Jack: Welcome to Excess Returns. I’m Jack Forehand, joined by the better half of our hosting team today, Kai Wu of Sparkline Capital. And today we’re really lucky to have Dom Rizzo on. Dom’s the portfolio manager of T. Rowe Price’s Global Technology Equity Strategy and the firm’s technology ETF, and we are gonna talk about the thing everybody’s talking about these days, which is AI and everything that’s going on there.
And we’re gonna get deep into the technology and what it might mean for the economy and a lot of different things, but we’re gonna start at a high level, because it’s funny, when I put these questions together, it was like three days ago, and I was gonna talk about the big correction we’re having in these types of stocks, and then two days later we’re back to a rally.
But I did wanna ask you about that because one of the things I read in our prep for this is you had a quote that said, “This feels a lot like the 1998 sell-off to me, which proved to be an incredible buying opportunity.” So I’m wondering if maybe you could talk about what we’ve seen recently and that quote.
Dom: Yeah. Well, first off, thanks for having me, guys. It’s great to be here. And for the listeners, I think it’s really important you know that I love the question list that you guys sent over.
Jack: Well, thank you.
Dom: So I’m really excited for this pod. Look, well, first off, let’s take this with a grain of salt. I was five years old in 1998, so this is not personal stock picking experience. But from someone who loves markets and cycles and studying bubbles and reflexivity—look, if you did rewind to ‘97, ‘98, I think there’s a couple similarities.
So one is really high momentum factor in terms of day-to-day trading volatility and month-to-month trading volatility. And if you just look at the statistics from June and July, June was literally in the top 4% for momentum factor for monthly returns, and then July was in the bottom 1%. So we’re in a high vol, high momentum-driven market, either on the way up or on the way down, and that has some similarities.
You have some geopolitical shocks. So Asia financial crisis was more ‘97 heading into ‘98, but Iran war and questions around oil pricing and where that’s gonna end now. So I think that there’s some similarities there. I think there’s some similarities in the technical trading around large hedge funds having some issues as well, right? Long-Term Capital Management was many, many multiples bigger from a systemic risk perspective than situational awareness. But, you know, I’ve seen estimates of situational awareness’s public gross exposure being north of a hundred billion dollars. And that was obviously unwinding through much of July now, in hindsight. So I think there’s some similarities there.
There is a fairly big fundamental difference with ‘98, and that is actually in the fundamentals themselves. So if you go back and look at the semiconductor industry in 1998, you saw an 8% revenue decline, right? That was mostly driven by the decline in memory pricing. But then if you look in 2026, industry sources have something like a 64% revenue increase. So, you know, there’s similarities—like all great things in history, it rhymes, it doesn’t repeat. But we are clearly not seeing that decline that we saw in 1998 from an overall revenue perspective, yet we’re seeing almost similar price action, right? You know, the SOX went down 40% in 1998, and we saw a 30% drawdown recently in the SOX. So some similarities, not perfect, but I think overall thinking about this as a 1998 style correction that results in an even stronger follow-on is kind of a good mental framing.
Jack: Another implication of your 1998 thesis is that we are nowhere near the end of this capital cycle. I think you’ve talked about we’re about halfway there. So this would look more like a speed bump on that thesis than the end of the road.
Dom: Well, you know, one of my favorite charts is actually just the performance of the Nasdaq from the launch of Netscape through the end of 1999, 2000. And if you look at that and then you put the Nasdaq from the launch of ChatGPT through today, it’s literally the halfway point, right?
So ChatGPT was launched roughly three years ago. I know this because I took over our global technology strategy on December 1st, 2022, which was the day after ChatGPT was released, November 30th, 2022. So, you know, in life, sometimes you get lucky on timing, and that ended up being one of them, right? Semiconductor background, taking over the global technology strategy. And then we were AI on and semiconductors on, right? And so that’s been, in hindsight, the right place to be.
