← Excess Returns 
Excess Returns
The $5 Trillion Question | Kai Wu on the Risks of the Mag Seven's Big AI CapEx Bet
01:07:18
Loading summary
Transcript93 lines
- [00:00]
A
These companies need to be making trillions of dollars in revenue, you know, five years plus out in order to justify these investments. What you find is a pretty shocking underperformance of these companies that are aggressively trying to grow their businesses as opposed to those that are not. Alphabet has 21% Microsoft 28 and then Meta 35 which is higher than the average utility at 28 today and higher by the way than the, you know, capex spending of AT&T AT the height of the dot com bubble in 2000 which is only 21% 4 most dangerous world in the English language or this time is different. What you find is that over the next 20 years after the crash their sales actually 11x so they actually did realize the promise of the Internet right? They, they grew very quickly. The problem is that at the same time their multiples collapsed.
- [00:54]
B
Hey Kai, welcome back to Excess Returns.
- [00:57]
A
Great to be back.
- [00:59]
B
We always love having you on and this topic in particular is going to be very interesting. So you recently and for people that follow Kai or even if you don't follow Kai, Kai has for a number of years been putting out excellent, awesome, interesting and I think very useful and important pieces of research on Sparkline Capital which is the website that his firm is, that's his firm site and so on there. There's a whole series of articles and just recently you put out an article. The title of it is Surviving the AI Capex Boom. And we're going to talk through all the different I think things in this article and why it's important for the markets, the Mag 7 and just you know, everything that you wrote about the historical context with big boom and bust and infrastructure and stuff like that. So this is going to be a really I think interesting topic and it's very topical because this is, there's tens of billions if not hundreds of billions being spent on this, you know, development in this build out with, with AI. But before we get into all that obviously Sparkline Capital is the website the research is on there. You can also jump off to the ETF set Kai's firm manages and just learn more about Kai and, and his background and the research that he's doing. So please go there and, and check it out. Okay, so I wanted to just, we there's going to be a bunch of charts in this episode so if you are just listening you're going to get much more out of this if you go to the, the YouTube channel, our YouTube channel but maybe to set it up. Kai, I was just what made you there was obviously Something that you, I mean it's all in the news and everything. But what was the genesis of kind of looking at this? Was there something that just like struck you or were you just like interested in like how this build out compares to maybe what has happened historically? Like give us the context here for that.
- [02:59]
A
Yeah, I think it's like the intersection of two threads. So as you mentioned, a lot of my works are on intangible assets and what I see is this transformation of the economy from industrial to information based and the rise of the intangible economy. And so this is an interesting kind of booking of the trend where you have the largest companies in the world which are these kind of asset light money machines like the Googles and the Amazons that just print money, start to turn away a little bit from the model that's brought to so much success. And the question is why is that and is that concerning and if so, why? And the second thread is funny because I actually wrote my, my economics thesis on credit cycles and my advisor was this guy named Ed Chancellor who was also my manager at GMO for a period of time who's very famous for his work on bubbles. He wrote the book Devil Take the Hindmost, but also on capital cycles. He's basically written the book on that. I've been thinking a lot about capital cycles and the role of the supply and demand of investment in various sectors through time and using that lens in this paper of course to try to explain and make sense of everything that's going on in the AI sector.
- [04:11]
B
That's great. So help us put this in context. Like when we're looking at, in this first chart, you know, we're looking at the spend of, you know, Meta, Microsoft and Google and Amazon, like how much are we talking about in terms of, you know, what these guys are putting up and expected to put up in terms of the money in this infrastructure development.
- [04:33]
A
Yeah, I think what's so interesting about this chart is the acceleration we've witnessed over the past year, maybe a year and a half, six quarters let's call it. So obviously over time these companies get bigger and bigger, so naturally you'd expect their CapEx to increase. What you're seeing is huge acceleration of CapEx and it's now running at what's expected to be $400 billion over the past over one year, which is a lot. And what's crazy is it's not meant to abate. It's not like a one time thing. Oh yeah, let's just build the data centers and then kind of move on. Most analysts, whether it's McKinsey or Morgan Stanley City, expect this to continue, if not increase further over the next five years and to reach levels cumulatively of somewhere between 2 and 5 trillion dollars. And keep in mind that there's a lot of downstream consequences here too. It's not just the spending of these firms. It's also the utilities and the power and the data centers. There's a lot of work that needs to be done to kind of build out the infrastructure to make AI to kind of realize the promise of these exponential scaling laws.
- [05:41]
B
One of the points you made in this initial section was the type of, or the amount of money that would need to be generated to I guess justify this type of spend. So you have this Bain estimate, I think that something like, you know, $2 trillion by 2030 in order to justify it. And that's the question that I've been kind of wrestling with. And I think a lot of people are, you have all this money being spent, but like what type of return is going to be needed on that to justify the spend?
- [06:13]
A
Right. Look, this is not the Apollo program. This is not like a, this is not socialized money. This is money that's being spent out of private coffers, you know, and ultimately on the behalf of shareholders in these companies. Right. And so as a investor in a company like Meta, you expect a return on your investment. And that's the thing, which is if you're spending $5 trillion, I mean there's some assumptions around depreciation, useful life of these GPUs which we can get into later. It's kind of a hotly debated subject. You know, you come up with an estimate that's in the trillions. Yeah, these companies need to be making trillions of dollars in revenue, you know, five years plus out in order to justify these investments. Now the problem is that, you know, their estimates range, but like you know, 20 billion, let's say maybe 50 billion, is kind of where we are now. So you need like 100x increase in revenue to go from where we are now to where we need to go in five years to justify the build out that's being planned now by these companies. And that's not entirely impossible, but it's a pretty big gap. And one that as an investor running this company is, you do have to ask that question, which is how likely is it that we actually do achieve that?
- [07:24]
B
How much of the stock market's gains have been driven by this small group of AI driven or I guess AI oriented companies.
- [07:37]
A
Yeah, so I cite this statistic by Michael Semblis at JP Morgan and he says since ChatGPT's release, which is like November 2022, I think 75% of the returns of the S&P 500 have been driven by AI linked stocks. So that's these magnificent seven stocks like Google and Microsoft and Nvidia, but then also a handful of smaller names that matter less obviously based on their weight, 80%, I think of earnings growth and what was it, 90% of capex. So these companies are kind of single handedly moving the markets as they have in many ways for the past decade. But ever since AI has become a theme, this very real investment is moving the stock market and the economy too to some extent. I think some other researchers have come up with estimates that maybe around 1/2 of the GDP growth of the US over the first 2/4 of this year can be attributed to this AI spending and then the downstream ramifications of it.
- [08:34]
B
Do you personally, when you see this type of like concentration or you know, a small handful of stocks, does that personally like concern you when you look at the market?
