it's two worlds colliding, because you've basically got the world of this capex boom towards AI. And then actually we've got the real world impacts of it on the environment. And I don't think this is something which. is hitherto received a lot of airtime. The hypers don't want to publicise this necessarily. But when you look at the electricity consumption of data centres, it's going up a lot. Elon Musk, sometimes prone to hyperbole, you know, basically said, we're going to run out of electricity in the next year or two at this rate. But my attention was drawn to this

through one of my former portfolio holdings, which is a company called Vistra. And Vistra is one of the largest electricity generators in the US, so it's based in Texas. You know, my original investment thesis on Vistra was that there was a very tight supply demand situation in Texas, amongst other places for electricity, and that as the largest owner of generation assets, it would eventually rise in value. Well, it did. Unfortunately, I sold the stock as a result. That happened.

You know, profitably, but it does happen. And since then, little could I have foretold that actually Vistra, an electricity generation business would become an AI play and it shares on a sort of one year basis that would actually outperform invidia because the hypers are scrambling to get electricity sources. And what shocked me and why I wrote this article is that Vistra recently announced that it's going to build two new gas plants and restart a coal plant that it had retired with the intention of converting it to gas by 27. So this highlights the environmental impact right now of this capex bin. So the real battle is going to be whether the AI world can be built at the same time as the

transition to so-called net zero. It seems to me from reading your piece that those two positions are incompatible. Not necessarily. So, you know, the AI protagonist will tell us that AI will help us to be more efficient and lower our carbon emissions. For example, through heating our homes more efficiently. Or there's an amazing study put out by Google, of course, but which shows that, you know, aircraft contrails, which are those sort of vapour trails in the sky, that if you could reduce those by a third using AI,

the emissions reduction would more than offset any electricity use. So they're going to tell us those things. And that's great. and that could be true. So I don't want this to be a sort of Luddite rant. It's not supposed to be. You actually mention it's not Ludditeism. You mentioned the so-called Jevons Paradox,

which is created by this Victorian economist, W.S. Jevons, where he argues in reference to coal and technology, the improvement in steam engine technology, while using individually less coal, actually increase the demand for coal. And you think something similar has happened. That's correct. In AI world.

I think that is possible, isn't it? Because even though each individual steam engine was more efficient, use less coal, it then led to much great to use of steam engine. There's something similar may happen with AI, where it leads us to be more efficient, but then we just use it a lot more. So I was struck by, if you look at the sort of individual stats on this, that a Google search uses 0.3 watts of energy, but a chat GPT search uses 10 times that.

But if you do an image-based consumption, it's apparently an image-based search could be the equivalent of charging up your smartphone in terms of power consumption. So quite a lot. So before we've sort of integrated this into the 9 billion searches that happen every day, there's going to be a lot of demand for electricity.

I heard that Microsoft last year had in two months use as much electricity as it had in the whole year before. So you can see how they... The hyperscaling is in electricity consumption. It's dramatic. And then there's the water as well. So, you know, we've got some 20-22 stats on that from the companies themselves. But, you know, Google used 5 billion gallons of fresh water in 2022, which was a 22% increase

on the prior year. Microsoft had a 34% increase in water usage on the prior year. And that was... This is two years ago before we've really probably got into a lot of this stuff. So I've seen some studies that say that, you know, a typical set of chat GPT searches use about half a litre of water. There's a town in Oregon where three of Google's data centers use a quarter of the city's water supply. You know, so it's going to put stress on all sorts of areas.

The real world is still out there. I mean, of course, you know, with electricity generation, if we can generate that electricity using solar and wind and, you know. So the International Energy Authority is claiming that at least for next couple of years, alternative sources of energy might supply enough energy to meet the demand growth. Yeah, that's right. And that's encouraging, right? But the example of Vistra shows that at the margins, they're desperate for power.

And even if you co-locate your data center next to a new, nuclear facility. What you're doing is you're pushing out another company in another sector who's going to have to get their electricity by, you know, dirty means. So in the AI-pilled paper, I cited the beginning that the Fed says that big tech will soon be buying aluminum smelters for their fixed electricity contract. This is the sort of world we're living. This is such a big story that I'm sure that this has been covered in earlier GIS and it will be covered in future GIS. It's absolutely fascinating, just a little glimpse into how Marathon is implementing these views. And I think these are broadly held views across Marathon, I take it. So tell me your thought. Yeah, so the investment

implications is firstly to recognize that this is a massive CAPEX boom. And we know through the history of capital and cycle investing, that are often, and that doesn't end well for investors. So we've got to have one eye on that, right? But, you know, there's different sort of layers of the onion that you can peel from a portfolio sense. I think we're benefiting from general electrification. So we say, for instance, Marathens's European holdings in like Schneider, Le Grande, some of the cable makers, electrical distribution companies, etc. But then I think the broader point I would make is that what happens when you get these booms is the entire conversation becomes consistent. assumed with them. It's like a black hole. It's a black hole of capital. It's a black hole of

investment attention. It's a black hole of everything we're talking about. And what that means is you want to be in everything else. And this is what Marathon did in the telecom boom when everyone was chasing. When value was extraordinarily cheap. It's still cheap today, not quite as cheap. We were in small cap basic materials and industrials. And actually, I would argue, that's where you want to be as small caps, emerging markets, everything else. Great. I hope we return to the subject another time. Thank you for your time today. I hope you will listen to the next edition of the capital cycle.

This communication is provided for information purposes only. Please refer to Marathon's website and the Global Investment Reviews for further information, including important disclosures.