Full Transcript: Azeem Azhar on Where AI Value Will Accrue

Reconstructing AI Demand, Stack Economics, and the Moats That Survive

Kai: Over the past few episodes, a central topic of discussion has been the multi-trillion dollar AI investment boom. Massive capital spending has helped propel markets higher while leaving stock indices increasingly concentrated in companies tied to the AI build-out. Some of the biggest questions facing investors today concern the sustainability of AI demand, the pace of technological progress, and where, importantly, economic value will ultimately accrue across the AI stack.

To help answer these questions, I’m glad to introduce my friend, Azeem Azhar, a technologist, investor, author, and entrepreneur who has been thinking deeply about the impact of frontier technologies for many years. Azeem is the author of The Exponential Age: How Accelerating Technology is Transforming Business, Politics, and Society, and the founder of the excellent Exponential View newsletter.

Azeem, thanks for coming, and welcome to The Intangible Economy.

Azeem: I’m so excited to have this conversation with you, Kai. I’ve learnt so much from your essays over the last couple of years, so looking forward to it.

Kai: Me too. All right, so let’s start with your report on the state of the AI economy, which you just released a couple of weeks ago. Congratulations, by the way. It was excellent.

Azeem: Thank you.

Kai: So the idea is this, that the supply side of the AI boom is relatively visible. We can observe chip sales, capital spending, and data center construction, but really the demand side is much harder to see. You attempted to reconstruct that demand from the bottoms up. What did you find?

Azeem: Well, we found that there was a real economy underneath all of this, underneath all of those orders that are improving Jensen’s collection of leather jackets, is a company’s spending real money. And our model, our estimates came in at, for the year to June 2026, a hundred and ten billion dollars for the previous twelve months, excluding China, being spent on generative AI.

And that includes, you know, Main Street firms buying from OpenAI or Anthropic or from Cursor or from Eleven Labs, through to consumers spending on their consumer products, whether it’s, you know, the big ones that we’re familiar with or some of the more niche products. And it also includes where, you know, OpenAI and Anthropic and other labs are spending on training capacity through the cloud companies, the hyperscalers and the neoclouds.

And that hundred and ten billion dollars, just to put it in some context, was the trailing twelve months. But if we annualize the June number, we got to a hundred and seventy-five billion annualized. And what that implies, of course, is that the growth rate has stayed high and, roughly speaking, the growth rate is about as high in the summer of 2026 as it was in December 2025.

That surprised us, by the way, kind of when we went into the work, we thought the growth rate was gonna slow down. And I think the other thing, just to put it in some context for your listeners, is it’s about three times faster than the previous sort of internet-oriented waves, the internet itself, mobile, apps and ads, and cloud.

So it feels real. You can slice off a little bit here and there if you want, but ultimately there’s this number. It’s north of a hundred billion dollars.

Kai: And so, you know, obviously there’s different components of AI demand. How did you go about, like, deduplicating and avoiding double accounting? So a dollar of, say, end user demand that goes to an application, how does that kind of filter through all the way to the bottom?

Azeem: Yeah. So we used a number of different approaches. In fact, I think the deck is sixty-five pages long. There are five pages of methodology slides. It felt a little bit like a physics PhD, in that sense.

So we approached this in a number of different ways. One is that we do kind of bottom-up assessments of the thousand or so companies that claim to be in the generative AI space. And some of those are more well-sourced than others. So, you know, when you see confirmed press leads, you can do something with that.

And that then gives you a number. And then we have to figure out, right, how much of that has been going down the value chain to other people. And this is a key point. ‘Cause if you spend ten dollars with Anthropic, they spend a bunch of that with Google and Amazon to host those models.

We don’t wanna count that twice. And, you know, to do that, we reconstructed and built financial models for the key players, whether it’s the big FM, foundation model labs, but also the AI hyperscale businesses within Microsoft, Google and Amazon. You know, these are not broken out consistently at all by the hyperscalers.

So that allows us to figure out roughly speaking, you know, what’s being passed through to a supplier. Now, that gives us one number, and that’s roughly that hundred and ten billion dollars that I talked to you about. We also have to do some triangulation. So there are always disclosures that can help you.

So how many chips has a company bought? How many gigawatts does it have? What do we know about the efficiency of the models? And, you know, under particular operating assumptions, how well will they work? And we use that as a triangulation mechanism to say, “Are we really in the, you know, in the right ballpark?”

And then, of course, we’ve been doing this now for a year, so we’re able to test whether our model is an overs or an unders model when new leaks come out. And I think one of the things that I would say, despite that number being absurdly high, a hundred and ten billion dollars, is that we were lower.

If you look at where we thought we would be in September 2025, or we thought we’d be in June 2026, we were more conservative than the number we ended up with. And, you know, particularly because Anthropic did very, very well. So of course, I’m gonna be very honest, this is not auditors going through the work and looking at regulatory filings.

Where there are regulatory filings, where there are audits we can look at, we’ve absolutely done that. But we’ve endeavored to source and give a confidence rating for every number that goes into the model and to describe why we made the choices that we did.

Kai: That’s really robust. Thank you. So let’s move into the next question, which is where in the AI stack we think the value will be captured. But before we do that, and before we get into the economics of each layer, maybe you can just briefly walk us through each of the four layers of your taxonomy. The chips, hosting, foundation models, or FM as you called it, and then applications. And then maybe talk a little bit about how, just really quickly, about how spending has been shifting amongst them over the past few years.

Azeem: Yeah, absolutely. So, you know, at the bottom of all of this are the chips, predominantly NVIDIA’s, but also now, you know, kind of Google’s own and AMD.

And the chips are a capital expense, so that $110 billion, of course, doesn’t include anything that’s been spent with AMD or NVIDIA. You then above that you have the hosting layer, which is the, you know, infrastructure operations. Amazon Web Services, Azure, but also these companies that we call neoclouds, so Nebius and CoreWeave and others.

And so these are people who are in the business of acquiring the physical asset of the chips, acquiring the power and the cooling, operating data centers, and effectively renting their capacity to companies who need it, which will be the foundation model labs, the most common of which is OpenAI and Anthropic, but it also includes companies like, you know, Eleven Labs again, which does some voice transcription.

And so above the labs is the application layer, and a company like OpenAI operates both at the application layer, which is what you and I are using with ChatGPT, but also the more direct layer where somebody is hitting the API for GPT 5.5 or 5.6, more directly.

And we separate that out because there’s a lot of activity, increasing activity at the app layer, which then is using OpenAI or Anthropic or one of the Google models to serve that. And at the moment, as you would expect, the value is at this stage accruing at the infrastructure level, and it is accruing down at the hosting companies, because ultimately if you need to use any of these apps, something has to power the chip.

