Full Transcript: Michael Mauboussin on AI, Base Rates, and Intangibles
How Base Rates and Intangible Assets Shape the AI Investing Landscape
Kai: Welcome to the first episode of The Intangible Economy with Kai Wu, where we explore how intangible assets, innovation, and technological change are reshaping investing. I can’t think of a better guest to kick this off than Michael Mauboussin. Michael, welcome.
Michael: Thanks, Kai. And by the way, I’m just super honored to be one of your first guests and I’m a huge fan of the work that you do, so I’m really, really looking forward to our conversation today.
Kai: Me too. A lot of my research has been inspired by the work you’ve done over the years. I think this will be a lot of fun. Okay, so let’s kick it off. Dive right in. One of the biggest questions today, of course, is how investors should be thinking about navigating the current AI boom. Companies are spending trillions of dollars on AI infrastructure, presumably with this expectation of huge returns at some point in the future. You recently published a piece called “Bayes and Base Rates,” assessing the plausibility of these expectations. Talk me through what the main question you were trying to answer is, and why you felt base rates provided a useful framework for doing this.
Michael: Yeah, maybe just to take one step back to make sure all the listeners are on the same page. You know, usually when you think about how do you make a forecast, the common way to do that — and probably resonates with most people even when I say it — is you kind of do a lot of work. You know, you do ground-up work, you gather lots of information, you combine it with your own experience and inputs, and then you project into the future. So usually when you see an analyst forecast, or even a company forecast, that’s typically what they’re doing.
Another way to think about how to make a forecast is to use so-called base rates — also called the outside view. And now rather than building it up from the bottom, you’re saying, let’s think about this problem as an instance of a larger reference class. So you’re basically saying, like, in history, what happened when other people or organizations were in these situations before? How did that all turn out?
So in the fall — to your point — I was looking at all these numbers. Two of them that really popped off the page: one was some of the stuff OpenAI was saying, and the other was Oracle within their cloud business. But just to give you a sense of the OpenAI numbers, specifically in 2024, they did revenues at $3.7 billion and they were forecasting for 2029 $145 billion. So that’s a 108% compound annual growth rate. And the question then becomes, like, how many companies of that size have ever grown 108% compounded annually for five years?
So for our reference class, we went back to every US public company since 1950. So this is basically the Compustat database. Our reference class was companies with initial revenues between two and $5 billion. You want to start at the right kind of level. And we asked — and by the way, it turns out when you do that, 75 years of data, you have about 18,900 firm years or things you can examine — and the answer is no company had ever done it before. Right? And it turns out the average growth rate was around 7%, the standard deviation around 10.6%.
By the way, the feedback I got: people called me up and they’re like, “Oh no, you should be looking at just technology” or “You should be looking at just software.” And of course, all those numbers are in there. So if you remove the slower-growing things, it does move the mean up and does move the standard deviation up. But it still obviously doesn’t change the basic fact that no company’s ever done it before. So with the standard numbers, it’s like a nine-and-a-half standard deviation event, right? Something that seems implausible.
So my point was less that they can’t do it — it’s not like a physical impossibility, to state the obvious. Base rates are not things that are handed down on tablets from on high, right? These are living and breathing distributions. But when you’re considering the possibility of something like this, you want to bear in mind that you’re talking about something that’s never been achieved before. That’s probably not going to be your base case, like your most likely scenario of unfolding. It may be a scenario where you place some probability on it, but unlikely to be that.
Now, what’s interesting — amazingly, even last month, OpenAI actually revised up their estimates for their 2029 revenues. So they now have it at $185 billion or $184 billion. So actually, if you stay with that 2024 base, they’re now looking at 118% compound annual growth, which is actually pretty extraordinary.
So now the next question would be something like, is that possible? Is it plausible for them to do this? And if you want to revise your views up and be bullish on it, there are a couple of ways to think about that. One is in 2025, they actually did $13 billion of revenue — that’s 250% growth. So their growth rate in the first year of the five years was well ahead of expectations.
The second is to look at basic diffusion, right? So you say, how fast is this technology diffusing versus other technologies? As we know, sort of famously, ChatGPT got to a hundred million users in two months. It took TikTok — which was also extraordinary — nine months. Took Instagram 28 months. Took Facebook four and a half years. So they really are fast out of the blocks in terms of diffusion.
I’ll just say, Kai, that we, you know, like to think — when we think about total addressable market — we kind of come at that with three different lenses. The first lens is sort of that bottoms-up build-up: you know, how many customers could possibly use this, how much are they gonna pay, and so on and so forth. So that would be a very standard way to do this. The second is to look at diffusion models. And diffusion models are often underutilized in the financial community, but they’re super interesting to look at to see how diffusion works. There are simple models — even things like the Bass diffusion model — it’s a three-parameter model that can give you some sense of that. And then the last is base rates. So we try to triangulate these three different things to think about what the potential is.
So again, I’m not saying any of this should be ruled out. In fact, they’re running ahead of pace on this. But by the same token, just acknowledging that it’s a tall order to get to those kinds of growth rates they’re talking about.
Kai: That’s right. You’re saying that, look, this is not saying that this couldn’t happen, but if it did, it would be unprecedented.
Michael: Unprecedented. Yeah. And then for OpenAI specifically, just sort of to say, why is this difficult? One is you have to have a great product, and we could talk more about that. But they have to have a great product — obviously it’s very competitive and these things are all related to one another. The second is you have to have great people, right? So you need great people to create that product, and there’s obviously a very dynamic market for talent right now. And then the third is, in their particular case — as you mentioned in your intro — they have to raise an enormous amount of capital.
The latest number I’ve seen is they’re going to burn through something like $218 billion of cash before they go to free cash flow neutral. And if you go back to even recently, you know, the big Ubers and so on and so forth — none of those guys burned anything close to that much money. So it means you have to raise a lot of capital. You have to have a great product, have great people, and raise a lot of capital. You know, those are the three legs of the stool. All three of them are kind of hard to do, but doing all of them simultaneously — that’s the trick. And by the way, they’re on pace to do it all. But that’s the challenge.
Kai: Yeah. Lots of hoops to jump through. That’s a really interesting framing of the question, which — with how much money’s at stake — we should be thoughtful as investors as to what is the likelihood that they managed to nail all these difficult things in succession or together.