- [08:45]
A
Yeah, I mean that, that's an interesting question. Like obviously we have concentration now of, you know, the top seven stocks, the MAG7, at 33% of the index in the dot com boom. For context, at the height of the dot com boom in 2000 it was around 20%. So you know, 33 versus 20, so very concentrated. I think that's undeniable. Now, like there's a couple of things to think about. So on one hand you can argue that these companies are, you know, conglomerates that you could spin out YouTube, you could spin out AirPods and they could be standalone companies. And so we shouldn't necessarily, you know, think too much about the size of any one company. It's because it's just a conglomerate of many small companies. On the other hand, I think there is some risk. There's obviously the idiosyncratic risk associated with one single executive or one CFO doing the accounting for a single company. I think there's also the risk that these companies tend to be very correlated, that they're all kind of like doing the same thing together in a kind of competitive arena, which we can describe more in the future in this talk later. But these companies are all kind of tech companies that are operating a handful of oligopolistic services and they're also all in this case investing in a single market, which is AI. Right. They're all competing over this pie which is obviously tricky to the extent that that ends up going south. I also think that there's a valuation argument too. People have. I was actually emailing with Chancellor about this earlier yesterday about the fact that maybe concentration isn't bad in and of itself. What happens is when your index is one third a single category of companies, all of which trade at relatively high multiples, that embed this expectation of an AI future, that is what is potentially an area of concern. So it's more the valuations than it is necessarily the concentration inherently. But these things tend to go hand in hand because indices being cap weighted naturally will increase their weight towards stocks as their valuations increase. And so you do tend to see concentration be high at the peak of a evaluation cycle.
- [10:44]
B
And we'll maybe get into the. Just one comment on that. Like when all those companies are also doing business with each other and you have all these circular deals, it probably adds an additional element of risk to the mix, you know.
- [10:55]
A
Right, yeah. They're all like in bed together.
- [10:57]
B
Yeah. And we'll come back to that point in a little bit. But one of the things that I, Jack and I both really value this about your work is that you, you are trying to take like historical context and looking at like history and, you know, what's happened in the past to try to draw some parallels to like where we are today. And in this next SE section, you sort of talked about other capital sort of booms that our country has seen. And this next chart, you kind of are putting it in the context of the spend for railroads, Internet and AI. And so I'll let you sort of talk to this.
- [11:38]
A
Yeah, look like today's not the first time that we've seen the private sector build out the infrastructure for a brand new technology. You know, 25 years ago, 30 years ago, it was the Internet, which of course changed the world. And then before that, a hundred years beforehand was the railroad boom in the us. And so what you can do is you can say let's put this current boom into the context of the prior ones. What you find is that relative to GDP, that $400 billion is about 1.3% or something. So that's a number to start with. And then if you go back to the height of the dot com boom in 2000, it was about 1%. Railroads are much higher actually. So it was closer to 6% of GDP was being spent on building out the railroads in 1872. Now the difference though is that the depreciation is very high, different between railroads and GPUs so people estimate that GPUs depreciate at around 5 year useful life. So 1/5 per year is what's assumed in most of the filings for these hyperscalers. Other researchers would argue that it should be closer to two to three years because of the accelerating replacement cycle of Nvidia GPUs. But we can, for the sake of being conservative, assume, take Amazon at their word, five years. Fine. Okay. The thing is that railroads last 30 years. You lay the steel and even if there's a bust and no trains are driving over it, it's still going to be there. And if you make those adjustments and it turns out that on an annualized basis, we're actually spending more on the AI buildout as a percentage of our GDP than the fiber room in 2000 or railroads 150 years ago. So yeah, we're right up there in this, you know, at, at the record levels of spending for kind of the recent history of the United States.
- [13:28]
C
That GPU depreciation thing is so interesting because I don't even know that like the people that are using them necessarily know how fast they're going to depreciate because a lot of that is about, right. How the level of innovation going forward and how fast these things are going to become obsolete.
- [13:40]
A
Yeah. And can they be repurposed for inference or other tasks that may still have value? Right. Because if you were to say, oh well, let's just depreciate after one or two years, which is kind of the replacement cycle for these things, sure, maybe you can't train state of the art frontier models with them, but they're not useless. You wouldn't want to write them down to zero. So it's a little bit tricky. I mean, it's interesting too because actually what's happened is the opposite, where over the past several years, if you go to the filings for Amazon and these companies, they've actually been increasing the assumed useful like these chips over time from like two to three years to five to six years. Now is that the right thing to have done? Obviously it helps them from an accounting standpoint to do that because they can depreciate these assets, the capital spending over a longer period of time and therefore impact their earnings less. But is that actually how the technology is working? And I think this is a really heavily debated topic. A lot of smart people have weighed in on both sides of the debate.
- [14:31]
C
So this next chart is really interesting, the capital cycle schematic, because I think that kind of gets at the flow of this whole thing and how it might play out. So can you explain what we're seeing here?
- [14:40]
A
Yeah. So this is actually from Ed Chancellor's book, Capital Returns. And what it describes is this dynamic where you have this exciting new technology with railroads. It's the Internet, it's AI. And companies rush to make investments to kind of win that land grab and win the AI race, so to speak. And then what happens is you find investors reward companies for these visionary investments. Right. Remember when Oracle announced the deal with OpenAI, their stock went up by 36% in one day. So the investors are still very excited. Markets are excited about companies who are putting. The bigger the deal, the better. The more money you say you're going to invest in AI, the more your stock price will go up. Maybe it's backwards, but that's the way the market is today. So what happens naturally is that every company wants to be an AI company, just like every company wanted to be a dot com in the 90s. We had hundreds of companies launch railroads, railroad companies spread up in the 1800s. And so you see more and more investment into the space. Ultimately what happens though is you have just too much investment. Demand maybe increases, but it just cannot increase enough to meet supply. And you end up with excess capacity. This glut of fiber, glut of rail, that leads to a collapse in prices. We saw, I think 85% of fiber was unused after the dot com bust, which led the price of, you know, bandwidth to fall 90% over the subsequent years. Right. And that's obviously bad for the builders of this infrastructure. If you own an AI data center, you own a railroad, you own a telecom, not great. When prices go down 90%, you find many of these companies will enter financial stress. If to the extent these companies took on leverage, as some in the dot com boom did and most in the railroads did, you know, you would face the legitimate risk of bankruptcy. And many of these companies did go bankrupt after, you know, the capital cycle turned. Yeah.
- [16:40]
C
One of the interesting things, and I think this will be a thread we'll kind of pull out throughout this, is this idea is, is there something that's different about AI than these previous things? And I don't know exactly what it is, but I mean, we're kind of building towards a super intelligence or something like that. And does that change some of the stuff we can look at in history? And do we have to judge AI differently because of that? Do you have any thoughts on that?
- [16:59]
A
I mean, look like there is a state of the world that whereby the progress on the research side of these AI models surprises to the upside. Quicker and more powerful than we could ever have conceived. Half the workforce is laid off instantly and then the rest of us follow where maybe it's just the last invention and everything changes. But that I don't think would be my base case, at least not over the next few years. And of course these assets depreciate pretty quickly. This is a lot of money being spent in a Space and the 4 Most Dangerous World in the English language. This time is different.