But if you think about where we are from the spending cycle, I think we are actually right at the point of acceleration on high numbers, right? And I think that’s hard for some people to grok ‘cause the numbers are already so stupendously big. If you look at the hyperscalers in CapEx spend, they’re gonna grow, say, 75% in 2026 to roughly $800 billion of spend. The Street, broadly speaking, thinks that we’re gonna see 20 to 30% CapEx growth next year, which, obviously depending on where you end up, but call it like 1.1, 1.2 trillion of CapEx.
I actually think we’re gonna see an acceleration in CapEx next year, so growing faster than the 75% growth we’re gonna see in ‘26, which will put us at kinda 1.5, 1.6 trillion of CapEx spend for the hyperscalers. And what’s so amazing about that is I actually think that will happen in a relatively moderate pricing environment for memory. So a lot of the increase in 2026 was memory going from mid to high single digits of the overall capital budget to, say, mid 30s of the overall capital budget as pricing went up hundreds of percent in memory. I don’t think we see nearly the memory price inflation that we do in ‘27, and despite that, I still think we see an accelerating CapEx number next year because the ROIC is just so attractive for the hyperscalers.
Jack: Was there anything we learned from these recent earnings reports? We just had all the big tech earnings reports, and it seems like the market liked Microsoft, it liked Amazon, and they disliked other ones like Meta. And we don’t wanna talk about the individual companies, but was there anything at a high level you learned about the overall ecosystem from those reports?
Dom: Yeah. Well, I’m very happy to go into why I think the stocks reacted the way they did.
Jack: Yeah, that’d be great.
Dom: I think it’s really a function of what the market’s perceived ROIC is on this capital build-out, right? And so let’s take Amazon first, ‘cause that was the clearest articulation by any of the hyperscalers on what the CapEx really would translate into, to revenue and cash flow, out of Jassy.
So one, they broke out the different pieces of the CapEx, right? There’s long lifecycle assets and short lifecycle assets. Long lifecycle assets being property and actual infrastructure, and then the short lifecycle assets being the chips, the networking equipment, the data center spend, right? And Jassy laid out on the call that on that short lifecycle piece, which is the piece that everyone would naturally worry about the most, that they break even in roughly two to three years, and then they last five to six years of useful life, right? So two to three years to get your money back, and then two to three years of great cash flow.
And when you think about what that means, it means that even on the shortest lifecycle piece of this, the company should have very, very strong ROICs, right? And then when you take that in the context of AWS growth accelerating to 37% at 50% incremental operating margins, you don’t just have management saying ROICs are gonna be strong. You see that coming through in the acceleration of the revenue growth and the EBIT margin expansion as well.
And then you take that in context of Microsoft Azure growing 43%, guiding to accelerating to 45%, in context of Google Cloud Platform, GCP, growing 82%, also at 50% incremental operating margins. I think you have a lot of evidence that the hyperscale CapEx business is hitting that inflection point. And then I think the cherry on top of all this was Satya Nadella kinda retweeting the Morgan Stanley analysis of 30% ROIC for the broader hyperscaler universe. That tacit blessing I think gives people comfort that this immense capital build will also be coupled with revenue acceleration and operating margin expansion out of the major hyperscalers.
Now, if you wanted me to take a step back and say, “Hey, that sounds really good. Where are some potential pitfalls that we should really think about?” The main pitfall is how much of this is actually just driven by the AI labs, OpenAI and Anthropic, right? OpenAI is already something like 25% of Azure revenue. It’s probably mid-30s or a higher percentage of the backlog, probably even higher as a percentage of net new bookings. So I think there’s a real scenario where the labs grow so ridiculously quickly that they could eventually aggregate frontier intelligence and put pressure on the hyperscalers. But that scenario is a few years out at least, and in the short term, we should have very, very strong fundamentals out of the cloud companies.
Now, you kinda alluded, some of the companies reacted well to their earnings and some companies bad. Obviously, the company that acted poorly after was primarily Meta, and the thing that’s tough for the market to get around with Meta is that they’re investing all of this to just make the core business better. And the core business has already accelerated, right? If you go back a few years ago, the world was thinking, we used to talk about Meta, and we used to talk about, oh, what about the China Q-Comm, there’s comps, right? You know, there’s a world where because of advertising from Temu, Meta may decelerate to high single digits.