You did another interesting analysis in your paper, which drew on this work studying large-scale projects. I grew up in Boston, so I know the Big Dig, of course. When you think about the base rates implied by this study — and tying it to how the AI boom has become so capital intensive, and there’s so much complexity to the build-out and the rollout of these data centers — what is the lesson from this?
Michael: Yeah, and this is another great example of base rates. By the way, Kai, you’ve probably seen this experience. You say, you wanna remodel your kitchen or something, right? So you go pick out all the appliances and you get a contractor, and then you mention it to a friend and say, “Here’s what I think it’s gonna cost and here’s how long I think it’s gonna take.” And what does your friend say immediately without knowing about the details? Whatever it is, double the cost and double the time, right? So people have a reflexive reaction to that.
There’s an economist named Bent Flyvbjerg who wrote a book called How Big Things Get Done, along with Dan Gardner. And it’s really — it’s a book I would highly recommend. I think it’s a terrific book. And Flyvbjerg is famous for building a database of 16,000 projects. And so this is, you know, 20 different sectors, 130-plus countries. As you pointed out, the Big Dig is of course on that. But think of the Sydney Opera House, hosting the Olympics, you know, the Channel Tunnel — all these kinds of things.
And so the question with these 16,000 — and by the way, he’s the guy that coined the term reference class forecasting — so the question is, of these 16,000 projects, you know, how many were done on time, how many were done on time and on budget, and then how many were done on time, on budget, and delivering what they’re supposed to? And the answer is: on budget is less than half of them. On budget and on time is less than 9% of them. And on budget, on time, and actually delivering what they promised at the outset is one half of 1%, right? So here again, just acknowledging that these big projects are very difficult to do.
Now, on the one hand, these AI data centers — as you point out — are complex, tricky things. On the other hand, they are to some degree modular, so modularization actually helps you a little bit. That would be the bullish case. The other complexity — with permitting and energy and cooling and all that — that would be the downside.
Now, there’s an article in The Economist — again, I don’t know if this is right or wrong — but it said about 25% of AI data centers in 2025 were delayed for some reason. And the estimate for 2026 is somewhere around 30 to 50%. So the way to think about this is just: stuff happens along the way, right? Very rarely does anybody lay out a plan and have that plan executed perfectly. There are now 16,000 projects that have been studied to demonstrate that that’s not something you should expect, and you should expect some problems along the way.
Kai: Interesting. So yeah. Let’s go back to an earlier piece you did on the impact of intangible assets on base rates. We’ll get into intangible assets in depth later in the conversation, but that was a really interesting piece because — at the time — investors were quite skeptical of the ability of companies like Amazon to put up sustained growth at such high rates for long periods of time. And your argument was, hey, look, the economy’s transformed. We now have more intangible assets, which have unique properties — say, scalability and stickiness — that potentially reshape the distribution of what’s possible. Both faster growth on the upside, but also more risk on the downside.
As we think about AI as almost being an extension of the technological frontier beyond cloud computing and many of the other intangible assets that Amazon levered on its way to glory — you know, should we be — when we think about, you mentioned some folks pushed back on your base rate saying, “Hey, you should just focus on tech companies only.” And yes, while the forecast baked into these capital investments are unprecedented, maybe AI is the next technology, the next frontier of human intellect and one that perhaps benefits even more from intangible assets in a way that does shift the distribution further. What do you think of this argument? And I know I’m comparing you to yourself in the past, so...
Michael: Exactly. No, I think it’s fabulous. It’s a great argument. And just to summarize what we found — and I think how your work is consistent with this — is that when you think about intangible-intensive businesses versus non-intangible types of businesses, it turns out the average, the means and medians, are not that different for those distributions of returns on capital or growth.
But what is very substantial is that it’s just a much larger standard deviation. So you have more really great businesses and more businesses that go bust, right? So on average you’re seeing about the same thing, but you get much fatter tails.
I wanna unpack — there are a couple of things to say here — but one is on Amazon. You know, I totally agree that that’s a beneficiary. Now, in retrospect, that’s a right-tail event. So they’re growing much faster than what people anticipated.
There was another company in the 1990s — I dunno if anybody remembers this — that was actually talking about their asset-light business and how intangibles were gonna really save them, and a company that had actually extraordinary growth rates. This company, by the way, from 1995 to 2000, grew 61% compounded annually. So that’s not a hundred percent, but 61% compounded annually. And by the way, they ended 2000 with revenues of a hundred billion dollars, right? So they went really — they grew really fast. The name of that company is Enron. And within nine months, that company filed for bankruptcy. So we sort of forget about the other sort of asset-light slash intangible stories where they basically go completely belly up. So to me that’s — you know, that’s an illustration of both sides of that. And you’re absolutely correct to point out that both of those things are likely to happen, and perhaps AI is going to accelerate those things to some degree.
Now, the other thing I’ll mention — and Kai, I think this shows up in your work as well — is that really in this century, so let’s say in the last 25 years or so, large companies — and these are large, like you think about these are the top 10 companies, the Magnificent Seven — these companies are just growing faster than what we have seen big companies grow in the past. And that’s a really interesting phenomenon.
One of the inspirations I take is from work of Jim Bessen. He was an economist at Boston University where he basically argues, hey, you know, the catalyst of this is these guys spending an enormous amount of money on proprietary software that allows them to gain both economies of scale — old-school economies of scale — and differentiation in a way that companies couldn’t do in the past. And then finally, that technology doesn’t really diffuse. So like the secret sauce at Amazon, or the secret sauce at Meta, or the secret sauce at Google does not get spread to the world. As a consequence, that’s allowing these companies to both build these moats, but also grow at rates faster than what we’ve seen historically.
So, as I like to joke, usually when we point out a trend like this, it’s about on the verge of reversing. So maybe it’s gonna reverse, but that does sort of explain a little bit of the whole Magnificent Seven thing and why these big companies have done really well. And, you know, as a little reality check on this one — one of the things we like to look at is, you know, the top 10 companies in the US public equity markets. You know, what share of the market capitalization are they? And the answer is about a third, right, at year end. But they’re by our reckoning about two thirds of the economic profit, right? So if you say economic profit — return on capital minus cost of capital spread times the invested capital — you could do that for the whole market. It’s a positive number, but two thirds of that is gathered by these 10 companies. Just 10 companies.