But what started to happen over the last few quarters has been the amount of value being captured by the foundational apps, specifically Anthropic and OpenAI, has started to increase proportionately. So if we go back to the first quarter of 2025, we reckoned about 9% of value was accruing to those labs. By first quarter of 2026, we thought that was 11%, so that’s a 20% uplift. But of course, revenues had tripled in that time.

And you can start to see that now when you see the gross margins that Anthropic and OpenAI are able to achieve, and they are able to secure sort of a decent capture, a decent amount of that value. And at the same time, the amount of value that’s going up in the apps layer is also increasing.

And again, what I would say is that that is a classic maturation of a new technology. You know, the value starts down with the kind of hardware, and then it moves into the apps. But what I would also say is we don’t know if this is where it ends, right? Because the AI market could play out very differently to the internet or to the mobile market.

Kai: Yeah, no, I wanna echo what you just said there in terms of the classic maturation. You know, I’ve done a lot of research on historical cycles as well, and that’s one of the things you do see, that necessarily in the beginning of a technological revolution there’s a build-out phase and there’s a deployment phase. So maybe that’s a clue, maybe not. But as to kinda how, you know, what inning, I guess, of the build-out we may or may not be in.

Azeem: I think what is also complicated here is the degree of vertical integration, vertical competition that we’re starting to see. So, you know, this starts with Google building their own chips. But now everyone’s building their own chips, right?

Amazon is, OpenAI is, Anthropic is as well. And there’s more cross-stack competition. So Microsoft runs Azure, it’s hosting OpenAI, but it also offers rival models, and it offers its own models. And OpenAI and Anthropic are trying to secure more compute capacity that is more owned and operated.

So there’s gonna be this really interesting, mixed economy where, you know, companies will be both contracting out to specialists, you know, the CoreWeaves and the Nebiuses, to do that, and contracting out to companies they’re competing with and competing with those companies as well.

And, you know, you are such an expert historian on this question. It’s quite complex, right? And I’m not quite sure where it will end up. I mean, I think in truth, the bulk will be this sort of segmented set of layers. But there will be enough reason for particularly the biggest foundational apps to try to get at least part of their businesses vertically integrated.

Kai: So I wanna get back to that point of vertical integration in a second. But just to kind of step back, for the listeners, you know, we have these four layers, right? And you have a great slide where you say that pricing power is not about stack position per se, it’s about competitive pressure, right?

Azeem: Right.

Kai: So how much competition does there exist? You mentioned NVIDIA kind of stands more or less alone at the bottom, but then there’s different layers of competition. So maybe you could just, you know, talk high level about, for each of the four layers, which do you think has the most competition, at least today? And then as we move forward, which faces the most potential competitive pressure?

Azeem: Well, I think the hosting companies are the ones who have their hands around the real choke point because you ultimately need to host. The amount of capacity that’s available is pretty low.

It’s as much about whether data centers can be energized. And whether or not open source or closed source models, you know, get higher or lower margin market share, you’re still gonna need to run things off those chips. The hosting companies obviously have to acquire chips, and I would say that the chip business, whether it is the GPU itself, like what we think about as NVIDIA, but also the memory, is a business that has tremendous, tremendous demand attached to it. So I mean, I think we’ve already seen this play out quite a lot in what’s happened with the prices of those companies.

And the interesting thing that I noticed, and we haven’t released this yet, but when we try to model out AI demand five or ten years, the biggest, tightest bottleneck ends up being around memory, not compute. So when you buy an NVIDIA GPU, it’s got the logic chip, and it’s also got this high bandwidth memory attached.

And that is a market which is dominated by three companies who are sold out a year in advance. So if you think about the resilience of a bottleneck on the actual hardware, there’s something to be said for, you know, memory and GPU, but there is a different risk there, which is that if you end up at the top of a cycle, we know what happens to memory stocks, right?

They give up all the gains of the previous, you know, five or 10 years. So there is this sort of like weird risk that is not as persistent in my simplistic thinking as I bought the chips, I’m depreciating them over six years, I can rent them, I can rent them to OpenAI, or if open source wins, I can rent them directly.

So I think there is something in that particular market. And, you know, you have to remember just how strong hyperscale businesses were up until five years ago. You know, the 50, 60% gross margins on these businesses, and of course, they’re giving that up at the moment to establish themselves in the AI business.

So what about apps and foundation models? Why do I think there’s more complexity there?

Kai: Before we go there, can I just make a comment on the memory side of things? So you had a really good slide, and I wanna just make sure I call this out for our listeners, where you showed like the cost of a data center in 2021 versus today. And so memory, if I’m quoting you correctly, was 2% of the cost in 2021. It’s now 18%, right? So this is obviously a bottleneck. The SK Hynix, they just joined, you know, the US markets because investors really want a piece of the action here. It has become a legitimate bottleneck.

Of course, you know, high margins and profits attract competition, so we’ll see how that plays out. And then it’s interesting too on the hosting side, you mentioned, you know, the kind of oligopoly that was AWS, Azure, and GCP, you know, pretty good traditional cloud business. But now we have, you know, Oracle coming into play, we have the neoclouds, Meta, I guess, is in the game now, and, you know, Groq and SpaceX.

So I mean, it is interesting that you’re having kind of a... We’re seeing the next phase of the competition where eventually you saw consolidation around the three big players in the last cycle. I think it’ll be interesting to see. Right now there clearly is more competition on the hosting side, whether that settles down and there’s kind of a shakeout, and we end up with a similar market structure moving forward.

Azeem: Okay, let’s talk about that. I think that’s a really interesting question. Like, why would we have a shakeout? I don’t think that there is a network or ecosystem advantage to being on any given hyperscaler. If you look at the historical trajectory, AWS came out first. When Azure launched, the thing that they did was that they bound themselves quite tightly to the Microsoft enterprise ecosystem, which gives you a little bit of a lock-in.

But at some point, you just wanna win customers, and so that lock-in starts to sort of teeter and disappear. I’m not sure there is going to be a clear lock-in on any given hyperscaler or neocloud. And they’ll end up with different categories of expertise because, you know, we think of like, oh, yeah, AI data centers.

But you’re doing very different things if you need kind of guaranteed low latency co-located training compared to inferencing, which can be much more distributed. You can move your workloads around, and I think that that somewhat favors a more heterogeneous and fragmented set of players in the hosting layer. And the other thing to note is that demand is so tight that these companies can make quite a lot of money.

I mean, I think the SpaceX S-1 was quite telling there because in the S-1 they sort of declared they’ve got these two contracts with Anthropic and with Google to serve AI, and they’re pretty big. They’re like a billion plus a month. And if you look at how much capital they have deployed thus far to support these contracts, it’s something like twenty-six billion dollars, and I think the Anthropic contract alone would pay that back in twenty-one months.