- [17:40]
C
That's very true. So what I love about your work is you bring together my world of factor investing with this world of technology. And that's what we're getting out here at Exhibit 6. Because there's been a lot of research in terms of high asset growth firms and how they perform. So can you talk to this?
- [17:53]
A
Yeah. So I took this chart straight out of the Fama French work. So you know, obviously Ken French and Eugene Fama won a Nobel prize for their work on asset pricing. They had this five factor model, right, where it's market size, value, quality and the profitability. And the last factor is what they call the, let's call it the asset growth factor. And what that is is saying, let's look at companies that had experienced high asset growth over the past trailing year relative to those with low asset growth, and then look at how they perform over the next year. They'd rebalance and so on and so forth. So this chart shows that return stream over the past 60 years. And what you find is a pretty shocking underperformance of these companies that are aggressively trying to grow their businesses as opposed to those that are not. I think it's like 8.6 or something percent percentage points of underperformance per year, which you compound over 60 years gets you to a relative return of negative 99.6%, which is not good. So these companies have historically underperformed. And I think this is to some extent evidence of the capital cycle.
- [19:02]
C
And this next chart we're seeing the same thing, right? With high capex stocks.
- [19:06]
A
Yeah. So obviously there are different reasons why companies can grow their balance sheet and you can spend that money on a variety of things. Wanted to focus specifically on the most relevant variable, which is physical infrastructure. Physical capex. And so looking at that one variable itself, saying companies that have high changes in their capital expenditures versus those with low, how have they done? And you find a very similar effect. One thing to note is that you do see if you look closely, kind of acceleration of the underperformance into the dot com bust. And that's as we know, due to the fact that a lot of telecoms and other companies that were aggressively building out the Internet suffered undue losses, disproportionately large losses when things turned. But what's really interesting is that even outside of that one period, you find pretty consistent underperformance for these companies throughout the time period.
- [19:54]
C
And you see it across sectors as well. Right. What did you see in the sector data that you thought was interesting?
- [19:59]
A
Yeah, I thought this was important. So typically when you think about the capital cycle, you think of a sector versus sector thing. So. So shale boom in 2010, telecom boom in the late 90s, early 2000s. But it turns out that there's a lot of within sector dynamics too. So take a given sector, say materials. If a single company in its sector is just for some idiosyncratic reasoning, decides it wants to take over the world and build out its business and it kind of aggressively goes for it, that company will tend to have on average underperform its peers that are more conservative. So you're seeing both a cross sector but also within sector effect. And what's quite interesting is that this operates. We looked at the data across all 10 sectors and found that in every single case this effect exists, that there's underperformance by companies that are seeking to aggressively grow their capital base. And by the way that communications, we know that's kind of a crazy situation, maybe that's a fluke, but that's not even the worst one. Materials are actually even more impactful than within communications. And you even see some, you know, both asset light and asset heavy industries in this mix too.
- [21:04]
C
Yeah, and this next chart is interesting to me as well because this is sort of showing what we're seeing now where capex growth has been recently. And then it shows basically what we would expect. Right. Which is not what you've seen historically, but what you have seen now, which is technology is where all the CapEx growth is.
- [21:17]
A
Yeah, I mean that's the interesting thing because historically technology has been a relatively capital light industry, but we started seeing like a reverse over the past year and a half with this big surgeon in building out the AI data centers where, you know, you have 35% year on year growth in it, 33% and communications 25 in discretionary, which is largely Amazon. So the Mag 7 and these big tech firms include Oracle, you know, are really driving a lot of this just from a kind of total dollar standpoint but then you also see the downstream effects in real estate, energy and utilities. These are firms that are kind of building out the power and physical data centers required to supplement the chips and the models and all the other good AI stuff.
- [22:03]
C
I don't know if you've looked at this, but is this a lot faster than the dot com boom? I mean I was just coming out of college in the dot com boom, but it does seem like the pace at which the spending is increased is even faster than it was back then.
- [22:15]
A
Yeah, I think it depends exactly on how you look at it. I mean there are different components of this, but it's definitely. Yeah, definitely, you know, a big outlier event historically.
- [22:26]
C
And one of the interesting things to me too, and you can correct me if I'm wrong about this, but it does seem like there's more like the people that are spending the money, there's more of a winner take all feeling now than there was in the dot com boom. Like I don't think any of them thought like they were going to win the Internet, but it seems like some of these firms as they think they're working towards like AGI, they think there might be like a winner take all dynamic here and maybe that's encouraging more spending. Do you have any thoughts on that?
- [22:48]
A
I think it absolutely is. I think you know, these big tech CEOs have seen, you know, how the mobile social networks have worked out, you know, Uber versus Lyft. Right. They see the dynamics of digital competition over the past two decades and it's led them to this understanding that network effects and scale matter tremendously. And that's why, Right. A VC company would subsidize Uber at a loss for years and years and years because it will eventually become profitable and you know, be one of the, I guess to an oligopoly. Right. So I think firms are taking that same approach now towards AI even though it's a much more physical infrastructure based industry than the kind of the pure software businesses that we've seen the past couple decades. Yeah.
- [23:32]
C
As we move to this list of the firms, the highest capex increase, this would be shocking to anybody I think if we looked five years ago because these are all the firms you would expect not to have of the highest capex increase.
- [23:44]
A
Yeah, I mean, so you see Oracle here as the number one in the chart with a 249% increase in capex over the past year. Oracle is an interesting example because they were a little late to the AI race, but they've kind of come in and they've carved out like a really nice market for themselves being this kind of more neutral cloud AI provider in the space. And they've just been really aggressive making deals like the OpenAI deal I mentioned and others just, you know, they want to be, they're aggressively going for, for, you know, to make to become a player in the space that's historically been dominated by, you know, Amazon and Google and Microsoft, you know, and Meta and Amazon are of course on this list as well as is Nvidia. You also see a few guys who are kind of more in the, who provide other infrastructure for data centers. So not the chips themselves, but you have like Arista, Amphenol, Micron. You also have lam research on the semisconductor side. And then you see Palantir companies that are more like users of AI, but that are aggressively using it to presumably win government contracts.
- [24:47]
C
In this case, you mentioned that Oracle deal before. I believe the money they've pledged to spend is significantly more than they even have at this point. Right. I think they're going to have to issue debt or something as I recall, just to spend the money. So to your point before, it's like a race to say who's going to spend the most money even if you don't have the money right now.
- [25:03]
A
Right. And the wild thing is that the market rewards that. Like the market's like, tell me how much you're going to spend. The more you spend, the more I'm going to make your stock go up. Which is kind of interesting because that's not what I expected to happen. Yeah.
- [25:18]
C
For those of us who lived through the dot com boom, it was kind of a similar thing, not just with what you're spending, but if you put dot com in your name. And so there's a tendency to want to relate it back to that. But also I try to keep myself checked like on the other side of this and say like, I don't know, I mean, maybe something is different about AI. I mean these are very, very smart firms that are spending a lot of money. It'll be interesting. None of us know. We'll see how it plays out.
- [25:40]
A
Exactly.