So in fact, their market cap is actually punching under the weight of their actual economic profit contributions, which tells you the market probably doesn’t believe they’re sustainable forever and so forth. So, so these are all sort of interesting things — it’s almost a counterbalance to this argument that, you know, we’re overweighting these big companies. They actually, at least to date, have had fundamentals to justify, or at least you could argue they could justify, their positions.
Kai: Right. And to tie this back into your base rates analysis, right — perhaps another way of saying what you’re saying is, due to the Jim Bessen argument that we have more scalability due to the rise of intangible assets, companies today — maybe the correct reference class isn’t historical companies with one or even ten billions of revenue. Maybe it’s ten to a hundred billion. Because we were all saying — people were saying back then — “Oh wow, Amazon’s a hundred billion dollar company, there’s no way it gets to a trillion.” And then it did. So relative to the past, due to the presence of intangibles, you’re now seeing these mega companies form on the back of the scalability of these assets.
And perhaps AI is just the next in line. It’d be interesting to see, because we all know that if you look at venture and private market returns, there are certainly examples of companies that do manage to grow at rates similar to what OpenAI thinks they can do, but at much smaller scale historically. I guess the question is, can OpenAI take that same growth rate and apply it at much larger scale, due to the fact that maybe the TAM of AI is bigger? And technology has certain properties that we can get into as well, that are almost an extension of many of the features we’ve seen in the Web 2.0 or internet businesses of the past couple of decades.
Michael: Kai, one thing I’ll chip in on this — and I agree with you a hundred percent — there’s a sort of, if we think about the big LLMs, the frontier models, you could sort of put them into roughly two different camps, right? One is the new companies coming along, starting off by themselves. So this is OpenAI and this is Anthropic. And then the other camp would be, sort of, the legacy guys building their own models. And that would obviously be Gemini and Google as an example. You know, these guys have very different financial resources behind them.
And so this goes back to the way I might think about this — the Clayton Christensen argument about whether it’s a sustaining or disruptive innovation, right? So if AI comes along, it’s actually a sustaining innovation for Google or Alphabet, right? So it actually makes a strong player actually stronger. And by the way, I think there were a lot of concerns — and I think those don’t evaporate — that AI itself could undermine the goose that lays the golden eggs for Google, which is their search business. That doesn’t seem to have completely happened, at least not yet. Versus these new guys that really seem like they’re disruptors.
So it’s a really interesting question, like: is this gonna be — is AI just gonna be a wind in the sails of these big guys that helps the Amazons and Microsofts and so forth? Or is it something that unseats them? So that’ll, that’ll be interesting to see.
And then by the way, the dichotomy in financial resources is also very marked, right? So, you know, Meta and Google and Amazon and Microsoft have a lot of money they can spend on stuff. Whereas these other guys really have their, you know, tin cup in hand and they have to raise a lot of capital. xAI solved that by merging with SpaceX. But those guys still have enormous capital needs just to sort of stay in the game, which is — yeah, which is, which is really interesting. It’d be fun to watch.
Kai: Yeah. I mean, this is a central question of markets today. And I guess it’s a perfect bridge into my next question.
One of the big questions with major technological shifts is less about whether or not they create value, but more who captures that value, right? We have seen many historical cases where a new technology comes around and it’s transformative for society, but the progenitors of that idea don’t actually make any money — and perhaps they even just go bankrupt. So as we think about the AI value chain — you mentioned the hyperscalers and certain model labs, but we can even expand that further to include chip makers like Nvidia, data center operators, the model providers, applications, and then ultimately their customers. If you’re an enterprise company using Claude Code in order to enhance productivity, that would be in the stack as well.
So as we think about the way that value — let’s assume that technology is indeed transformative and it does create a pie of value — how do we think about that being distributed across these things? What are the factors that ultimately will determine where this profit pool accrues?
Michael: So, you know, one of the frameworks I really like a lot on this is the Brandenburger and Stuart framework. And just to make sure everyone’s on the same page, it sort of lays out these four different dimensions: willingness to pay, which is the consumer’s maximum price at which they’re indifferent to buying the good or keeping their money; price for the company; cost for the company; and then willingness to sell — the supplier’s price. So you have these four different markers. The difference between willingness to pay and price is consumer surplus. The difference between price and cost is the company’s surplus or economic profit. And what lives between cost and willingness to sell is the supplier surplus.
Okay. So to your point, Kai, I think historically what you would have to argue is that almost all this ends up going to the consumer ultimately. And that’s because of competition, basically. Right? So you and I both have LLMs or we have some sort of a product and we’re competing with one another, and we want to gain market share. Obviously we try to have the lowest — you know, we wanna focus on our marginal costs — but basically we lower our prices in order to do that.
So I think, I think at a high level, my assumption is a bunch of that value should probably accrue to — we should, ultimately it’s gonna be consumer surplus, right? And there’s already some really interesting research about the magnitude of the consumer surplus that’s already been generated by all this stuff.
Now that said, I think there are a couple of interesting things to consider as well. One is, if you go back to Michael Porter’s sort of classic Michael Porter work, he talks about, you know, sources of competitive advantage, you know, and the big ones being cost leadership and differentiation. But he’s also got this concept called operational effectiveness, which is, you know, the stuff that you do that everybody else has to do — we all have to do to compete. And he basically argues that’s not a source of advantage, right? Because everybody eventually will do the same thing. It’ll get commoditized and so on and so forth.
However, if you look at microeconomics, it turns out that actually doesn’t seem to hold empirically. In other words, some companies are just better at doing the same thing as other companies are. So, for example, really high-quality managers can run a manufacturing facility almost twice as productively as a less-skilled manager. And my guess is — and I think this should be a really big focal point of investors — my guess is that companies themselves will embrace AI with differing degrees of effectiveness. And that operational effectiveness is actually going to be really important as well.
So when you’re talking to companies, you know, one of my colleagues likes to joke that when you’re talking to a company about their AI strategy, you figure out if it’s PowerPoint or Python, right? And I think there’s still a lot of PowerPoint AI strategies versus Python AI. Well, I guess we don’t even know — Claude Code, whatever. But that basic argument, I think, is a really, really important one to bear in mind. So I think ultimately this stuff.
Now that all said, we also know the other thing to think about is sort of first-order and second-order effects. You know, there are now suppliers and we obviously know the semiconductors and so forth have done very well. Nvidia is obviously the most valuable company in the world as a first-order provider.