Okay. Well, that’s not the world we’re living in, obviously, right? We’re gonna see mid-twenties growth out of Meta this year. But this is about making their core business better and AI being an existential platform that Meta has to win in order to remain very strong going forward. Not necessarily accelerating a business that’s so obvious to see like it is Azure or AWS or GCP.
And so that’s the slight difference. I actually think all these companies are making the right decision by spending. When you have a platform shift like AI, and it’s structurally capital intensive, one, you have no choice but to spend. But then two, I think the returns of spending are actually extremely high, as we’ve already seen demonstrated. So that’s what gives me confidence in this acceleration in CapEx heading into next year.
Kai: I think you are right to point out that downstream of the near term demand are the labs, Anthropic and OpenAI, and then they themselves, of course, have customers, right? So going back to your question on ROI, the cloud companies have obviously been realizing some ROI. But the question, I guess, is more on sustainability of that demand vis-à-vis the end users, right? Because ultimately, the end user pays Anthropic for the tokens, who then pays the cloud hosting companies.
You know, we’ve obviously seen some news around pullbacks to, say, Uber, maybe even Coinbase, around people thinking less about token maxing and more about, “Hey, how much are we actually getting in ROI per unit of token spend?” How do you think that trend... And of course, pressure from Chinese labs and Chinese companies and open source as potentially being an alternative, pricing wars maybe even, with OpenAI and some other model companies introducing lower tier models that are more price efficient and compute efficient. How do you feel like that plays into this thesis, not just with the hyperscalers in general, but the whole AI complex all the way down to memory and chips?
Dom: It’s a great question. It’s probably the thing I think about most right now, right? Because it has so many ramifications to what happens to the chip ecosystem, what happens to the hyperscaler ecosystem, what happens to the application software ecosystem, what happens to infrastructure. Every single question is really downstream from this concept of what do the frontier labs look like in a few years.
And let’s just take a step back and think about the past year. When I did the Anthropic round last year in the global technology strategy, they were doing five billion dollars of run rate ARR. This is last summer, right? The latest rumors on Twitter or whatever, any third-party data source, would put that number well north of $70 billion. Okay, I’ve never seen a company grow that fast. I mean, how many ServiceNows is that? That’s like three ServiceNows or something, right? It’s really stunningly fast growth.
And why is that? It’s because coding as a use case went completely vertical, and then Claude Code and OpenAI Codex really just caused complete virality within the enterprise, right? And why did coding go vertical? It’s because you clearly made your average coder 20 to 30% more productive, and your most successful AI native coders probably many-fold more productive.
And what’s so amazing about that is, ask any engineer in the world right now, has that resulted in any decline in their day-to-day work, or are they working harder? I mean, every software engineer is working the hardest they’ve ever worked right now, despite being 20, 30, 40% more productive, right? And so when you’re in the world of making white-collar labor more productive, I think there’s very, very high returns because enterprises are willing to pay for that productivity inherently, right?
And so how big is the coding TAM, right? It’s a question I get all the time. Well, there’s thirty million people who write code for a living. Let’s say we pay them each a hundred thousand dollars a year, so we spend three trillion dollars on coding knowledge work as a society globally. Let’s say we make them each 20% more productive, that means that the three trillion dollars of spend is six hundred billion dollars more productive.
And then you say, okay, what right does OpenAI or Anthropic or Cursor, I would argue, maybe, have in that six hundred billion dollars of spend? I don’t think it’s crazy to say half. It may even be more than half, right? And so that’s three hundred to four hundred billion dollars, and then you compare that to the overall application software business being three hundred to four hundred billion dollars of revenue. And you’re like, “Wow, coding in and of itself is as big as all of application software,” and that doesn’t even include HR, legal, finance, all these different TAMs.
Okay. So then your question is, what if you have open weight models commoditize that TAM? What if you have N minus one models come in and the returns aren’t to frontier intelligence, but they’re to N minus one models that are simply good enough? Or what if that’s way too much money to spend on coding, and enterprises aren’t seeing the return, right?