And then you’ve got the Google contract as well, and you’ve got these assets that you can depreciate reasonably over six years. So there’s quite a lot of money that is around, and you can make quite a bit of money from that. And because there isn’t a natural reason for there to be a network effect on a hosting stack, and the market is kind of heterogeneous enough in terms of use, I can imagine it being a little bit more fragmented than we’re used to seeing in a, you know, in an IT business.

Kai: Hmm. Interesting. Okay, so let’s move on to the next... I cut you off, so I didn’t mean to do that, on to the foundation model and application layers. You know, obviously we’re seeing a ton of new open source. I think Thinking Machines came out last night with a new open weights model.

You know, how serious is the open source competition? And we also saw, I think it was the Meta model and OpenAI now trying to compete more on price as opposed to being at the frontier. And, you know, a lot of CTOs and such are kind of pushing back against the rising costs, right? Of doing AI in their bottom line. So talk to me a little bit about competition, which it seems to be accelerating significantly into the frontier model layer.

Azeem: I mean, there’s loads of competition, and there are things that we do today without going to a frontier model company, and we do that within my own team. We use the GP models, we use the Kimi models as well. Here’s the way that I think about this. The database market was kind of similar to this, right? So you had Informix and Ingres and Illustra and Oracle, and now there’s just Oracle. And then you have the open source databases, so Postgres and MySQL, and Oracle’s actually bought MySQL now.

But that market splits about 50/50 by usage between proprietary and open source. You know, why is that? I think it’s less to do with technical capability, and it’s more to do with assurance and verifiability and ecosystem support and all of those types of things. And so I do think that given how complex AI is to run and how slow decision cycles are, that there’s still lots of room to be a proprietary, more expensive model because of the other things that you’re providing.

When we hear companies switch to, you know, 90% open source models, it’s businesses like Shopify and Airbnb and Coinbase. This is not mainstream, right? This is not CVS or UnitedHealth or Hertz. Like, these are the most sophisticated technological big companies in the world who were well ahead of the curve.

And I think the other thing to think about is that the proprietary companies have the choice to adjust their strategies if they so choose. I mean, they may not do, right? They may be messianic the way Steve Jobs was messianic about the one-button mouse. But on the other hand, they may say, “Listen, we don’t wanna just leave all the value on the table, so we’re going to change our pricing.

We’re gonna accommodate data requests a little bit more.” I mean, they might, they might not. So I agree that what open source does is it provides both pricing but also business model pressure on the proprietary model companies. But that’s what it is. It’s pressure. It’s not the door closing on their businesses.

And the market has got so far to go. There’s a fantastic company called Ramp, which is a payments platform, and they provide data on how American companies are spending on AI. And I believe—gonna get the sum, but we have it in the report—the median American company that goes through Ramp, which is 70,000 businesses, but they are skewed to being tech forward, we’re spending $11 per employee per month on AI, which is a very, very low number.

I mean, we’re really in the foothills. So market share arguments, I think we’re not at the Coke versus Pepsi versus Dr Pepper stage of a market that is not gonna grow anymore. The market’s gonna grow a lot. And so I don’t think this really changes the trajectory of growth for the proprietary model companies, you know, in the next couple of years at least.

Kai: So you mentioned the move towards vertical integration by the model labs. I mean, do you think... Could one argue that that might be a defensive strategy with regards to if they’re seeing pricing pressure at the model layer and they’re saying, “Well, why not build a really cool harness and then start kind of going verticalized,” you know, Anthropic for finance, Anthropic for legal, for health, for design, right?

And then there’s the question around the pushback, right? You know, Alex Karp was on CNBC talking about how a lot of the customers of labs should be concerned about the risk of IP exfiltration and, you know, them taking your data and then building a business against you. So like, you know, maybe there’s kind of a natural equilibrium.

You mentioned the database space kind of having some, you know, there’s room for both open source and non-open source. You know, each customer kind of has their own priorities and trade-offs around these things. But, you know, I’m curious there too because it does feel like that’s been a very cohesive push by the labs to kind of try to attack, you know, well, you mentioned hosting and self, but really the application layer.

And that might go hand-in-hand with what you’re saying about a lot of value is moving up the stack, which, you know, you want to own the customer, that this is kind of a standard thing you see when these technologies roll out.

Azeem: Yeah. I think it is. It’s partly that, I mean, there are some precedents. I mean, simply on compute it is about three things. It’s taking a slice of the economics that you’d otherwise give to somebody else. The second is it’s about being able to create your own capacity or the thing that actually drives your business. And the third, I think, will be about silicon model integration being more tight, right?

Being able to juice out efficiencies that you might not otherwise. But the verticalization upwards I think is super interesting. What Anthropic has done had really, like, impressed me. They went out with a coding tool. Coding works very well because the customers are smart, they’re willing to pay, and you can test whether the coding is working or not.

And then I do think it makes sense to verticalize because as you go through the adoption curve, early majority customers like to see vertical solutions. So it’s much easier to go to a biotech company and say, “Hey, we’ve got this AI that you recognize generically, but here is the harness for biotech,” than it is to just say, “Here’s a bunch of machine intelligence, do something with it.”

Right? So I think that makes sense. So it is partly defensive, but it’s also partly about moving, especially with finance and legal, to where willingness to pay is the highest.

Kai: Yeah.

Azeem: I think Alex Karp is making a very, very good point because there’s a real tension between the model supplier just supplying you a model and being your competitor, and I think the place to look is Amazon Marketplace, right?

So we know that when Amazon ran the marketplace, there’s been a lot of reporting and court cases about this, they would start to compete with their sellers based on data and activity they could see on the marketplace. And, you know, sellers did stay because the distribution was still worth it. Now this is not exactly identical because if you’re training the AI as you go through that, I think that there’s a real risk that the labs don’t just skim a bit of your margin, but they take your space.

But then let’s look at it from the perspective of the company, the customer. I think what great companies have is tangible and intangible understanding, explicit and implicit understanding of what it is to be successful in that vertical. And I would be very, very hesitant to just assume that the AI provider I’m working with is going to always work on my behalf.

And I think that the really smart companies would want to be logging the traces of information they send out so that they have a trace library that if they choose to move vendor, they can sort of bootstrap that choice. And I think the other thing they may want to do is really insist on how they can get not just their data back, but the learning that has emerged from that data back and how do they audit that that has already happened.

And I’m sure conversations like that are happening, and they’re not gonna be straightforward, but they’re the sort of thing that applies pressure on anyone running a proprietary product at this point.