Now, what’s interesting of course is Nvidia is relatively concentrated in their customers, and almost every one of their customers themselves are trying to develop their own technology to weed themselves off of the reliance on Nvidia. So they’re — you know, we’ll see — it’s obviously, that’s probably all priced in. I’m not saying anything that anybody else doesn’t know. But that’s also something to bear in mind.
Now, Nvidia — by the way, it’s been — the performance of that company financially in the last three or four years is really up there in the hall of fame of great financial performance.
Kai: So I want to go back to something you said, right, which is that economic theory would suggest that competition should drive excess returns to zero. And we certainly — we were talking about this before the call — have seen this in the LLM space, right? Where at one point OpenAI had a huge lead and that was eventually eroded by the progress of their spin-out Anthropic, and then Gemini, and we’re seeing open-source competition as well.
But to take the other side and going back to your Michael Porter framework — barriers to entry is kind of an important concept here, right? To what extent — if you were to make the reverse case and say, Nvidia currently has CUDA, they have a big lead in chip making, of course people see their margins and that’s an opportunity they want to go after — where would you see potentially barriers to entry at various points along the AI stack?
Michael: Yeah, well, the first thing to say is, you know, the LLMs, as far as I can tell — they’re, you know, there are now a handful. You pointed out that ChatGPT was really ahead of everybody else. By the way, ironically, the technology was developed at Google, which is interesting. But there are now — you know, however you want to count it — somewhere between three and six models that are roughly doing similar stuff. So that doesn’t really feel like right now anybody’s really distinguished themselves.
I also don’t — it doesn’t feel like a network-effect business to me, right? The network effect is: the value of a good or service increases as more people use that good or service. You know, whether you use Anthropic or OpenAI, whatever — it doesn’t affect me that much. Now, that was also true of Google and Google search, right? There really — I don’t think search was about network effects. Search was about economies of scale. And that’s a slightly distinct concept.
So there’s an economies-of-scale thing. I think it’s also important now to go back to your point on where are the barriers to entry. We like to say in finance that financial capital is not a barrier to entry, but it feels like this is probably — right now, you know, just the ability to raise capital to fund all these operations seems extraordinary. Now when you think about the big hyperscalers — and the hyperscalers are doing the traditional cloud stuff as well as AI stuff — Amazon: ridiculous resources. Microsoft: ridiculous resources. Google: ridiculous resources, right? So trying to pony up, you know... Oracle is like number four in that business and, you know, they’re in a much weaker financial position. They have very large aspirations, very large goals specifically about their revenue growth. But whether they can pull that off financially, I think remains to be seen.
So I don’t have a good answer to this, but I think it’s a really interesting question. And then the other one is just talent. You know, a lot of this is about talent. You think about, you know, how did xAI get going quickly — the answer is they hired a bunch of people from OpenAI. They basically have the secret sauce, or some components of the secret sauce, in their minds. They go back and replicate what they’re doing, right? So you’re seeing a lot of talent sloshing around. That’ll be another interesting component to all this: who can preserve and maintain the talent.
And by the way, it goes back to capital raising. A lot of the compensation of these people is in the form of equity. Many of these companies, some of the companies, are private. So of course when they’re up rounds in private markets, it looks like at least your wealth feels like it’s going up. If and when a bunch of these companies go public — if by whatever, for whatever reason, their stocks go down versus up — that just makes the market for talent that much more difficult, right? So you’re what you’ve promised in terms of compensation, what people’s expectations are in compensation, aren’t met. Maybe that means you give out more equity, which is dilutive for the other shareholders. So all this stuff will be fascinating to see how it all unfolds over time.
Kai: Interesting. So you’re saying, of the various moats that Porter talks about, capital requirements and economies of scale are almost in your mind the most important thing here. It’s less of a technical moat. You mentioned human capital, but not brand interestingly. And when it comes to capital raising and just the ability to — if you’re Sam Altman — take a trip to the Middle East and come back with a trillion dollars or whatever, it’s really interesting.
And I think it gets to the next question I have, which is around the game theory of this whole thing. Because in some ways, one barrier to entry is just the threat of a massive war by which the competitors — especially the smaller ones — get bled dry, right? If you’re Google, this is the hand you might consider playing. So let’s talk about the competitive dynamics around especially last summer, when we saw a series of escalating announcements. A firm saying, “I’m gonna commit X.” Okay, they get X, that goes over the top — two X, four X, eight X. This whole dynamic — and you pointed this out as well — in past capital cycles, I think about the telecom bubble, the railroads, we’ve seen a consistent pattern of overinvestment into infrastructure, which has led to poor or net poor returns, bankruptcies for the builders.
And I’m sure that all these CEOs have seen the data and they know the historical pattern, and yet they persist in doing this. You mentioned something really interesting in your paper around a kind of preemptive strategy of deterrence. Maybe you could walk me through that — it’s a really interesting potential take on what this could actually be.
Michael: Yeah, no, Kai, you’re exactly right. And I, you know, encouraged me to go back and reread Porter. And by the way, this is like from the 1980 Porter book and it’s actually pretty good stuff. And it’s this idea of preemption, right? Which is — you want to try to lock up the market, the resources of the market, essentially to discourage competition, right?
And so by our tally — I might be off by a bit — but OpenAI did 15 deals in 2025, right? And you remember in the fall there was just a flurry of them. It felt like multiple deals per week. And these were very large deals coming out. And basically if you’re a competitor looking at this, you might be saying like, “Gee, how am I gonna keep up with that?”
Now what’s interesting: already this year, the OpenAI-Oracle deal — they’re not gonna expand Stargate, which is one of their big AI data center projects. And in a way, they were expecting this. Just this week we had Sora, they shut that down. They had a billion-dollar deal with Disney, a bunch of copyright stuff as well with Disney, that’s gone. So you’re seeing some of these things being walked back. And by the way, even many of these staggering headline numbers were not really financial commitments — they were sort of more aspirational to some degree.
But it is — it’s like, you know, this is just the peacock flipping its feathers, showing its feathers, or any animal becoming large to try to scare everybody else off. Now the other thing — again, just to echo this point — you know, the challenges that both Anthropic and OpenAI have is they have to raise a lot of capital. And so, you know, you can sort of signal all you want to the world, but at some point you have to be able to back it up with having money, you know, some sort of money to do that.
