AI: Mad Maths?

Welcome to another episode of the Capital Cycle podcast. This is Edward Chancelay here. And I have with me Charles Carter, a European portfolio manager at Marathon. Good to see you, Eddie. So Charles, we're returning in this podcast to the investment topic that has dominated the news for nearly three years, namely the huge excitement around artificial intelligence and the capital spending boom that's been unleashed, a spending boom that gets larger and larger as every month passes. From a capital cycle perspective, these developments are unsettling to say the least and reminiscent in many respects to the investment splurge, which accompanied the dot com mania of the late 1990s. Yes, I mean, I think, you know, it's obviously a lot of excitement in the stock market, which is quite narrowly confined to this area of air at the moment. And it's quite natural for capital cycle investors to start worrying because with frenzy sort of enthusiasm, you can get massive investment, or we are seeing massive investment, which often breeds unforeseen and excessive competition and in the end disappointment for investors. So how do you think investors should respond to the current frenzy? Well, I think for a generalist investor, there are two courses of action that are when confronted with a new investment boom like AI. One is to bone up as quickly as possible on the new technology and then try and talk plausibly about trillions of tokens and post training, scaling and reinforcement learning and so on. And the other way is really not to bluff once way into a technical discussion, but to think about the fundamentals. And from a capital cycle perspective, this means thinking about capital flows, industry structures and potential return on investment. And the other aspect of the question is to look at the motives of the participants and to get an understanding of why they are behaving in this way and indeed whether it makes sense. So let's discuss the scale of investment that's currently underway in the AI field. So Morgan Stanley have come up with some estimates and they have a figure for cumulative investment in data centers over the period 2025 to 2028 of $3 trillion, excluding the cost of energy. McKinsey, they have a figure of over $5 trillion up to 2030 and City have figures of 2.3 trillion for hardware and 1.4 trillion for R&D up to 2029. So these are very big numbers. Big numbers indeed. And Charles Zuhramber, 20 years ago, you and I put together the first marathon book, Capital Account, which contained real-time analysis of the.com boom and its aftermath. What struck me very forcibly about some of the global investment reviews of the late 1990s and early 2000s that we included in the book was it was quite clear, as I say in real-time, that the revenue expectations embedded in some of the TMT stocks were wildly unrealistic. And there's one piece I remember which described the value of telephonic as internet business that at the time was called Terranetworks. And one of the marathon authors pointed out that given the then current market cap of Terranetworks, it would require the company to gain an unfathomably large share of Latin American total consumer spending. And that didn't happen. And eventually Terranetworks, share price having climbed nearly 10 times from its 99 IPO, the stock declined 98% from its peak. Now, do you think that the return on capital currently being invested in AI is unlikely to generate the markets implied expected return? When it comes to the return on investment, Mr. Market seems to believe the money is well spent, given you see, I think, a $12 trillion increase in the value of 10 technology firms in the US since the release of chat GPT in 2022. That's according to the economist. And so either the stock market is rewarding the new investment with what looks like a Tobin's Q ratio of between two and four times, that's the incremental market value divided by the investment outlay. Or it's taking a much rosa view of the core legacy technology franchises in these businesses. And it's probably a combination of both of those. With regards to Tobin's Q, the replacement cost of a business, one of the points we make in both of the marathon books is that when investments are trading at a high Tobin's Q, in other words, at a premium to their replacement cost, there is a huge incentive to carry on investing. Yes, I mean, green for go. So tell me, what do you think will actually drive the returns on AI capital spending in your view? Well, I think in the short term, short to medium term, the actual return will depend just as it always does on the life spans of the assets and the sustainable profit margins. And that's a function of the demand for AI solutions and services and the degree of competition. So McKinsey estimate that 60% of data center investment goes into chips and hardware. Currently, these assets are being depreciated by the hyperscalers over a period of roughly five to six years. But you think that this depreciation schedule is optimistic, given new technological innovation and new chips that come along? Yeah, and actually, the time period has been extended despite what appears to be an accelerating cadence of new launches. So Nvidia is releasing a new chip platform, if you like, every year now. That said, Amazon has recently reduced its depreciation period saying that it is due to the increased pace of technological development. So if we're trying to calculate its better returns, let's do the math, as the American say. Yeah, so I mean, just very simply, if you assume the assets generate no economic profit after five and a half years and that the 3 trillion figure for Morgan Stanley is back-end loaded, then the chips and hardware investment alone would need to generate a net cash flow of over $500 billion in 2028 