Kai: Yeah, and I wanna talk at length about the diffusion and adoption question. You know, how does the technology diffuse through financial, biotech, all these other industries? Before I get to that, though, I wanna ask one more question about your report. So quick detour. We talked about this, but one of the slides that you put together that I love and I’ve cited so many times, it shows the 1860s US railroad build-out, and it shows how basically these railroad companies, they obviously created a ton of value for society, but they themselves failed to capture much of those profits.

In the final slide of your report, you outline several ways in which AI’s value could similarly accrue mainly to the users as opposed to the providers of AI infrastructure. Can you maybe walk us through those scenarios and explain why in this state of the world, the profits might actually accrue outside of AI stack as consumer surplus?

Azeem: Sure. So the way that would play out would be that general purpose AI models are able to effectively replace the app layer upwards. So you don’t need to go to a Salesforce, you don’t need to go to a specific, you know, drawing app powered by AI. You can just go to your general purpose chatbot.

At the same time, open weight models, so the open source ones that anyone can host, get good enough or more than good enough for most people to want to use. And so now we don’t need the apps, and we can just go to a general open weight model. So both of those things have sort of squeezed out margin that either app vendors or proprietary models could have enjoyed.

And then if models become progressively more efficient, they might need less and less compute, which will mean that the kind of compute checkpoints and the sort of rents that the neoclouds and hyperscalers can charge, especially at peak times, starts to diminish. But so at the end of that, everyone’s margin on the supply side has really got squeezed down.

But we as consumers are getting exactly the benefit that we had wanted and would’ve got otherwise, which is we’ve got agents running and doing our financial analysis or, you know, we’ve got coding happening. And you can imagine, like, one way to think about that is for many of us, we just go to a no-name, super cheap domain host who then offers us for $3 a month, “Yeah, we’ll host your website as well.”

And we don’t think, “Oh, let’s go to Google Cloud and get, like, the bulletproof version,” because essentially there’s been enough competition, and most of those technologies are open source and non-proprietary, that the price for that has come down hugely, hugely, hugely.

And then I look at my Squarespace bill, and it’s 25 bucks a month. I’m thinking, “Why on earth am I paying 25 bucks a month for this? I should be paying $2 a month.” So I think that that is a way that, you know, consumers could capture a lot of the surplus in that kind of world.

Kai: Interesting. Okay, so let’s talk more about that part of the equation. Let’s zoom out of the AI economy and focus instead on how AI is impacting the kind of real economy, so to speak. The buy side. You know, obviously you showed in the report that demand’s super robust, but if you’re a skeptic you might argue that, well, the demand’s being driven by these unproven pilots, and this now the token maxing trend.

You know, ultimately the sustainability of AI demand will depend on whether enterprises can truly convert the spending into economic value. You know, one of the puzzles, for economists especially, is why this rapid AI adoption, which you showed, has not yet translated into the firm level productivity.

You wrote a piece called “Why AI Isn’t Showing Up On Your Bottom Line,” in which you cite some of my favorite work by Paul David and Erik Brynjolfsson. You compare AI adoption with the electrification of factories, which proceeds in these three phases: the light bulb, group drive, unit drive. Walk me through these phases and talk to me about where most enterprises are today.

Azeem: Yeah. Well, you know, the Paul David and the Brynjolfsson work is absolutely fantastic, and actually we’ve got something coming out in a week or two after we record to take this a bit further. But, you know, what happened with electricity was that the very first factory owners understood its power, and they hung light bulbs in their factories.

And you juice out an extra hour of a day from somebody. Factories were powered by these central steam engines, which then used belts and pulleys to distribute the power to the extent they could to elements of the workshop. And the next stage in electrification was to replace the central steam engine with an electric motor.

But you still got the configuration limitations of how much you can transfer power through a belt and a pulley. And the final moment is the moving assembly line, and it’s kind of Henry Ford, which is the unit drive idea where you’re essentially redesigning the entire process with the assumption that there is electricity.

So you can start to do things like move a manufacture through stages with highly specialized workers along that production line, which is what he did. You cannot go from light bulbs to the moving assembly line by adding more light bulbs. And so in the same way with AI, you can’t go from copilots to an AI native firm by just adding more copilot licenses, right?

It’s just no matter how many you add, you won’t get there. And it’s really complicated, as we saw with electricity. Most companies are at stage one or stage two, and the advanced ones are at that stage two because I do think the entirely AI-enabled enterprise—the first ones we’ll see will be sort of de novo companies, maybe Anthropic is the first one that is really, really like that.

And why is that? It’s because in order to take advantage of AI’s productivity, you need to let AI work in a loop where it can move, work much, much more quickly without the kind of human time and dollar cost in between. So every time we step in with a bit of verification or a decision that a human needs to make, we effectively slow down the whole process, and you can only make marginal benefits.

I think about this in terms of we love going on family holidays with my wife’s parents, who are absolutely kind of vigorous and exploratory, but they’re in their 80s. Doesn’t matter how neat my sneakers are, I’m still gonna move as fast as the slowest person in the group. And so for companies who put AI into a particular process, at some point there’s gonna be a non-AI process they’re gonna hit, and now any speed benefit you’ve got, you’re gonna lose, so you’re back to just getting a cost benefit.

So I think that’s the way that we look at this. But there is an additional complication, which is that Erik talks about the J curve, the productivity J curve, and, you know, versus work that’s done with Jones and others.

And the point being that with any investment, you are in the red before you turn that investment and you go into the black. I don’t think that you can tell the difference between a company that’s doing really well with AI and one that’s doing badly with AI just from how much they’re losing with AI.

Because an unsuccessful company is just in the red. A successful company has got lots of pilots that are about to turn the corner and go into the black, but right now they’re all loss-making, and so they also look like they’re in the red. And so if Erik’s J curve theory is correct, we should expect to see lots of companies saying, “We’re not seeing results at scale yet.”

But at the same time, you should see them continuing to invest because they’re seeing indications that they might get results.

Kai: Right. Yeah, I mean, I think this gets back to Erik’s idea of the complementary intangible investments required to actualize and unlock the value in a general purpose technology such as AI. Or just having access to ChatGPT, that’s great, but... And maybe it only enhances my individual productivity, but due to the congestion piece you talk about, that’s not really gonna change my organization until I completely rewire, you know, the layout of the factory floor and such.

Azeem: But you work with this data. Can I ask, I wanna ask you this question because you look at this data yourself, right?

Kai: Yeah, sure.

Azeem: So we should be seeing an increase in intangible investments and maybe even a delta since ChatGPT came out. But I’m not clear whether we measure it through the BEA frequently enough, with enough detail, or we’re measuring the right things to actually see that. So how should I understand that given that’s your expertise?

Kai: So one thing I’ve been looking at a little bit is job postings. And you can look at employee profiles as well to say, you know, is it the case that firms are rotating their workforce away from jobs that are substituted by AI and towards those that are complementary or augmented by AI?