I think the other thing — and this goes back to your point about historically this has led to some difficulties and bankruptcies — the challenge here really is you have to make huge commitments before you learn about the economics of the business. We don’t really know what the economics are. And that’s, I think, the thing that’s giving pause — even to the hyperscalers, right? I think that’s what’s giving pause to everybody. You’re seeing these CapEx numbers ramp up a lot, and I think the companies themselves believe they’re gonna do fine, but they often almost think internally. They don’t think game-theoretically. They’re not thinking about what everybody else is doing — they’re thinking about what they’re doing. And it remains to be seen what those economics actually turn out to be.
So that’s what the risk is: you know, you have to basically put up a big ante before you know the hand. And time will tell, I guess.
Kai: Yeah. And I think for the average investor, the big risk is that most of us are in the S&P 500, which is a cap-weighted index — of which a third of the index is in Mag Seven, these hyperscaler companies. And almost a half is in the collection of AI infrastructure companies, which includes Oracle and some of these other names, right? So it’s like Game of Thrones, right? These guys are playing a high-stakes game of poker. And we as the investors are kind of sitting here, perhaps unwittingly — unclear on that — with half our money being bet on this gambit paying out.
As you say, it remains to be seen. But certainly something of concern to investors today.
So I’d like to switch gears now, to think more high level — not just about AI, but what drives modern businesses today. And we touched upon this a few times: this idea of intangible assets. I think it would probably be helpful for the listeners just to maybe start from the basics. Can you maybe just define what intangible assets are and why do they matter? Like, why do investors even care about them?
Michael: So there are two aspects of this, and by the way, again, Kai, I, you know, I think your listeners probably know this, but you’ve done some of the best work on this that’s out there. So I would also recommend that everybody go — if they haven’t, I’m sure they’ve read all your stuff, but if they haven’t, they can, they can go back and revisit it and it’ll be to their benefit.
Well, an intangible is basically something that’s not tangible, right? So a tangible asset is something that you can look at, touch, feel. An intangible doesn’t have those characteristics. So the kinds of things that would typically be in that bucket would include things like software code, research and development, advertising, training of your employees.
The key issue is that intangibles typically show up on the income statement as an expense in the SG&A line — selling, general, and administrative line. So rather than a physical asset — like you buy a machine, if you’re a company, or a factory, it goes on the balance sheet and gets depreciated over time — in these cases, the intangible is expensed immediately.
Now, just to give a little bit of level-setting — these are our data, so other people might have slightly different data. The first thing I would say, this is actually not our work. Going back to tangible investments — these are by macro economists — tangible investments were about 1.4 times intangible investments in the late 1970s. So call that about 50 years ago, something like that. And by our numbers, intangible investments now are — it’s almost the flip — about 1.5 times tangible, right? So if you’re just taking a 50-year point of view, things have really changed quite markedly.
By the way, Kai, I think you know this, like we may have talked about this, but if you look at the crossover line — the point where intangibles crossed over tangibles — it was right when the original big Fama-French paper came out in ‘92, ‘93, something like that. So right when they were talking about price-to-book is when price-to-book started to lose some of its relevance.
We run these numbers for the US public equity markets — these are our estimates. We estimate that investment SG&A excluding R&D — so this is the component of SG&A that would be considered an intangible investment — was $2.2 trillion last year, to give people some calibration. Our estimate is that CapEx was $1.7 trillion. Investment R&D — our argument is not all of R&D is investment, some of it’s maintenance, but that investment piece was $700 billion. And so, you know, you take $700 billion plus $2.2 trillion, that’s $2.9 trillion. And that gets you back to that ratio I was mentioning a few moments ago.
So these are very, very large numbers. And not to say it snuck up on anybody, because it really didn’t. But you, you have to take a step back and understand what these trends have looked like. And by the way, you can see it if you just look at the composition of companies by sector or industry in the S&P 500 — you’ve seen that mix shift, you know, toward technology, toward healthcare, away from materials, away from industrials, and so forth.
Now, the second point, Kai — and you’ve, you’ve already touched on this a couple of times — but I’m gonna just, I’ll amplify it. There’s a wonderful book which I’d recommend — I think both of us are fans of this — called Capitalism Without Capital by Haskel and Westlake. For folks that want to get initiated, this is a really nice book. And I’ll say they, it was a nice piece of marketing where they talk about intangible assets specifically and they call it the Four S’s, right? So they, they had to wiggle some stuff around to make it work. But it’s a nice way, it’s a nice way to remember this.
The key is that, look, none of the laws of economics have been repealed. There’s nothing magical here in any way, shape, or form. It’s just that intangible assets have different characteristics than tangible assets that you just need to be aware of. And as we’ll see, there are kind of pros and cons to intangibles. Just as we saw with that distribution, we’re seeing faster growth and faster decline than we saw in the past.
So the first is something you’ve mentioned a couple of times: scalability, right? Which is, it’s typical that an intangible asset has a high upfront cost, but once it’s established, replicating and distributing it tends to be relatively cheap, right? So writing software is just the classic example of that. Things like network effects also would fit into that bin as well. So you can grow much faster than what we’ve seen before because you don’t have those physical constraints as you used to have. Again, that’s the good news.
The bad news: there can be obsolescence, and that sort of leads to the second S — and that’s sunkeness, which is once you’ve invested in something, if it doesn’t work out, that asset tends to be not very valuable. By contrast, if you invest in a physical asset — you know, you start a restaurant and you buy a building, you put tables and cash registers and all that stuff — well, if it doesn’t work out, those assets still have value because they can just be sold to someone else who’s doing basically the same thing. So you could think about: recovery values might be better with tangible assets than they are with intangible assets. So sunkeness — and I almost put obsolescence in that bucket as well.
The third one is what they call spillovers. And the idea here is that it’s really difficult to protect intangible assets — it’s easy for people to take them. By the way, you can get into big issues about intellectual property, people stealing this stuff in different countries and so forth. But basically, the idea is best practices disseminate very quickly with intangibles.
So, you know, the example I often like to talk about is the iPhone. You know, it was launched by Apple in 2007. It was a very different form and function. By the way, Nokia had more than 50% share of the smartphone market at that time. And within really short — you know, they had patents and all this stuff — but within very short order, everybody basically had a phone that had the same basic function and features of the iPhone, right? And so that’s a classic example of a spillover.