just to meet the cost of capital on the new equipment investment. Is it realistic to expect sufficient demand to generate such profits? Well, yeah, if you assume the data center operators need to generate 20% free cash flow margin to justify their share prices, that implies a revenue requirement of $2.5 trillion. And if their customers in turn need to make similar margins, that implies consumers and businesses would need to pay close to over $3 trillion for AI services in the not-too-distant future. That's equivalent to 10% of current US GDP and 5% of the global labor cost based on OECD data. And again, it's just to cover the cost of the equipment and doesn't include other investments like energy infrastructure and R&D. City is 1.4 trillion number that I mentioned earlier and all those $100 million sign-on bonuses or sign-on salaries, which are nice if you can get them. So how does that compare to current AI revenues? Well, current AI revenues are modest compared to these figures. Open AI, which has the lion's share of weekly daily active users for its chat, but has a current revenue run rate of about $13 billion. And the information forecasts that grow to $200 billion by 2029. City have an estimate of total application and your revenues of just over $40 billion in 2025. And they have that rising at 80% a year to reach $780 billion by 2030. So it's a falling short of the target that you mentioned earlier. Yeah. And where are those revenues expected to come from? Well, I think in the case of the city numbers, it's largely predicated on rising demand from enterprises through application programming interfaces, APIs, giving access to the large language models, enterprise customization, partnerships, that kind of thing. And then for UBS, on the other hand, I mean, they're much more skeptical on the enterprise adoption, noting in their case that total third party AI product revenue for the listed software companies is only around 2.5 billion at the moment of which Microsoft is more than 80%. Probably a problem of market definition. Yum. And one of the problems with AI today, which is what we also had in the TMT boom, is it was unclear what the eventual and profitable applications from the new technology would be. You think that the uplift to profits from the early AI adopters has been pre-underwhelming to date to say the least? Well, yeah. I mean, there have been a couple of reports that recently, one from MIT, where they looked at-- [BLANK_AUDIO] 300 publicly disclosed AI initiatives and found that 95% of the projects failed to deliver any boost to profits. Another report from McKinsey found that 71% of companies reported using generative AI and more than 80% of them reported that the technology had made no tangible impact on earnings. So even with cities' bullish forecasts, huge losses in the AI ecosystem seem highly likely over the next five years. As you say, this huge gap between even their figure and the three trillion revenue figure that you would need to justify the investment. And of course, you can say, well, this is just early days and there may be J.K.E.R.'s effects and no doubt when properly applied, AI is going to bring a huge improvement for productivity. And I heard a case recently of a pharmaceutical company that was using AI to reduce the time it takes to format and analyze trial data for new pharmaceutical products. Previously, they had a very manual process which took three to four months and with AI, they could reduce that time by 80%. So you could argue that these hyperscalers are investing for the long run, which marathon should appreciate? Yeah, I think the counter-argument is that the investment logic is that you know, you don't need to generate profits in the short term, they can continue to incur huge losses as investors look upon the current way of investment as a stepping stone to a future with agentech, AI tools for video and yet unknown kind of applications of superintelligence. I think it can also be seen as a subsidy to consumers and businesses to start using AI in the hope of charging them later. You know, it's a bit like what happened a few years ago when the venture capital industry was subsidising, take away delivery drivers. And I remember Charles Stierntu in the dot-com boom, there was one company that was launched, that was paying people, giving them computers and expecting to get its revenue from a bit of advertising when they surfed the web. There's that same build and they will come, subsidise now and profits may turn up sometime in the future. So let's talk about the subsidy, what you think about the subsidy. You know, you can look at it almost, I mean people talk about training, it's like a very expensive college education, you've got your undergraduate, your postgraduate and then your PhD and after years of investing further years in the future you'll get your payoff. And in that technology race with growing demand and a lot of uncertainty, investors don't mind about not having short term returns if there is this idea of a long term payoff. And I think crucially that patience is supported by the idea that you might have winner-take-all economics, which has been a feature of the tech industry in the US, certainly over the last 25 years. Yes, over the last 25 years, but not actually for many of the companies that participated in the TMT boom of the late 90s, early 2000s, the companies with the exception of Amazon, who stock fell 90%, some of them like Google and Facebook appeared long after the dot-com bubble. Yes, I mean, you're going to have very few winners and a lot of losers. Yeah, so the eventual outcome, as you suggest, would depend on the future structure of the AI industry rather than demand in itself. Yes, from a capital cycle perspective, the demand question is really secondary. What matters