- [25:41]
C
One of the things we've talked about a lot is, you know, the, the Mag 7 have sort of been referred to as this asset light money machine. And one of the things we're seeing shifting in front of our eyes is this idea that they're asset light companies because they are spending their cash now like they haven't ever in the past. So can you talk about what this chart showing us?
- [25:58]
A
Yeah, I think it's helpful just to contextualize how dominant the Mag 7 have been. So over the past decade, they've compounded at 27.5% per year. And I did the math for this paper, that means they created over $23 trillion in wealth for their shareholders, which is just an ungodly amount of money. And as the JP Morgan statistic showed, three quarters of the return to the market of the entire S and P can be true to these seven stocks know, over the past couple years. So these guys are kind of like the all important pillars of the stock market. And then if you're an investor in the stock market, a third of your money is in these companies. So it's been great for you up until now, but now what happens? So this table you mentioned here, you know, shows a couple statistics. So the first thing we show is just how profitable these companies actually have been. So over the past decade, the return on invested Capital of the Mag 7 has been 22.5% versus just 6.2% versus for the rest of the S&P 500. That's a pretty big gap. And if you look at return on equity or free cash flow margin, it paints a similar picture. And look, it makes sense. These companies don't need that much tangible capital. You don't need that much physical capital in order to generate these outsized earnings. What you need instead is you need intangible assets. And so that's the bottom of this chart where we show four different intangible intensivity metrics. We show that, you know, as a percentage of their revenue, they're spending, you know, much more than the average company on R and D and marketing. They have a lot more patents, they have, you know, a much greater share of their workforce is in high tech areas like AI. And so across the board, these companies are asset light. They don't use physical capital, they use intangible assets to generate these earnings. And that's worked out really well for them.
- [27:40]
C
Yeah, well, I think that's one of the things people miss when they compare this to the.com is like, you know, you just want to say, oh, those companies were overvalued. These companies overvalued. Like these are much better businesses. I don't know, you can correct me if you think I'm wrong, but these are significantly better businesses than the businesses I was seeing in front of me back then.
- [27:54]
A
I'd say that is true. Yeah, they're better businesses. They have better balance sheets, they are more profitable. Obviously there's going to be core weaves in there. There are some companies as well that are more like the global crossing analogy. But your core companies that compare AT&T 25 years ago to, you know, Google today, I mean Google is definitely the superior business.
- [28:17]
C
Yeah. So as we talk about how that asset light nature is changing, we've got this next chart which is magnificent, magnificent 7 CAPEX to revenue and you can just see that's gone off the charts recently.
- [28:26]
A
Yeah. So like that's the catch. Which is like, okay, these companies have done really well on the back of being asset light. But you know, the unfortunate thing for shareholders is they're now starting to turn their back on that model. Moving from asset light to absolute heavy. The capex, the revenue ratio of this group of stocks has increased from in 2012 it was 4% to 15% now. Right. With that massive acceleration more recently and the streets to be believed and what these CFOs are saying is to be believed, that's not coming down anytime soon. These companies are committed to the ag race.
- [29:00]
C
Do you think longer term they feel like they'll be asset light again? I mean you said it's not coming down anytime soon, but do they feel like know once we get to the other side of this, like there'll be asset light companies again? Or do you think these companies are forever changed now?
- [29:11]
A
I think they're, I think they're forever changed. I mean the depreciation of these chips is pretty fast. It's not like, oh, we build the railroads one time and they're good for 30 years if you build these chips and they're good for somewhere between two to five years. So you're always going to need to be refreshing the chips which are about 60% of data centers, 65%. So you have the, the, you know, physical component, the grid, et cetera that's being built out, that's you know, know more permanent. But yeah, you're always going to need, need to be recycling these things, you know, moving forward as you know, the, the frontier what is, you know, the highest end GPU continues to increase exponentially.
- [29:43]
C
Sometimes you see a chart and it's just like completely, it's like eye openening in terms of how serious or how much the change has been. And this next chart was one of those for me. Because you compared these companies to utility in terms of capex to revenue. What'd you find there?
- [29:56]
A
Yeah, so what this shows here is it lines up all of mag 7 from low to high. So you can see Nvidia and Apple have pretty low capex to revenue ratios whereas like Alphabet has 21%, Microsoft 28 and then Meta 35 which is higher than the average utility at 28 today and higher by the way than the, you know, capex spending of AT&T AT the height of the dot com bubble in 2000 which is only 21%. And so that's the question which is like is are the max 7 becoming quote unquote utilities? Obviously that's a bit of a aggressive statement and you know, there's plenty of qualifiers but at least from a very high level that's what this would suggest that you know, they're transitioning from being kind of like someone else do that. I'm going to focus on these kind of high margin things to. All right, well someone needs to build this out, might as well be me. Let's go build out the AI cloud. And I think what's concerning is that if you look at the data, so similar to the fama French work, right where they look at like the performance of you know, companies that have high growth in capex, if you instead look at the level of capex so look at like for each point in time the companies with the most capital intensives, the most capex to sales versus the least, you find that those companies consistently underperform their peers. Asset light businesses are and historically been better. Asset heavy businesses have been worse. You know, it requires a lot more upkeep. You have to keep spending money just to kind of offset depreciation whether it's due to wear and tear or to obsolescence, you know, due to technology. Technology these companies, they tend to be in more competitive arenas where you know, physical infrastructure is just, you know, there's lower barriers to entry. It's easier for folks to come in there and compete with you like Oracle just coming in. And now they're likely to be a top player in the AI cloud, not, not as hard as more intangible intensive sectors. And that's of course what leads to the capital cycle dynamic which whereas if the barriers to entry are low and competition is high, then when returns are high or hype is high, competitors will come in and it'll create an excess capacity and that'll lead to lower prices, lower profits and then they'll lead to a washout over time. So I think that's very much likely to happen given this transition of these companies from this asset light to this asset heavy business model.
- [32:13]
B
Kai, this wasn't in the paper and I don't know if you know this or not, but what is meta like, what are they spending so much on? Do you have any idea, like, what they're plowing the money into?
- [32:25]
A
Yeah, I mean, AI data centers. I mean, Zuckerberg is a very aggressive CEO. I mean, we saw this with the metaverse, right, in 2022 when he made big investments in the Metaverse, which he ultimately did pull back on. But in that case, he didn't really have any competitors. I think the dynamics here are very different with AI than what we saw in 2022. That was almost like he was going out on a limb himself and trying to make it a thing. Whereas today everyone is doing this. And I think it's highly unlikely that we see these CEOs turn back.
- [32:56]
C
And I have to assume some of this is a function of feeling like he's a little bit behind. I mean, we've talked about the $100 million people he's trying to hire. Maybe the person who feels like they're a little bit behind is going to go a little more aggressive on the spending.
- [33:07]
A
Yeah, I mean, there's certainly that element too. I think there's like a big kind of fomo. There's a bit of an ego thing here too. You know, a lot of different, like, kind of personality dynamics at work.
- [33:19]
C
So this next chart was interesting because this gets at what has been one of the greatest things about these companies, which is the free cash flow they generate. And you're looking at how that's changing here as they start to spend money on this.