And you do see evidence, by the way, of that. I mean, it’s still pretty early, so I haven’t been able to empirically link it to any uplift in terms of fundamentals. I was actually literally looking at this last week. But yeah, I think that’s one of the mechanisms you might see, right?

So you’re asking the question of what channels would complementary intangible investment occur through, right? I mean, also like layoffs and reorganizations, potentially hiring outside consultants. You know, there’s a bunch of different channels, but you kind of step back and ask the question, which is, again, having access to OpenAI’s API is not a competitive moat, right?

The moat is being able to harness it in a powerful way, you know, incorporating all your proprietary data and whatever intangible moats your business itself might have. But yeah, I think the macro data is still pretty new and, you know, again, to the J curve idea, I wouldn’t expect it to be showing up in the data.

So I don’t think it’s like a disconfirmation of this idea that AI investment is going to improve the economy. The fact that we haven’t seen it yet, you know, if we don’t see it in 10 years, then, you know, maybe there’d be some questions. But you know, I also do think at the bottoms up level that the CEOs, like there is some irrationality around it where they kind of feel like the pressure from the board to engage in AI theater and such, but there’s also the genuine view that, look, if this isn’t an investment that I think will ultimately improve my business, then like why am I doing this, right?

People are pretty capitalist, pretty rational. And so I think you have to just trust that the companies aren’t completely, you know, in AI psychosis or whatever.

Azeem: Yeah. And my sense in talking to CEOs, both in Europe and the US over the last, you know, year—I’m quite lucky I get to speak to between sort of 60 or 100—is that they are talking about it with a greater deal of acuity, right?

They’re closer to it. They know what good looks like. They know where they’re making mistakes, what bad looks like. They’re not more skeptical. They’re just more capable about all of this. And they also say similar types of things, which is that it’s easier to get to the tens of millions of dollars.

It’s proving harder to get to the hundreds of millions of dollars and beyond. But not many I spoke to, if any, have said, “We’re gonna slow this down, and we’re gonna stop trying to get this to work.” In many cases, they’re being more and more ambitious, even if they’re a little bit skeptical about now just giving everyone a Copilot license.

And I think that shows, again, a maturation. That this all comes back, though, in a way to timing because, you know, you’ve said, well, we might see this in the data in 10 years’ time. Well, that may be too late for a market that is, you know, behaving the way that it is.

And, you know, I think that’s true for a lot of the data that we would make sense of this. We’re only gonna see that data when it’s too late to make the decisions that we would have made had we had that data earlier. And I think that’s one of the things that we’re trying to do is just trying to provide a bit more light for people until, like, robust statistics emerge.

Kai: Yeah, I mean, look, you mentioned you work inside the organization with all these CEOs, which is really an interesting privileged view. But if you are a public markets investor, right, like myself, and you’re just looking at publicly available data, alternative data, aside from, you know, I mentioned some job, labor market, potential indicators, what other things should I be looking at, or what should investors be looking at, you know, as we kind of look at the cross-section of companies, both across industries and within industries against their own peer groups, to say which of these companies are potentially on the path towards electrification?

That they’re kind of the early adopters of the technology that might, you know, give them a competitive period of advantage against their more laggard peers.

Azeem: Yeah. That’s a really great question. So one of the things that we’re trying to do is, the first signal is what does that demand curve look like? And, you know, if the demand curve is not growing, that gives you a very, very easy answer to the question.

The next thing that we do is we look at S&P 500 earnings transcripts, and we essentially extract where people are talking about AI and where they are talking about it with quantified claims. So if you’re quantifying a claim, then at some point maybe that’s a little bit less theatrical. We haven’t yet actually gone to the level below, which is to scale that claim. But we do try to categorize a claim by, you know, whether it is product development revenue, simple cost savings.

And what you want to see is you want to see a mix of those claims. You want to see more and more lined up towards, you know, revenue, new markets type of activity. So that’s another piece that we look at. We also look quite a lot at this... I’ve talked to you about Ramp data, which is sort of, you know, old data. One of the useful things about the Ramp data is that they’re pretty good at categorizing relative levels of spend, so you can see where the median to the top decile spend is.

And you want the median to rise. You also wanna see what’s happening within the top decile spend. And if that diverges too much or if the median doesn’t rise, I think that’s telling you something, not necessarily about how useful this is for Main Street, but how difficult they’re finding it at this time.

And then I guess the last thing is, ultimately the quality of conversation that comes out of the CEO, CFO when they’re talking needs to start to talk with that sensibility. And the sensibility is about how is the data that we have and the know-how we have, the implicit knowledge, being turned into something that is useful.

And I think, you know, a few people have talked about this. I think obviously Karp has talked about it and Brian, whose surname I’ve forgotten, at Coinbase has talked about it as well.

Kai: Brian Armstrong.

Azeem: Yeah, yeah, that’s right. And that qualitative signal for me is also an important one, where the CEO is actually saying, “This is what the business is all about,” as opposed to saying, “Oh, my chief data officer is gonna tell you about our data lake.”

Kai: We just talked a lot about organizational redesign. Obviously many companies recognize that they may not internally have the capabilities to do this, and hence the hot trend today of the so-called forward deployed engineer, right? Everyone from frontier labs to hyperscalers to the traditional management and IT consulting firms are investing heavily in building out teams of forward deployed engineers to help their enterprise customers design and implement AI workflows. So how fruitful do you think these efforts will actually be?

Azeem: Well, I think it’s a sign of how difficult it’s proving, right? To get really meaningful results.

If you look at what’s going on, you’ve got Microsoft, you’ve got the foundation labs, you have these deals with some private equity. I think, hasn’t TPG done one with OpenAI as well? And if you go to some of the strategy consultancies of the sort of big three, they’ve been pretty honest about how much of their business is now, you know, AI transformation business.

It’s a lock-in mechanism, as much as anything else, I think from the perspective of the deployment companies or from the model companies. That is about lock-in, and about being able to get closer to embedded services, and it creates the optionality of whether you can get, you know, value-based pricing.

Because the real blocker is all about that reconceptualization. And that’s where Alex Karp’s comments he’s made, it is really part of that tension. I also... You know, the question you’d have to ask though is, how many forward deployed engineers would you actually need to shift an enormous American company with 100,000 employees, 50,000 employees, and 100 years of history?

I mean, it feels like those are quite difficult questions to answer, and I’m not sure any vendor-based relationship could ever answer that question.

So there’s even a question that might be kicking around, which is, is it more interesting to look at the mid-tier and smaller tier companies that can be rolled up in private equity and, you know, run in a way that is not dissimilar to how private equity has run these businesses. Which is they’ve come in and they’ve done operational improvements, they’ve brought in experienced hands to change processes and make these companies more effective.