The other example I always like to give is shooting in the NBA. There are these great charts on shooting. It turns out that, you know, since the late 1970s there’s been a three-point line, but it turns out that it was the last 15 or 20 years that front offices in the NBA recognized that three was actually more than two. And if you took the expected value of a three-point shot — even though the probability of making it was lower — it was a much higher expected value than taking a long-range two-point shot. And as a consequence, you’ve seen a complete migration in the spots from where the players take their shots. So now they’re around the rim and there are three-point shots. They really have gravitated to the higher expected-value places.
Interestingly, you can actually track the expected value of two and three-point shots, and at the peak, three-point shots were 14% more attractive than two-point shots because of mid-range shots. And that gap has almost completely gone away — it’s been arbitraged away by NBA teams, which is super cool. So put that in the spillover bucket as well.
And then the last is synergies. And I think this is a really exciting — it can be a very exciting concept. If you wanna be really bullish about the world, this is what you’d focus on. By the way, this is in part why Paul Romer won the Nobel Prize — I think 2017 — for his work on endogenous growth theory. So the idea is innovation basically is recombination of building blocks, and the more building blocks you have and the degree to which they’re digital means that you can actually innovate even faster than you did before.
And I think, you know, Kai, that’s one of the areas that, you know, I don’t know much about this area, but one of the areas that seems super exciting — for example, the application of AI in healthcare or medicine, right? So the question is, can we search the space and recombine building blocks in a way that’s vastly faster? We saw this when — you know, faster than what — protein folding, much faster than we did before. And I think that’s a really exciting area. So this idea of synergies — recombination of building blocks — is something that is defined by intangibles to some degree.
So, so yeah — you both have, you know, the fact that they’ve risen broadly speaking, and the fact that they have different characteristics. And that is really important for how you think about, you know, everything — all our metrics for valuation, our metrics for growth, our metrics for distributions of return on capital. All these things, to some degree, get affected by those basic observations.
Kai: Right. And you mentioned earlier, of course, the fact that the best-performing companies of the past few decades happened to also be the ones most leveraging intangible assets, whether it’s brand, network effects, human capital, or IP. And that can kind of explain, you know, why what seemed implausible as a trajectory for these firms was actually something they were able to achieve, due to their investments in this area, which at the time was well less studied than it is today.
One question I have now is that we’re starting to see a shift in terms of these Magnificent Seven stocks — away from this capital-light business model that led to so much success over this period, towards a more capital-intensive one. Their free cash flow has basically been close to eroded, given how much money they’re now pouring into the build-out of AI data centers via CapEx. How should investors be thinking about these companies? Obviously they still trade at reasonable premiums on the back of their historical success. But as they shift to more of a utility-like capital intensity, should that be an area of concern for investors?
Michael: Yeah. I mean, you framed it so well. I would just take one step back and say, look, what is an investment? An investment is an outlay today in the anticipation of future benefits, right? So the cash flows that I’m gonna generate over time, discounted to today’s value, are more than what I’m investing today, right? So it’s gonna be NPV positive — that is the level set for all of this, right? So whether it’s intangible or tangible, it doesn’t make any difference, right? Is this a good investment or not a good investment?
One thing I like to remind people of is that, you know, Walmart, which is one of the great companies of all time, had negative free cash flow for the first 15 years that it was a listed company. First 15 years. Now, it turns out it was profitable, right? It had net income, but it was investing more than it earned. So it had negative free cash flow. Do you think, you know, was buying Walmart when it first got listed — was that a good investment? The answer is a fabulous investment, right? Because the return on investment was really high, and when the return on investment’s really high, you wanna do as much of that as you possibly can while you can do it, right?
So that’s, I think, a really important illustration that even though Walmart was profitable because of the way the accounting works — you still have to assess the return on investment. And Kai, I think you put your finger on it — you said it just perfectly — which is that’s what people are worried about. If you have a view that all this spending is gonna deliver, you mentioned sort of utility-like returns, you think it’s gonna be something better than utility-like returns, then there’s an enormous opportunity in front of you. If you think it’s gonna be — because of all the spending and the commoditization of the goods or services — utility-like, and I’ll use the proxy of sort of cost-of-capital-type returns, then it’s gonna be value neutral, right? It may not hurt you, but it’s certainly not gonna be value creating.
I always thought this idea of asset-light was a bit of a — a bit of a myth — because if you’re just laser-focused on investment, you’re just realizing the investment was not on the balance sheet as it was historically — it’s now in the income statement. So that to me — you know, and this is what I always like to say, and I say this to my students quite, quite directly — like, at the end of the day, as a financial analyst, your job is to figure out how much money is the company investing and what’s the return on investment. And if you understand those two things, that’s the whole gig, right? That’s the whole gig. Because you’ll understand growth rates, you’ll understand economic — you know, profitability, you understand return on capital. And then the degree which you can figure out how long they can do this trick — you know, invested returns — that’s strategy. That’s the whole gig. That’s the whole gig, right? You know, multiples will follow that and so on and so forth.
So I think we, you know, we’re just wanting to not lose sight of what we’re ultimately here to do, which is figure out how much is being invested — doesn’t matter where it’s going on the income statement or balance sheet — and what the return’s going to be.
And I think to your point, I think there’s just enormous amounts of uncertainty about what those returns are likely to look like. You can paint a very positive picture. You can paint a very negative picture. And I think it remains to be seen.
Now the other thing I’ll just say — you know, from my work, this is the work I’ve done with Rappaport on expectations — you know, the other thing I find to be very useful is to sort of go backwards and say, if the stock price is at X, you know, what do I have to believe about the future states of the world to solve for that stock price? And then you’re doing an over-under. So it’s like, I think we should be more optimistic than that — which means you buy it. I think, no, we should be more pessimistic — which means you should sell it. So rather than pinpointing what the future’s gonna be, maybe an easier way to do it is to say, what do I have to believe about this stock? And then say, I think they’re gonna do better or worse than that.
Kai: Hence the base rates.
Michael: Hence the base rates.
Kai: Yeah. So you said something really interesting to me, which is around the idea that the term “asset light” is kind of a not-value-neutral term, right? Because effectively when someone says asset light, what they mean is we’re gonna count physical CapEx as an asset. But intangible investment — whether through marketing or R&D — is not considered an asset. And then you said something in addition to that where you said, look, all that matters as an investor is what kind of investments are they making and what’s the ROI on those investments?