- [33:28]
A
Yeah, I mean, look, the most common counter argument is the one, you know, to this being a concerning development is the one that, you know, you already brought up, which is that these companies are so much better than the, the, the railroads for sure. And then the companies in the dot com, that these are not telecoms, these are, you know, real profitable franchises. And that manifests in, you know, free cash flows that have just been, up until recently, extremely insane. And. But what you're starting to see is concerning because over the past few years, as the capex boom has taken off, you know, you can see the net income line has continued to increase, but the free cash flow line has started to turn over. Right. These companies used to pay big dividends, and you can see now they're laying employees off. They started doing that. Like, I think a lot of this is because they're realizing, you know, we're not going to, we need this money to build our data centers. Like we don't have, you know, the luxury of just sitting on, like, you know, infinite cash hoards, like we used to. Right. Like the investors used to criticize these companies, you know, even several years ago, because they were, like, you know, carrying on too much cash on their balance sheet. They weren't. They were, you know, and now we have this opposite problem.
- [34:29]
C
You mentioned these circular deals in the paper, and those are really interesting because those have been in the news a lot recently, and there seem to be a lot of opinions on all different sides. But first, maybe can you just explain when we're talking about circular deals, what are we talking about here?
- [34:40]
A
Yeah, so this is these. These are situations where one AI company will either invest or, you know, either through debt or equity in another AI company who might be a customer or a. Or supplier to them. Right. So it ends up creating this situation where all these companies are kind of interlinked, whether, you know, through equity or debt, in this, you know, web of. Of intricate deals. And it kind of ties them together in a way where they kind of rise and fall together.
- [35:07]
C
Yeah, it was interesting. I was listening to the all in podcast because it's a value guy. That's. That's where I get my tech information. And Brad Gerstner was defending these, and what he was saying is what's different than these from these, from circular deals in the past is that there is an underlying demand there. And so as long as there's an underlying demand there, you know, these deals are really. They are reflecting, like, legitimate revenue, legitimate demand.
- [35:27]
A
Yeah, I mean, I think that, look, there's two things going on here. There's two things that are potentially concerning with deals like this. So the first one is like shenanigans where you're kind of like a shoddy company and you're just trying to make a deal to prop yourself up when you're otherwise failing. I don't think that that's what's happening now. I think that if you're at Nvidia, you want to invest in and support these companies because OpenAI, for example, because their success also leads to more success for you. I mean, it's not like I don't think that there's any kind of concern there. I think where the concern is on the second piece is this idea of entanglement, where, yeah, the Mag 7 have amazing balance sheets, and that's not going away anytime soon. But what's happening is that they're increasingly entangling themselves with firms like OpenAI, which are aggressively investing money they don't have. Oracle, as you mentioned is taking on a lot more debt. Core weave is very financially constrained. And to the extent that that's the case, they do put themselves at risk, even if it's not showing up on their balance sheets. And on that note, by the way, you know, there's something I mentioned, which is the deal from last week where Meta took on $27 billion of debt on to the for the building of their Hyperion data center, but they did it off balance sheet. So they went up an SPV take on the debt. It was the largest forever credit transaction or offering ever. And you know, they, they, they don't have that. That doesn't affect their balance sheets. Not, you know, credit rating agencies won't look at that, but they are obviously backstopping that deal regardless. So I think that's a good example of something that, you know, has happened in the past too, that is, you know, potentially a bit concerning. But again, it's less about like, does it signal fraud? It's more that it like just ties the fate of companies that have, you know, very solid balance sheets together with those that have less solid balance sheets.
- [37:06]
C
This next chart gets into the arms race that we've been talking about, which is this idea, well, you're going to build this big of a data center. Well, I'm going to build this big of a data center, and then somebody's going to come over the top of the bigger data center or more GPUs. And you talked about how that relates to the prisoner's dilemma. So can you talk about that?
- [37:19]
A
Yeah. So the Prisoner's dilemma is a classic game theory problem that you probably learned about in college. But that does describe a lot of dynamics in many ways, including the AI arms race. The setup would basically be this, which is today we have this, basically a cozy oligopoly where these big tech companies have kind of divided up the, you know, tech, various tech verticals. And you know, they take turns, two or three of them kind of own each one. And as a result, there's very little competition and they have very high profits, which is great for them. Now the problem is that now AI comes along and AI is this new disruptive technology, at least in the minds of these CEOs, that, you know, risks basically collapsing all these markets into one where it's not like there's like shopping and, you know, productivity. It's. They're all just the same market. If you have the best AI agent, you win. And so now it's like winner take all going to your point jet and if it's winner take all now, it's become a lot more competitive. And so they're going to have to vie for this new market. And the challenge is this, which is. And this is where the prisoners dilemma comes in. Because in the ideal world, what would happen is that these companies would kind of agree collectively to moderate their investment, say, look, we don't want to disrupt ourselves. Like, this is a pretty good situation we have right now. Let's invest in the next technology, we'll do it slowly, and we can maintain our monopolies at its quo. The problem is that any single company's unilaterally incentivized its effect. In other words, they're incentivized to aggressively invest instead and try to capture the market. And what's happened in practice is that you have this, you know, you have Google, who invented the transformer, but they don't really want to disrupt themselves. And then you have Sam Altman in OpenAI who's like, wait a second, we have nothing to lose, let's go for it. And so he's like going all in on this stuff and he's basically bringing along these, these much bigger players with him. Because if OpenAI invests all this money in building out artificial general intelligence, well, these other firms can't afford to sit on their hand and watch them do that. They have to play too. And so what that leads to is a different equilibrium where now everyone is aggressively investing. And what that means is you end up with, at least in theory, an oversupply of capital, an over investment, where now we have a stable equilibrium that is suboptimal, where these companies are kind of net losers. But, you know, because of the incentives to any individual actor, this is just where we end up. And that's the capital cycle idea where you have, you know, an overinvestment and overbuild excess capacity over overly high competition, falling prices, and, you know, ultimately low profit pool for the, for the industry.
- [39:59]
C
How much truth do you think there is to this winner take all thing? I mean, do you think there's, do you think that is the most likely outcome here, that like one firm is going to dominate this whole thing? Or do you think the most likely outcome is all these firms will, maybe they'll develop different niches or something and they'll all have successful businesses and maybe it won't end up being as winner take all, as people think.