So for me, FDEs are both a sign of a market maturing. They’re a sign that the technology is difficult. The fact that it’s difficult probably means it’s important. I mean, electricity was harder than staplers, right?

Kai: Mm-hmm.

Azeem: And which created more value. But it doesn’t necessarily mean that you’ll, you know, snap your fingers and all these American companies are gonna feel like they’re AI native in five or 10.

Kai: I’m just so curious too, like all these stocks with the Accentures and such that have been taking this huge beating, if that is, you know, overblown concern or not. I mean, I don’t know. It’s an interesting thing.

Azeem: Well, I think the market segments in very different ways. So if you go look at average revenue per employee, the lower revenue per employee companies who are in systems integration or effectively IT services, that feels like it’s much, much more competable with respect to AI, right?

So you could just get Claude Code to do a lot of SI work for you, even if it’s odd COBOL. And I think that’s really, really different to what, say, TPG will have set up with OpenAI or the new Microsoft company, when they think about FDEs in terms of the kind of quality, training and skill set of the engineers who are being put in place.

Kai: Right. You think a higher end... I guess the only, then the kind of argument would be more that if you’re working with an OpenAI-funded entity, they’re gonna try to jam their technology down your throat and not gonna let you use an architecture that might involve, let’s say, Anthropic, their rivals. So I guess there’s, again, there’s trade-offs down the line, right?

Azeem: There’s gonna be trade-offs, and that is gonna give, you know, competition is a great thing. So what’s that gonna mean? It’s gonna mean that if you are gonna insist on a lock-in through to OpenAI, you’re gonna have to offer something extra in order for a customer to accept that lock-in.

Kai: It better be a really good product.

Azeem: Yeah, a really, really good product, which then you may be willing to do.

Kai: Yeah, I think that’s right. And I think this serves as a perfect segue into the next question. So in your book, The Exponential Age, you talk about how the big winners of the exponential age, companies like Amazon, Google, and Apple, use intangible assets such as intellectual property, human capital, brand, and most importantly, network effects, right, to achieve unprecedented scale and just dominate over the past decade or so.

So if we accept the premise, which I think we both do, that intangible assets are increasingly the source of competitive advantage, then the key question for investors as we look forward to the age of AI is will AI reinforce or erode each of these intangible moats?

So for example, on one hand, we’re seeing software stocks, they’ve taken a huge beating over the past year as investors increasingly think that the AI coding agents are going to erode their technical moats. On the other hand, one could argue that AI actually increases the value of human things, say like trusted key customer relationships or proprietary data, distribution, maybe some specialized forms of human capital.

Right? So let’s talk about this, because I think this is really important for investors today, which is which intangible moats do you think become stronger or weaker in AI economy?

Azeem: I love your question because it is about that rotation, isn’t it? About which ones will be better or worse.

So I would say on brand, my guess would be that brand would strengthen because in a world of like infinite AI-generated kind of slop, we will kind of move towards things that we recognize. And people will pay for that assurance, and maybe the assurance is just through brand, or maybe the assurance is through something that’s a little bit more concrete.

On network effects, I think this is quite interesting because agents might weaken consumer network effects, right? If I’m mediating all of my activity through an agent, then I’m not gonna be visible within the network. So I think that that becomes less useful.

On IP. Look, so IP is a place where I think both you could feel very, very rapid erosion of value, and that’s where the software companies sit. And you might see that erosion of value as well if the data that you have is actually not distinctive enough, because it’s extraordinary what the models can now do even without any data.

But you can’t... I mean, I’m sure we’ve all had this experience, right? The model sees a lot about my customers even though I’ve not told it about them. So then the question is, like, what IP do you have that is sufficiently proprietary? And I think a lot of industrial businesses have that.

But more importantly, do you have a mechanism to deepen and improve that IP that’s kind of really explicit? Some kind of loop where you keep generating within your own business data that improves your proprietary view of the world. Even then, I’m not sure. I think some of the AI scientists think that the models will become so generalizable that they’ll need N of 1 to be able to make sense of the world. I’m not a scientist in that way, so I can’t say that they’re gonna be right. But my sense would be that if they’re not right, data of that type remains important.

And then, you know, human capital is gonna be really intriguing, and know-how. Anything that’s average is gonna become useless, and I think a lot of companies are gonna discover just how average much of what they do is. But knowing what good looks like in a given domain will become much, much more valuable, not less valuable. And certainly when I’ve spoken to the strategy consultants or, you know, bankers who work in complex situations, they’re actually finding their judgment more valuable, more useful, just as they find the judgment of their juniors less useful because the AI helps it there.

So I do think that the AI will hit these things differently. I should say something about software companies. I mean, I do think that, you know, you’ve seen ServiceNow, I think, has done reasonably well, and you’ve seen Datadog do reasonably well in all of this. And, you know, Salesforce has not done well.

But again, for me, it’s really a question about how well can they rotate their business models. Because if you are an enterprise who’s used Salesforce for a decade, all your data lives there, and it would make sense for Salesforce to figure out some kind of commercial relationship where they are hosting the AI over your data.

The question I think is just gonna be what ends up happening with margins, which I guess is perhaps what’s behind the fear that’s hit the stock price.

Kai: That’s right. Yeah. I mean, look, you know, I’ve written about this as well. I think, like, the idea is that if everyone has access to AI, as I pointed out before, it’s more the complementary intangible assets that give firms their moats.

And so then in the case of Salesforce, it comes down to, you know, how strong truly is the switching costs? How strong truly is their brand, their customer relationships, and potentially some network effects? But as you step back from the complementary intangibles to AI itself, you know, you wrote in your book that AI is the ultimate intangible asset, and you ascribe to it, like, data network effects.

Like, do you still feel that way? It’s, you know, five years later, like, if I’m using OpenAI and switching to Anthropic back and forth across models, like how much network effects truly do these models have? You know, ‘cause I think a lot of it depends on can they train on customer data? If yes, maybe, but to the extent that customers are not, you know, the Alex Karps of the world are not super excited about that idea, does that maybe obviate their disadvantage?

Azeem: Yeah, that’s a really good point. So back then, the idea, you know, we’d written “The Wealth of Machine Learning.” GPT was at GPT-3.

Kai: Yeah, it was five years ago. You were way ahead of the curve when I went on that call, yeah.

Azeem: Right. And Anthropic—that team was still working in OpenAI, and so there was a very simple model. You’ve got a data flywheel. You put machine learning into product. Product does better, gets more edge cases. That data goes back in and improves it.

But I think you’ve made a really important point, which is that these new generalized transformer AIs, they are much, much more interchangeable. But there is something around, I think, how fine grained you end up being with some of these models. So, you know, I know people who are now applying models and they’ll say, “Well, I’ll use OpenAI’s GPT series here, because it happens to be particularly good at this type of activity. But if I’m needing a long-running agent task, I might be using the, you know, the Claude model.”