Now you’ve done a lot of work on the issue of how accounting obscures intangible investments — how it perhaps elevates physical CapEx but doesn’t quite give proper treatment to intangible investment and thus the intangible assets formed on the back of this investment. Maybe walk me through this, because I think this is quite important, and I think this is what leads to the misconception that asset-light businesses don’t have assets. They do have assets. The assets just don’t show up in the balance sheet because of accounting quirks.
Michael: That’s exactly right. And by the way, you know, even before I get into this, I will say, and I think Kai, you and I have both lived through this — the devil’s in the details here. So there are a lot of judgments that go along with this. But let’s just say that there are sort of two or three steps.
Number one is you have to break down SG&A — so selling, general, and administrative costs — into some basically two components. One is a maintenance component, which is how much money does the company need to spend to sustain current revenues, or perhaps market share, however you wanna think of that. So maintenance component, and then the other component is an investment component, right? So that investment component we can think of — deem it to be a discretionary investment in pursuit of value-creating growth, right? So that segregation is the first big thing we need to do.
The second thing is, once we have that investment piece — as you point out, this asset that we’re building internally — then you need to determine or estimate an asset life for it, right? So just like if you put a machine in your factory on your balance sheet as a five-year life or a three-year life or seven-year life, you’re making some sort of a judgment. We have to do the same thing for an intangible investment. And then of course, all you’re doing is you’re putting that investment on the balance sheet. So what we’d wanna do is ultimately put that on the balance sheet just like we would something else and then amortize it, right? So you’d depreciate physical assets; you amortize intangible assets.
So maybe I could try to give a really simple example to make this a little bit clearer. Let’s say you can buy a machine that costs $500 with a five-year life. And let’s say it’s NPV positive — the cash flows are great, we wanna make this investment. If it’s a machine, what do we do? We put $500 on the balance sheet, and then for five years in a row, we depreciate $100. It’s a great investment, and that’s how it shows up. You see $100 on the income statement, and then your gross property, plant, and equipment goes to zero net.
Okay? Now say you’re gonna acquire a customer. And let’s pretend the customer’s gonna be around for five years — they’re gonna churn after five years. They have the same exact cash flows as the machine. Well, what are we doing now? The answer is we’re expensing $500 on the income statement, right? All the cost upfront and none of the benefits show up. In fact, we’re saying it’s an NPV thing — we should bring on as many of these customers as we possibly can, right? But the more that we do that, the more money we’re gonna lose, right? So on the one hand, one looks really attractive, and the other looks really unattractive from a pure income statement standpoint.
Okay? So when I say the devil’s in the details, the key issue is, how do we think about this maintenance versus investment component? And how do we think about asset lives?
Now this is a very, very active area of research in academia — strategy professors have taken it on, finance professors, accounting professors. The paper we’ve used the most is a paper called “A Better Estimate of Internally Generated Intangible Capital.” I love this title, by the way — a better estimate, no matter what you have, this is better. This is by Iqbal, Rajgopal, Srivastava, and Zhao and the paper came out in Management Science in the last year or so. They go through the Fama-French industries and give you their estimates of the intangible component of SG&A, the investment component of R&D, and then they give you asset-life estimates. So you can tailor this by Fama-French industry, which is really good.
Now the other thing I’ll just say is: if you make all these adjustments — and it’s a lot of work to do this — it turns out that free cash flow doesn’t change, right? Because free cash flow is just net operating profit after tax minus investment — that equals free cash flow. Well, what we’re doing when we go through this exercise is we’re actually increasing NOPAT, we’re increasing earnings, right? Because we’re taking away an expense and we are adding amortization. But typically, if it’s a growing company, that net will increase earnings. And we’re increasing investment by the same exact amount, right? So you’re increasing one, increasing the other — your free cash flow doesn’t change.
So the question is, I think, why are we going through all this effort, right? If the free cash flow doesn’t change — and I think that’s, to some degree, a fair argument. But I go back to the basic core, which is: as an investor, as a business person, right — if you’re trading, it doesn’t make any difference. But if you’re a business person trying to understand a business, I would pose the question: is it relevant to you to understand how much money the company is truly investing and how much money they’re truly earning? And if those things are important to you — and I would argue they should be — then I think this exercise is actually a worthy exercise to go through.
Now that said, you can use proxies. You can do things like, you know, free cash flow yield. That’s not gonna change as a consequence of this, but you’re gonna be one step removed. You’re gonna be a little bit more blind about understanding how you got to the free cash flow, the path of the free cash flow. And I think in this case, understanding that path is actually a pretty valuable exercise.
So that’s why we, we go through this whole thing. And I just, I just have to believe this is a step toward the truth — a step toward understanding economic reality. By the way, I think it leads to much more quality conversations with management teams, for instance. By the way, even management — I don’t think they understand. They don’t know these numbers. If you could walk in there and say, “Hey, your SG&A is a hundred dollars — how much is maintenance, how much is discretionary investment?” They don’t, they just don’t know. They don’t think about it that way. So in some ways, we’re actually doing something that’s distinct from what management themselves are doing. They’re using a lot of inertia — they’re just doing what they did last year, plus or minus a little bit.
So that’s another really interesting point to think about: this is not, this is not like in the day-to-day language. People don’t do this normally. I think the market sniffs it out, by the way. But I don’t think people do this normally.
Kai: So this distinction between maintenance spending as opposed to investment is actually quite important because — sure, free cash flow is what it is, it’s unchanged — but it doesn’t tell you the future growth prospects for a business, right? There could be two businesses with the same free cash flow yields that have very different investment profiles.
And I think you’ve done some interesting work on this actually, where you looked at tangible investments on the CapEx side against depreciation, and you kind of found that actually companies — the depreciation kind of understates how much is maintenance as opposed to growth. Because most of the companies that are very tangible-capital-heavy are older businesses, more mature businesses, for whom the obsolescence and the wear and tear on their physical machinery is actually faster than estimated by depreciation. And that leads to systematic issues with investing in these businesses and perhaps explains part of the reason why they’ve underperformed historically.
And then on the flip side, intangible-intensive businesses that have been doing the same thing — investors have actually not been giving them enough credit for how much they’ve been investing in future growth, whether it’s R&D developing a new drug or some new software. And that could potentially explain part of the reason why intangible-intensive companies have outperformed historically. Does that tie back into why free cash flow is helpful as a way of solving some limitations, but not enough?