- [40:15]
A
So I think you have to be very careful to define, like the different, like layers of the stack, right? So like there's like the model layer, there's like the chip layer and so on and so forth. You know, I think most people think about the model layer as being kind of the most competitive, right? Whereas you know, I was doing this like in 2019, 2020, I was training my own large language models using open source, like BERT for example, the original Google one, which was like a GPT1 level model or. And then over time, you know, you saw OpenAI kind of quickly, you know, pull ahead with their closed source models, you know, two and then three in particular was very good. And so there was a long period of time where there was a period of time when they had their models were significantly better than everyone else's. But then you found Gemini caught up and even the open source models are pretty good now. And so there's an argument to be made that maybe what will happen is that there will be commoditization at the model layer where it doesn't really matter. Everyone will just train on the same dataset. It's just going to use the Internet and Wikipedia, so on and so forth, and you can have your own in house, RL teams, et cetera, but it's going to be pretty similar in that state of the world. And again, it all comes down to competition that there's not gonna be much profits then in that category because there's too much competition. But there might be profits in a different layer of the stack, say the guys who use the underlying intelligence but then train and fine tune on a specific domain dataset or have distribution to end customers who really want access to that product or something like that. We'll see. Also at the chip level, like right now, Nvidia had the monopoly for various reasons, including Cuda. Like obviously Google is trying to develop TPUs and other companies as well. OpenAI made a deal with PMD, their chief rival, a week after the OpenAI deal. So look, there's competition entering that space. It's the capital cycle. If you see high profits in a sector, you're going to want to go as a capitalist, compete against that and maybe you'll succeed, maybe you'll fail. But to the extent where there is less barriers to entry and less differentiation, you're more likely to see kind of perfect competition and zero profits than others. I think it's going to be interesting to watch how it plays out at the various levels of the stack. But I do think that this tendency towards an overinvestment is quite likely. You have Larry Pages on the record saying I'd rather go bankrupt than lose this race. These guys are really very much committed to trying to win this race.
- [42:42]
C
Just as an aside at the model level, I do feel when I use them now, certain models are much better for certain tasks than other models. And so maybe there will be a nature where, you know, certain models will gravitate towards certain types of tasks and be better at that than all the other models.
- [42:55]
A
Yeah, I think that's, that's entirely possible. I mean a lot of the companies now are working on like these like vertical, for example, like you saw the announcement, I think it was last week that like OpenAI is like, you know, hiring the ex investment bankers to like train their model on Excel for financial modeling. Right. And we're seeing that as well in like the legal and accounting fields. And these companies are starting to, to do that to build a kind of vertical applications of their models. So that could be where we end up where maybe like one model is the best at like financial applications, another one the best at, you know, taxes or something. So that, that's, that's possible as well.
- [43:31]
C
So just one more for me before I hand it back to Justin. You've got this interesting chart and we went way back here, we, we looked at railroads. You looked at the cost of railroad construction and earnings and contribution to gdp. So what do we take from this?
- [43:42]
A
Yeah, so this is from Azim Azhar, who wrote an excellent blog. What I wanted to do was just go back here and understand where is value captured. Going back to the AI stack idea. It's all about competition in these big infrastructure booms. And this is an interesting case study of the railroad boom. If you look at the purple line, you can see the annual cost of railroad construction and it fluctuates, right? So like around, you know, 1870, it looks like there was a bit of a peak. And then the panic happened and it collapsed. And it peaked again, then collapsed, booming bust. Boom, bust, booming bust. And if you look at the green line, that's the revenue or the earnings, I'm sorry, of the, you know, railroad industry through that time period. And you can kind of see these periods where like the cost, they go through a building cycle and then they crash. Right? And each time that happens, you see this, you know, big wipeout, hundreds of railroads go bankrupt and there's a consolidation and then they kind of do rinse and repeat. So that happened for a period of 20 years, right. That the industry really struggled to make it through. And then eventually they made it to the other side and profit started to climb and costs kind of stabilized the build out more or less finished with these long useful life for rails and then moving forward. Railroads are a industry. They're not extremely profitable, but they're not bad. They exist. What was really interesting is you can see the third line on this chart is the contribution of the railroads at gdp. And that line basically went exponential even as early as the 1860s and 1870s and then just took off. What's amazing with this chart is the extent to which the railroad industry itself didn't make any money in round numbers, but yet they contributed such a huge share to gdp. So who actually won? That's the question. Obviously consumers. Obviously. You know, if you're a business trying to ship stuff from you know, the west coast to the east coast or vice versa, if you're a consumer trying to take a ride, you know, but, but so, so it's the customers, you know, whether businesses or individuals that did do really well. And like another example would be like the dot com boom. This is like a classic case study where as I mentioned after you know, global crossing failed, they, you know, the price of fiverr went down like 90% per, you know, went down 90% and that allowed companies like Netflix and Facebook to then you know, build out massive businesses on the back of this subsidy subsidized Internet. So basically what's happening here is the like two categories is the infrastructure builders, the telecoms, the railroads, the you know, and today the data center builders. And then there's the customers of those services which are the Netflixes and Facebooks in this example. And you're seeing because of the overbuild which what happens is that it means that the infrastructure builders get hit. But what then also happens, interestingly is they then subsidize their customers because like the competition brings down prices allowing these companies, their customers to thrive. And so that's why it does make you question without even getting to the final section whether or not where you want to position in the stack. Maybe infrastructure is not the place to be as we'll see. As I've shown in other places, um, in the beginning of a cycle, like think about the dot com boom. In the beginning of the dot com boom actually telecoms were the best performing sector. They you know, handily outperformed the general tech sector up until 2000 and then they collapsed 92% and never recovered. So there is an argument that it's dynamic. But if you look over the long arc of time in general you found that the railroads, the telecoms were not good investments. Yet the companies that they served were actually quite benefited by the technology. And to the extent that there is an overbuild, which maybe there will be, maybe they won't, you're actually positively exposed. Right. You're antifragile, I guess, with respect to CapEx, if you're in this other layer.
- [47:33]
B
And so I think that plays into this next part, which is where are the opportunities? And what you had done a few years ago is published this paper, Investing in AI, Navigating the Hype, where you actually built like an AI investment financial analyst to find where maybe some of these hidden opportunities or under the radar opportunities might be. So maybe before we get into what those might look like or what those are, just talk to how this, like financial analyst, like, what type of data did you use and what type of process did you use to come up with this investment strategy?
- [48:10]
A
Yeah, so look, there's like going back to the framework, there's two categories. There's the infrastructure players and then the early adopters. So infrastructure, they're kind of obvious. Nvidia, other chip makers, Core Weave, et cetera. Like JP Morgan had a list. But the early adopters are harder to find because when I say early adopters, all I really mean is the everyone else section of all other companies, the companies within that mass that are actually investing in AI and not just talking about it. So one thing that we've seen since that paper came out over the past two years has been an increase in the amount of chatter in both filings and earnings calls of CEOs talking about how they're doing AI. Everyone's saying they're doing AI because it's good for the stock price. But you need to get away from that theater because anyone can say they're doing AI who's actually investing in this technology. And that's where a lot of the unstructured data and alternative data that is described in that other paper comes into play. So things like trademarks, you would get a trademark if you're planning on launching an AI product, AI powered product, job postings or LinkedIn bios of your employees. So we're trying to find these kind of hidden signals for companies that are truly investing or positioning to use AI in their businesses as opposed to just talking about it or not even at all.
- [49:29]
B
Were there any names that kind of really surprised you that kind of rose to the top? I mean, I'm seeing like Caterpillar in here, Walmart. I mean these, these are big Fortune 500 companies, obviously. And on the S&P 500, that are, you know, trying to incorporate the use of AI to probably to help with operations, help become more efficient, do all that kind of stuff.