And why did that happen? That happened because of the data that they were trained on, and not just the data, the kind of intangible know-how that the teams had when they were training on those reinforcement learning traces. But equally, I mean, as you say this question and I think it through, the frequency with which those can switch, who could switch because they are not contractually bound not to switch, is evidence that maybe that network effect is not strong. You know, anyone who can switch quickly does, and people who don’t switch quickly are switching slowly because enterprise IT has done a deal with somebody.

So I think you could well... I think you’ve probably put your finger on something there, yeah, that we have to revise and review our view on just how strong that data network effect will be.

Kai: Got it. Interesting. No, I appreciate that, and I think that’s, you know, that’s being a good investor, being a good thinker is being willing to kind of revise assumptions, you know, as the state of the world changes.

So, you know, one more question on this theme. So, you know, one of the kind of common threads through all your writing is the idea that technology’s become increasingly abundant. And so if intelligence becomes abundant, then the question becomes: what is scarce?

Azeem: Right. What does become scarce? I mean, look, intelligence is becoming abundant in really wild, wild ways, and we just, like, have to think about the ways in which we now might work. My son sent me a photo of his year two university exam result, and I don’t know how to interpret them.

I was at university 30 years ago, so I just sent them to ChatGPT, and I got back a, “Okay, he’s done well here, and, you know, a little bit of work needs to be done here.” So when intelligence becomes abundant, there are gonna be complements to it that will become increasingly valuable, and some of those might be, you know, verification.

So I’m gonna stand up and verify in a legal context that this is the output that you can stand by. ‘Cause right now an AI model can’t do that. And that’s the difference between going to a law firm and getting ChatGPT to read your contract, right? Because a law firm is effectively putting their neck on the block.

So I think that idea of verification as it extends out into liability becomes really, really important. I think another thing that becomes important is... You know, people talked about judgment and taste, and these are really, really easy words to use. They’re very, very hard words to describe and to maybe even test for.

But there is something about the judgment that comes out of experience that does make a difference. And, you know, I think about my new book, which I’ve just submitted the draft for, and what I can do with the AIs in that field is so distinctly more capable than what I can do in areas that I’m not a complete expert in.

So there is that scarcity around judgment that I think will become more and more valuable, and we’ll find better ways of explaining what we mean. Right now, I just think, “Oh, well, how many years have you worked in the industry? Okay, you’ll have judgment,” which is a terrible way of saying—

Kai: It’s an intangible.

Azeem: It’s an intangible. And I think the third thing that will become valuable will be things that are human-provenanced. And you’re already starting to see this in writing, so you know, a lot of people are getting annoyed about AI writing, and they’re getting annoyed that 75% of LinkedIn is AI written.

And I can see there becoming, you know, provenance marks or brands that will serve on—post organic, post vegan, but you know, this is human created, and therefore there should be a premium. And certainly what we do as a team now is we experimented with sort of like using AI in little bits and pieces of our output.

And now the only place where we use it is in the charts. We don’t even use it for captions, we don’t use it for social media posts. All of that is human written. But kind of constructing a chart off data, it’s kind of programmatic anyway, we’ll use AI. I’m not sure if that’s a great business decision, but I do think that that sort of human ownership will somehow be progressively more scarce.

Kai: No, I feel like I’m the same way. Like, there are a lot of things I do that I could be using AI for, but I just don’t on principle. I don’t know why. Again, maybe it’s a bad business decision, but there’s kind of some kind of like artistry and some kind of pride that one has in doing certain kinds of work that I think is lost if you fully outsource it to AI.

Azeem: Yeah. And, you know, as an investor, I think that’s really, really critical because, you know, AI is all about the mid, right? It’s the Gaussian and it’s kind of the weight of where the tokens are sits in the middle. But you know, if you’re gonna be the world’s best investor, you wanna be on one of the edges in some way.

And if you are coming up with hypotheses, and again, I found this with my book, of course I’m saying something new, and so I’m saying something that is not easily represented in the training data. So I spent three years reading hundreds of thousands of pages of different research material and thinking about it, and when I’m working with the AIs on my book, I’m being yanked back into orthodoxy, yanked back into a market view, yanked back into, you know, a world where all you ever read is the FT and The Wall Street Journal, and you think you’re gonna get a distinct kind of perspective on the world.

And I do think that that is... I find that actually quite helpful. So I will sometimes be testing ideas from the book based on how badly the AIs respond to it. Because the more badly they respond to it, the more sense I have, which is this isn’t in their training data, so this does have some novelty around it. It’s either really good or really bad.

Kai: Yeah. You’re on the tail of a distribution somewhere, yeah.

Azeem: I would rather be on the tails, honestly, than stuck in the middle. If you’re writing a book or you’re investing, you just don’t wanna be in the middle.

Kai: Yeah, I think that’s a really interesting way to reframe my question. It’s not that intelligence is becoming abundant, it’s that the middle of the distribution is now becoming abundant. Which means that, you know, for those who are just starting out in their careers or trying to learn a new thing, like I’m trying to learn how to cook, let’s say. Like, I’m a beginner. So just becoming average is a huge advantage, a huge uplift for me.

If I consider myself kind of on the right tail at, say, investing, then yeah, relying too much on AI actually dilutes my alpha and I’m better off, you know, actively trying to push up against, you know, what the recommendations of an AI might be.

Azeem: Okay. So what you can do, just on the investing, the one thing that you can do, of course, is, you know, the AI does make it easy to kind of construct tools for you. So something that might have been difficult, like really time-consuming five years ago, like, you know, how do I map every supplier to this particular company, which is really a data research exercise, is something that you can now do very, very quickly for any company at a high degree of reliability.

And so that constructs a tool that helps you do something that otherwise would’ve been time-consuming, and maybe you wouldn’t have done it first time around.

Kai: Yeah. I think there’s two areas, you know, to build off that. One is just the kind of gathering and assembling of data sets into like a structured format, and then the other one is the presentation of that data into like a dashboard format.

Right? Like, I’m building this, like, web app, just kind of a nights and weekend project for me. I don’t know anything about, like, JavaScript, HTML. I’m more of a kind of quant on kind of the research side. But it’s put together, like, a really nice dashboard that I find very useful, right? In just a small amount of work. So it’s pretty incredible.

So then, so my next question is just stepping back even further about your book. You define this idea of the exponential gap, that technology improves at this exponential pace while our institutions, laws, and social norms adapt much more slowly at, say, a linear pace. Where is this gap most consequential today, and how would you propose that our institutions adapt to help close this gap?