is whether the industry tends to fragmentation or concentration over the next 25 years. And even if demand exceeds today's expectations, investor hopes can still be dashed by the onset of excessive competition. What have we seen to date? Well, there's an argument that economies of scale will drive concentration in the infrastructure and the platform segments, though, think chips, hardware, the large language models. And Nvidia clearly dominates its market at the moment with over 70% market share. And there have been really a few hyperscalers who've dominated the data center industry so far. But there are signs that competition is rising in data centers with players like Oracle and CoreWeave, et cetera, growing in the market. And competition between hyperscalers could easily escalate in a more capital-intensive field of battle, as it were, particularly if it turns out that the spare capacity, residing from either overbuild or productivity driven technological progress. While there may be capital barriers to entry, it's hard to discern major network effects, which have been such an important feature of the modes for these businesses in the past and to date. And data centers, you know, they are the high fixed cost and low marginal cost businesses. So when competition breaks out, prices tend to marginal cost a bit like airlines or all telecom industry after the dotcom bust. So, you know, it's the dumb pipe problem. And for large language models, the industry does appear to be fragmenting, not consolidating. So you've had new entrance like Grock and Deepseek. You know, companies, enterprises may find that they can get by just using small language models that they can adapt to their needs at a much lower cost. And on the application and the geographical level, the picture is really much less clear, but it's likely that there'll be many start-ups that will get funding and competition will be quite intense. Yes, now I saw a chart quite recently showing that I think that 80 or 90 percent of VC funding now is going into the AI space, which is not surprising. But if the outcome from all this incredible investment remains uncertain, how do we explain that it's going on at quite such a scale? I think partly the answer to that is behavioral. I'd like to quote a comment that was made by the Alphabet CEO on a earnings call back in July 2024 when he said, and I quote, "The risk of under-investing is dramatically greater than the risk of over-investing for us here." Even in scenarios where it turns out that we are over-investing, these are infrastructure which are widely useful for us. They have long useful lives and we can apply it across and we can work through that. But I think not investing to be at the front here, I think definitely has much more significant downsides. That quotation really hints to the innovator's dilemma facing the tech leaders. Nobody wants to be left behind in the next wave of tech platform change because of the fear of missing out. And the examples of Microsoft and also Intel in mobile. And before that, obviously IBM in PCs, I think are really scorched on the minds of the industry's executives. In other words, big tech faces the same problems that characterize as many management during adverse capital cycles as the analysts at TD Cohen put it last October. And I quote, "We believe the major companies in AI face an investment environment characterized by a prisoner's dilemma. Each is individually incentivized to continue spending as the cost of not doing so, are potentially devastating." And I think there's also this fear that broad adoption of agentic AI poses a threat to the company's respective strongholds in fields like search, productivity, shopping and social media. Bill Gates explained this dilemma back in February 2023 on a podcast. And again, I'd like to just quote, "He said, in a sense, the personal agent will replace going directly to Amazon or directly to Siri or going to Outlook, so that the fact that Google owns search, Amazon owns shopping, Microsoft owns productivity, Apple owns sort of everything on an Apple device. Once you get a personal agent, it kind of collapses those separate markets into, "Hey, I only want one personal agent." And of course, it can help me shop and plan and write documents and work across my devices in this rich way. And so, a decade from now, we won't think of those businesses as quite a separate. It's a pretty dramatic potential reshuffling of how tech markets look. So it's almost like a risk of this becoming like a great tragedy when the family members suddenly start killing each other. It reminds me of a comment, possibly, a powerful that circulated on the social media last year in which the alphabet found a Larry Page is reported as having said intently that he'd prefer for the company to go bust rather than lose the race for artificial general intelligence. And then, only today, someone sent me a comment that Mark Zuckerberg at Metta had made where he said, "After about the AI investment, if we misinvest a couple of hundred billion dollars so be it." So this sort of cavalier approach to all incredible capital spending, where we'll see. So a parting thought, Charles? Well, I think the AI investment boom in this sense can be viewed as a combination of this fear of missing out, the FOMO, and fear of cannibalization. Let's not generate an acronym for that. Well, it's not on air. So thank you, Charles, and we haven't even discussed the prisoner's dilemma facing institutional investors with regards to whether to invest in mega-cap stocks at current valuations. perhaps the subject for us. a future global investment review. Well, yes, at least it's a dilemma for active investors for passive your just a prisoner. Thank you, Charles. 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. [Music]