- [49:47]
A
Yeah, I mean look, I think the use cases of AI, if it is indeed a general first technology, should be basically everything, right. Whether you're, you know, on automating the robotics in your factory or your supply chain or you're doing predictive analytics on your customers, if you're a retailer or drug discovery, there should be use cases across all industries, not just in software. So the obvious names are like Palantir. Of course everyone associates Palantir with AI and they are using it. But you find a lot of other companies that maybe are kind of later in that journey or earlier in that journey, but that are relative to their say stodgier peers still incorporating technology in a way that's, you know, hopefully will meaningfully give them an edge over the, the laggards within their sector.
- [50:32]
B
And what about on the sector level? Like what did you find there that might be interesting? I mean obviously the AI infrastructure is all in communications and it. Oh, consumer discretionary too to some extent. But was there anything that sort of jumped out at you as.
- [50:47]
A
Yeah, look, there's a con. So this exhibit we looked at like for each sector, it's market cap. What percentage is in AI beneficiaries, either early adopters or infrastructure and then what percentage is not. And what you found is that on the infra side, communications, IT and discretionary, which is just Amazon really have a pretty high concentration. And then there's some in energy, real estate, utilities. So similar to what we saw with the capex growth chart. But what's kind of more shocking and more surprising is that the early adopters are way more spread out. Right. So of course, you know, technology and communications, those are kind of naturally going to be the first place you see, you know, adoption. I actually in that other paper did a chart where I showed based on an occupational level which sector have, which sector's workforces are most exposed to LLM automation and aside from was financials, which I guess interestingly shows up as a third on this list here, right. You know, all our jobs and then industrials, healthcare, consumer staples. So you're seeing like yes, there's less penetration within healthcare as in it, but it's not that different. Right. The much more balanced exposure than what we saw on the infrastructure side. And I think the same thing applies if you go to the geographic thing. So we basically do the same breakdown by side by country and we say for each country based on their own individual market, what percentage are AI beneficiaries versus not? And I guess on the infrastructure side, what you see is obviously we all think of the US as being kind of an AI market. And it is like AI stocks and infrastocks are a big percentage over this market. But you got to remember the supply chain is also global. You have ASML in the Netherlands, you have TSMC in Taiwan. And relative to their much smaller markets, they're actually even more AI exposed than the us. And the same point can be made on the early adopter side, which is, yeah, the US has lots of companies that are pretty modern and thinking about AI, but in fact Germany and Israel are actually scoring even higher as a percentage of their total smaller markets on AI early adopters. And then even after the US it doesn't drop off that fast. It's like Japan, Switzerland, China, Canada are all pretty serious about this. So I think the thing is that we all know that if you invest in the S&P 400, you're going to get a lot of AI exposure. But what's interesting is if you also invest in say the German DAX or whatever, you're actually getting pretty use in AI exposure too, just because the markets are so much smaller. Again, we don't think about the German companies doing AI stuff, but the market's just so much smaller as an overall percentage, you're still getting meaningful exposure.
- [53:19]
B
And what about capital intensity? That's the next chart with infrastructure versus the early adopters.
- [53:26]
A
Yeah, so this goes back to kind of the Mag 7 chart where we saw capex to sales go from 4 to 15. You're seeing a similar dynamic here where there's always been a gap between the more asset light early adopters, which tend to be more like software or whatever, and then the infrastructure players. So in 2015 it was like six percentage points for AI early adopters and then nine for infrastructure. But that gap has widened over time. So now AI early adopters have a capex to sales ratio of only 4 and infrastructure firms have like 15. Right. So that gap has really widened a lot and is not set to come down. Right. And to the extent that, you know, you believe the argument from earlier that empirically we found that asset heavy industries and businesses are inferior, they have lower returns on capital, they require a lot more capital just to kind of tread water because it depreciation. That's a concerning thing for this segment.
- [54:18]
B
Yeah, and that sort of plays into, you know, this next chart showing just how much of a premium valuation that these infrastructure Plays have relative to, you know, the early adopters. And I think in your paper you said that these valuation premiums maybe since 2015, I don't know if that's have, you know, gone from 32 to 137% which is pretty, pretty crazy.
- [54:44]
A
Yeah, I mean that's the other thing which is like I mentioned earlier that like in the dot com boom the telecoms actually outperformed the beginning of the cycle. Before they didn't. Right. And that's the same thing. Same thing seems to be happening here with infrastructure stocks. Nvidia, these obvious, you know, AI, big AI infrastructure companies have really outperformed over the past few years and kind of left behind, you know, these more under the radar names like Caterpillar. Right. I guess they've done okay actually. But, but they, but that as a, as a category that that group has not really appreciated in the same way that the kind of, you know, more memeified or hype, hypey infrastocks have.
- [55:22]
B
So walk us through this is like a walk down history lane kind of. So this is, you know, going back to the dot com boom and bust and sort of what happened with the NASDAQ index relative to the telecom index. And then as you go through that, you know, you make the point in the article and this is something that I think a lot of investors that didn't go through that period might not know that, you know, the fundamentals of a lot of those companies that made it through actually did really well for like a decade, but their stock prices basically didn't recover until like a decade later or until like maybe 09 to 00:10 whenever it was. So walk us through that because that's a very important thing to I think understand when looking about, you know, at these boom and bust cycles combined with valuations.
- [56:05]
A
Yeah, look, I mean this goes back to the idea from before that like, you know, just because I think AI is a transformative technology doesn't mean I want to make any money investing in the stocks that do that. And the same thing applies here where there's like, there's two ways to lose money. Right. One is, I would call it fundamental, which is you spend a trillion dollars building a data center that no one uses. Okay, so that's, you just like wrote off a trillion dollars of capex. But the other one is not even that. It's valuation risk, it's asset light companies, let's say, who aren't actually putting up that much money. But the thing is that their stock valuations, their multiples trade at pretty high extremes because investors are assuming there's going to be tremendous growth in the future. And the problem is that if those valuations revert back to a normal level, you will lose money and you may never get it back again. So even asset light companies and asset heavy companies as well face valuation risk. All companies have this. And in the case of the dot com boom, that was the main driver of losses forthese.comdarlings. so basically what I show here is starting in 2000, a basket of kind of highly high flying dot com stocks. And what you find is that over the next 20 years after the crash their sales actually 11x. So they actually did realize the promise of the Internet. Right. They grew very quickly. The problem is that at the same time their multiples collapsed. So their price of sales was 33. Going into the bust it fell to like 5 and that led to 85% return over two years. And if you look at the blue line, that's just a combination of fundamental growth and then multiple compression just mathematically. And what you find is that the, the total return went down 80% in the first two years after the bust and then it took another 18 years to claw your way back out to, to back to break even. But yes, they eventually made it right. But like it took many, many years to get back which is not a great outcome for investors.
- [58:03]
B
Right. And so you sort of then asked the question like if you were to take your framework for assessing intangible value and trying to buy those companies than you know, using a more sort of, you know, the way that you look at intangible value, which I'll let you talk through what those are, you know, what did you find there in terms of the, the hypothetical performance at least.