Azeem: It’s most consequential in the short term around the way that Americans are looking at data centers and how they are thinking about what a data center is and the arguments from both the local concerns, which is, you know, it’s gonna be noisy, it’s gonna be hot, it’s gonna, you know, be a constant buzz.

It might use up my water, through to concerns around what it might mean for, you know, climate and other types of questions. And the reason that that’s an example of the exponential gap is that effectively what those data centers are gonna end up doing is they’re going to, over the medium term, drive a ton of energy innovation.

And energy innovation has been so important for the success of the United States over the last hundred years, whether it was, you know, fracking and extraction of natural gas, whether it was the ability to scale out nuclear power in ‘50s and ‘60s, whether it was, you know, oil from Pennsylvania and beyond, and of course the Permian Basin, right?

So the US understands, JR Ewing understands the importance of a vibrant energy system, and the US energy system has been moribund for, you know, 10 years. And now there’s this injection of difficult problems to solve and rich customers willing to pay to solve them. And the narrative around all of that actually has got stuck in a general, “We don’t like big tech.

What is big tech doing? What is all this construction about?” So that for me is one good example. I mean, absolutely, like, I think it’s amazing, by the way, that both people on the left and the right are pissed off with data centers at the moment because I think that allows for a vigorous debate and discussion and ultimately a changing of the rules, right, that will happen. You know, that’s one example.

But I think the second example really, and the one that is longer term and harder, is about what is our relationship between credentials and what we value and what we incentivize for humans in their professional and educational development. And, you know, the favorite example is this Brown University economics professor who let kids have a midterm exam, which they could take back in their dorms, and people got 96%, which is, you know, that’s reasonable if it was an Indian university in maths, but not Brown University in economics.

And so he said like, “Listen, finals papers come in, it’s a closed book exam.” 18% of them, I think, dropped out. Scores dropped by half. Only three students got anywhere close to their original score. So the question here is: Why do the students think that the only thing that mattered was the high score rather than the actual learning?

And the reason is, and it’s true in the United States, as it is in the United Kingdom where I live, that education has long departed the process of education through to credentializing, and that credentializing where everything is oriented around the score, and even assessments have become about it.

And that is gonna be a real problem in a world of AI because we’ve just discussed how, you know, where you have a bit of judgment and expertise, you can use the AIs much, much better and you could get some, like, alpha, whatever alpha means in your world. But that gap where, you know, education, it’s not just what kids get taught in class, and I think many of the times they are taught critical thinking and so on, but it’s the way in which the output and the assessment focuses on something that a machine can do and can do well, and no one really cares whether a human did it or not.

So that for me is, like, the biggest sort of example for the exponential gap that we have to contend with right now.

Kai: Yeah, I mean, there’s so many challenges. AI’s just moving so quickly. And yeah, I mean, I think you’re very correct to at least highlight this issue, right, for policymakers and, you know, just people in society in general, educators, to think more about how they should be adapting the kind of norms and institutions to this kind of rapidly changing world we live in.

So as we wanna go close—

Azeem: Can I just share, I’m sorry. I wanna share one example with you, which is that I think that there’s increasingly now on remote interviews for developers that, you know, developers are asked to show that they’re paying attention without, like, ChatGPT running on the side, right?

They’re being kind of, you know, proof of human for the work that they have to do. And I think that one’s really, really important. It’s really important if you’re an employer as well because somebody who can think for themselves is going to be somebody who can use the AI tools better than someone who can’t.

Kai: Exactly. Yeah, no, that’s 100% correct. Um, yeah, so on the topic of AI tools, you know, as we wind to a close, I wanted to ask you this question. So I’m a firm believer that in order to truly understand a technological trend, you gotta roll up your sleeves and be an early adopter yourself.

Now, you’ve done a ton of fun videos and posts on how you are using AI in both your work and your personal life. What do you think has been the most consequential change that AI has made to your personal productivity or just the way you go about doing things?

Azeem: Oh, you know, for me it’s all about the ideas. So I record most of my meetings, and I have access to all of them. My researcher who worked with me for five years, she’s left to go and travel around the Central Asian republics for a year. She left me hundreds of Granola transcripts and all of her Google Docs already fed into an AI system so I can query, “What would Chantal have thought about this?”

And you know, it’s not given me a lot of time. It didn’t originally, but now what it’s done is it’s allowed me to go much, much deeper than I ever do on a subject. So typically now, I mean, I used to try to be first principles a little bit like you, Kai. But what I do now if I’m dealing with a question that I’m thinking about for my book or for an investment is not only can I get an AI to go off and get first-hand data from a dataset or crawling a, you know, 100 company websites, but at the same time I can go off and I can look at what the peer-reviewed academia has said about classes of questions like this.

So I end up with these bundles of dossiers and, you know, especially with the academic papers, I go back and read them by myself. But the biggest problem was finding them in the first place, right? And now the AIs can do that for me. So, as somebody who trades in ideas, it has been absolutely transformative because I can just get to the first-hand data that I like, stick it in the spreadsheet, analyze it somehow.

But I can also, you know, stand on the shoulders of all the giants who have looked at this question over the previous 200 years, and lean into that in a way that would’ve felt like a real privilege just three years ago.

Kai: Yeah. It’s really quite powerful as a research assistant. I think it was Ken Griffin who was on a podcast recently talking about how some of his employees are using AI to replicate academic finance papers to, like, say, “What’s the anomaly? Let’s make sure that this is, you know, reproducible and potentially additive in our process.”

I mean, I’m using it in the same way, by the way, as a kind of coding, using the Claude Code and Codex as a way of implementing and testing hypotheses that either are generated by myself or through other means. So yeah, you know, definitely a kind of a force multiplier, I guess, on, as you point out, people who trade in ideas.

Azeem: Hmm. Yeah.

Kai: So yeah, just my one final closing question which we ask everyone is, you know, what is one thing you believe about investing that the majority of your peers would disagree with?

Azeem: For a long time, the thing they disagreed with was the speed with which AI was growing. Back in 2015, I wrote an essay saying the demand for compute was gonna increase by a factor of a billion over the next 10 or so years. And I said, look, AI is gonna be even sharper for chips than Moore’s Law. And, you know, I think people have come around to that point of view.

I think the other thing that I would add that has maybe perhaps not unpopular with my peers, but most of my peers are venture investors who invest in private companies, is that increasingly I’ve been investing in public companies, and the reason is, it just takes so long with private companies.

It takes an enormous amount of time, and you have very, very little control for, you know, the sweat that you put in. And so that has been a shift for me over the last five years.

Kai: No, I like that. As a public markets investor, I’m a little bit biased, but I like that. Well, thanks so much for your time, Azeem. I really appreciate it. You know, this discussion was super fun, and I think we covered a lot of ground. We’ll have to have you back on at some point.

Azeem: Thank you so much, Kai. I really appreciate it.