Michael: 100%, Kai. Let me play this back to you just to make sure everybody’s on the same page, right? So a company has CapEx of a hundred dollars and their depreciation’s $50, right? Usually what we argue in finance is depreciation is a proxy for maintenance CapEx, right? So we need to spend that depreciation just to maintain our organization. So in that case you’d say $50 of the spending is maintenance, $50 is investment.
To your point, if the maintenance is actually higher than $50 — let’s say it’s $60 or $70 — that means there’s less money going to investment and more money to maintenance. And as a consequence, less money to investment. You know, all things being equal, same return-on-capital assumption means slower future earnings growth, right? So that’s why this is — you know, again, this may feel like accounting, you know, splitting accounting hairs — but this is actually really important because if you don’t have a good grasp on that, you may be overestimating future earnings growth.
Again, all things being equal, because you’ve miscalibrated the maintenance component. So yeah, I think that’s another really important thing to think through that most people don’t. And by the way, that depreciation is a proxy for maintenance CapEx. You know, there are two reasons we can miscalibrate this: one is technological obsolescence, and the other is inflation. You know, there’s always been technological obsolescence, but that comes in cycles. You know, inflation flaring up a little bit. And obviously throughout our thread, you know, AI and so forth, technological obsolescence risk — those things are much more tangible today than they were before. And so you really do need to think this through very carefully to understand what’s going on.
I think you’ve stated it really well, but just to be clear: you may, you may think more money’s going to investment than is. And again, with whatever return-on-capital assumptions you have, that means future earnings are less than what you think they’re gonna be. And that’s not gonna be good.
Kai: And just to kind of drive this home — so people don’t just take away the fact that this is an interesting accounting exercise. You mentioned Fama-French in ‘92, right? The idea that, as an investor, there’s a premium associated with buying low price-to-book stocks and avoiding or shorting high price-to-book stocks. It turns out hundreds of billions, if not trillions of dollars, are tied to indices and strategies on the basis of this idea: price-to-book, price-to-earnings, price-to-cash-flow.
Just to kind of put a bow on this discussion, which has been so fascinating — talk to me a bit more about how you think the rise of intangible assets and some of these adjustments you’ve discussed could potentially guide investors as they think about addressing some of the limitations of more traditional approaches, allowing us to keep our value discipline without simply excluding some of the most important modern businesses.
Michael: Yeah. So, Kai, I do wanna make a distinction. I think everybody will agree with this, but just to make sure that it’s really clear — the distinction between value investing and value factors. And these are somewhat different things. So value investing is buying something for less than what it’s worth, right? I think we all — that’s mom and apple pie, right? I don’t think anybody’s gonna disagree that that’s a good idea. The value factor is one of the ways we try to get to that, right? So they’re saying like, let’s buy statistically cheap things, let’s avoid statistically expensive things, and on average we’re gonna generate some premium to do that, right? And by the way, again, that’s very, very sensible.
In the Fama-French model, you know, one of the things they relied on was price-to-book — they use book-to-price, but basically price-to-book. The challenge is: if book value is less reliable as a measure of value than it used to be, then we could run into problems. And this is precisely the conversation we’re having now, right? Which is, if you’re expensing all your investments, you’re not adding onto invested capital, hence you’re not building up your book value. So as a consequence, the punchline is that book value is probably understated. In fact, I’m sure it’s understated for businesses broadly speaking. The adjustments we just talked through will increase your invested capital, hence increase your book value, hence lower your price-to-book. And that may reshuffle companies.
So there’s a really nice paper by Srivastava and Baruch Lev, and the paper’s called “Explaining the Recent Failure of Value Investing.” It’s in Critical Finance Review a couple of years ago. And they basically go through and make these adjustments. And what they find is it gives a huge boost to the value factor in terms of explaining performance.
And so, you know, these adjustments are not — you know, again, like you said, it’s not just accounting — this actually does improve the quality of the signal. And as you point out, the value factor is a very important signal used certainly in quantitative work. So I think this gets us, again, closer to what we were trying to do before, which is buy what’s cheap, sell what’s dear, earn some sort of a premium over time. So again, getting it to improve — boosting our signal basically by doing this.
Kai: Thank you. I could talk for four more hours, but since our time is winding to a close, I wanna just ask one closing question of you, Michael: what is the one thing you believe about investing that the majority of your peers would disagree with?
Michael: Yeah, I don’t know. I dunno if they would disagree, but there are a couple of topics that come to mind. The first one is this idea that dividends contribute to total shareholder returns. And you often see people like market historians saying, oh, dividends are, you know, a third or two thirds of the total returns over time.
Total shareholder return technically is a capital accumulation rate, which — a capital accumulation rate assumes 100% dividend reinvestment with no friction. So no taxes, no transaction costs, and so forth and so on. And if you accept that — you know, you have a hundred-dollar stock, it pays a $4 dividend, you now have $96 and $4 in dividend, and then you take your $4 and buy stock to get you back to $100 — it becomes very obvious. The price appreciation is the only thing that drives capital accumulation over time. So dividends actually play no role whatsoever in capital accumulation, capital accumulation. So I think that would be one.
The other one I’ll just say — which is related — is I think both dividends and buybacks are just wildly misunderstood topics. I don’t understand why they are so flummoxing for people, but they seem to be. You often hear things like buybacks in quotes “create value” or “destroy value” for the company, which is just mathematically nonsensical. What buybacks do is cause wealth transfers. So there can be a wealth transfer from the sellers to the buyers, or the buyers to the sellers, based on whether stocks are over or undervalued. But there’s no wealth creation, right? So again, the company’s worth a hundred, they buy back $4 worth of stock. Now the value of the company is $96. Right? It’s just $100 minus $4. There’s no wealth creation or destruction. They could pay a dividend, they could buy back stock, they could burn the cash in the parking lot. Doesn’t make any difference. Right.
So those would be two: the dividend thing on the role of dividends in terms of capital accumulation. And by the way, for most people, capital accumulation is what they’re after. And as a consequence, they misunderstand the role of dividends in achieving that.
Kai: Great. Well thanks, Michael. I really appreciate you taking the time to chat with me.
Michael: My pleasure, Kai. And again, I look forward to your future work. It’s always the best stuff out there.