- [58:28]
A
Yeah. So I mean let's talk about the framework first and like why it's even required. So like you know, your starting point should just be to use the tools that are kind of widely available, which is like price to book ratio. We mentioned Fama French as a starting point. The problem with like traditional value like that is that it is, it tends to be very focused on tangible capital as like a definition of intrinsic value. And what that means is that you're naturally going to be biased against companies that are asset light, whether it's, you know, companies with strong brands or human capital or more relevant in this situation, companies with strong intellectual property. And you know, the whole point of investing through a hype cycle is to be long innovation. Well, you're going to miss out on all those companies Right. You think about traditional value investors through the dot com boom. There are many famous examples of guys who just sat out the entire thing, which of course lead to missing out on a lot of upside. So is there a better way of doing things? Is there a way where you can remove that anti innovation bias, still be long innovation, but also have a value filter whereby when stocks get overheated, you can rotate out, take profits and move to other ones and then recycle. And that was the idea here, which is if we improve our valuation ratios to include not just tangible but also intangible capital. So you're now looking at price to a more holistic definition of intrinsic value. The idea would be that we can now approach these periods and just the tech sector in general in a way where we can now apply the value lens into the sector without just being always hating tech stock. Which is kind of what we show here in this again, hypothetical performance in this, in this chart here, which is to show that basically the gray line shows if you just took all Internet stocks over this period, what would happen on a relative basis, so relative to the market, and you would find that you beat the market for the first half and then you underperform the market after the bust. And it was a full beautiful round trip, like a perfect round trip, which is not a great outcome. So is there a better thing you can do? What if you instead did this, where you divided the universe up into a cheap half, an expensive half, unintangable value, and each month rebalanced, so you dynamically rebalanced. As stocks got expensive, you would kick them out, or as stocks fell on the bust and got cheap, you'd buy them back. That's what the blue line shows. And so if you do that, you achieve a decent performance, a consistent performance both on the boom and the bust outperformance against the market. The problem is that the expensive half of stocks basically fully offset those gains. So whereas the cheap half outperformed, the expensive half underperformed exactly offsetting the gains over this period. So what you want to do obviously, if this is to be the indication is you want to own the cheap half. If you just stay away from the expensive stocks, you can still be long innovation, still be long the intranet. And that beautiful green line, that fundamental growth line, the 11x in sales we saw, while also avoiding a lot of the pain from the bust associated with this multiple expansion compression risk. So you try and take valuation risk off the table. And that was the goal, that was what we tried to demonstrate here. And how that would have worked in this episode.
- [61:37]
B
Yeah, that's great. And people in that 2023 article, you can see kind of what type of names would have been in that portfolio back then if that was based on, you know, what your system was finding attractive. So if we were to, if we were to take this process and framework of intangible value and apply it to the AI universe, what would the portfolio basically look like today?
- [62:07]
A
Yeah, so about 20% would be in infrastructure stocks. So again, like, I'm not saying they're all bad, but just, you know, on a relative to how they were before, it's a little tougher place to play. So you still have Google and Amazon in here and Micron, the other example being given, and then the other 80% or so would be in early adopters. And that across not just technology, but also financials, communications, industrials, consumer discretionary, healthcare, consumer staples and energy.
- [62:37]
B
And what about how it's shifted? So in 2020 you had a certain framework or a certain type of exposures. And how has that sort of changed over time here?
- [62:48]
A
Yeah, so that's the interesting thing, right? This is all bottoms up. It's all looking at the valuations of these companies relative to their intangible augmented fundamentals. And so in 2020 the portfolio would have been almost half infrastructure of which chip makers, so like Nvidia and other semiconductor companies would have been a large component and then early adopters would have been maybe 60%. So that shifted obviously away from infrastructure companies. As infrastructure companies, as we saw earlier, have become more and more expensive as the market now fully values Nvidia, those companies have been kind of taken, rotated out of and then into more of the early adopters which continue to be under the radar and also more asset light.
- [63:29]
B
This is awesome stuff. Kai, what do you think if you were to try to bring all this to some type of conclusion, the way that investors should be thinking about the markets today and you know, investing in these types of companies and Also even the Mag 7 we've talked about a lot, like what do you think sort of the final conclusions of this paper are for you?
- [63:54]
A
Yeah, I think like, all right, so if you go to the capital cycle, like what's the mean insight of the capital cycle? The insight is that people underrate the supply side. Everyone's here trying to figure out what is demand going to be, is revenue going to 100x right. Will OpenAI continue to be the fastest growing consumer app ever and monetize their users through Sora? Maybe, maybe not. That's a really, really hard thing to predict. But what we do know is the supply side. We do know that these companies are making tangible investments into infrastructure. Can see the deals, these are, you know, multi year deals that have to be done. It takes years to build power plants and such. And we're seeing that it's, it's happening, right? And so you can project that forward, you know, hundreds of billions, trillions of dollars of capacity will be coming online over the next five years unless something drastic changes. And so I think that with that in mind, you know, and, and that's more a pretty low, a pretty narrow confidence band and with demand, you know, could really be anywhere. You have to ask the question which is, you know, if you are invested as a stock, a stock market investor in the index, you have 33% of your money or whatever in the Max 7 and probably 40 something percent, you know, across the major AI infra stocks. Like, is that where you want to be given, you know, what we learned about history and given, you know, the amount of capacity coming online that you need to, you know, see a heroic increase in AI revenues in order simply to justify that. And perhaps there are, you know, more, you know, there are other ways of investing in AI. So you still want to be long innovation, you still want to be long AI. I think, you know, like many people, that it will be a transformative technology, but you want to do so in a way that maybe is less exposed on the capital intensity side and perhaps less exposed on the valuation side as well.
- [65:33]
B
The one thing I'll say too is that these types of, I guess cycles, the boom and bust aspect of them, you know, a lot of times to your point, it's all, it's like people are asking questions later. It's like spend money, invest. And I feel like this is one of the first pieces I've seen at least that is really. And it's not questioning anything, it's just the reality of like the way that these things sort of shake out, at least historically and looking at how this is like transforming the market and these big companies like right underneath our feet. And so, you know, I think that it's, it's, it's great that there's people out there like you that are thinking deeply about this stuff because it's, you know, hundreds of billions, probably trillions of dollars, you know, decked up and invested in these companies and you know, the outcomes here are going to be pretty important and massive and so really appreciate, you know, you doing all this work and certainly sharing like all this with us. We're very fortunate. We get to sit down and our audience gets to learn from you. So thank you very much, Kai. Really appreciate it.
- [66:50]
A
Thanks for having me.
- [66:51]
B
Thank you for tuning in to this episode. If you found this discussion interesting and valuable, please subscribe on your favorite audio platform or on YouTube. You can also follow all the podcasts in the Excess Returns network@excessreturnspod.com if you have any feedback or questions, you can contact us@excess returnspodmail.com no information on this.
- [67:10]
A
Podcast should be construed as investment advice. Securities discussed in the podcast may be holdings of the firms of the hosts or their clients.