So one, I think enterprises are not stupid in general. I think actually most enterprises are quite thoughtful, and clearly they’re getting some value or they wouldn’t be spending this money, right? This concept that people would spend money and not get a return on it I find a little silly, but let’s put that to the side.
I do think what we’ve learned is that almost all the economic return is accruing to the frontier from a revenue perspective, even though the vast majority of tokens may actually be open weight or open source or N minus one to be more efficient, right? And I think a really good analogy may be Apple versus Android. But broadly speaking, I think we’re gonna live in an 80/20 world where 80% of the tokens are probably open source or open weight or N minus one, and 20% of them are frontier. But then the vast majority of the value accrues to those frontier models as they orchestrate the other models, as they make it more efficient.
And then I think there’s a real question structurally whether or not OpenAI and Anthropic are best positioned to give you that whole spectrum of models, right? Are they actually the most profitable and willing to give you... I don’t know if you guys saw the ChatGPT pricing decreases for the lower models last week. I think that was Sam’s bet, that you wanna be the most efficient at every level, the most efficient per task at Sol, Luna, and Terra, right? Like at all the different levels of the model. Or do you want a Palantir or a Microsoft to sit on top and help you kinda model swap to the most efficient model? I think every enterprise is gonna make a different decision on that. My gut is just being at the frontier with the frontier labs is the best way to do it. But let’s see. I think open question.
The last thing I’d address on your question—I mean, I’m going for a while, but I think this is probably the most important question, so I could do a full podcast on this—is what about examples like Uber and Coinbase? And I would say, “Hey, wait a second. Uber and Coinbase are not the average Fortune five hundred company,” right? Uber and Coinbase are tech-native companies that have the expertise to either buy billions of dollars of GPUs on their own and run these open weight models, get the most efficient pricing and efficient utilization out of the hyperscalers, and build their own harnesses on top of the models, which they can hot swap underneath.
I think that is quite rare. And I’ll just pull a random—I don’t think Pfizer is gonna do that, right? I have no clue what Pfizer’s plans are, but a random generic Fortune five hundred company, I don’t think, is going to do that. I think they are gonna either rely on a Microsoft and Palantir to try to do it in a model-agnostic way or simply rely on the labs. And that battle remains to be seen, but I think you’re seeing the world starting to line up on either side of that, and that’s really exciting from my perspective. I think I answered the question.
Kai: No, that was a great response. And I think you’re right to point out that end user ROI, that’s the game, right? If there’s no pie to slice up, then there’s nothing. What are we doing here?
Dom: And that I think comes from productivity, right? That comes from making the labor force more productive. That’s what unleashes the TAM.
Kai: That’s right. Yeah, I mean, if we’re not seeing productivity gains, then the debate over how that pie is sliced up is moot. So let’s presuppose for now that there are productivity gains, since obviously that’s a philosophical debate that would take three podcasts to cover.
But as we think more about the value chain, you already brought this up a couple of times. So I think this is really interesting, right? You mentioned memory, right? Over the past year, because of the bottlenecks, they’re taking a larger share of that pie. But you said something interesting earlier on, that you actually think that will alleviate. So that’s one thing that’ll be interesting to discuss. You mentioned the hosting companies, the applications, the model labs, maybe system integrators, the Palantirs of the world, Microsoft.
I guess that’s my question, which is, as we think about where there are moats and where there are... what will become commoditized, right? Because if you look back—and you mentioned you’re a historian, you’ve studied these past technological cycles—one of the weird things you see, and it’s kind of ironic, is that in pretty much all cases, or maybe almost all cases, you find that the actual builders of the infrastructure don’t actually make the profits, and that the profits actually flow downstream to maybe users or other parts of the ecosystem.
So let me ask you that, which is, obviously the frontier labs is where everyone focuses. But if we step back and ask, what are the non-obvious areas where profit pools might accrue, where there actually might be surprisingly more moat than one might think? What do you think would fall into that category?
Dom: Yeah. Well, I think so much of it does depend on this open weight versus closed weight debate, right? Because first off, I think it’s important that everyone realize when you say open weight, it does not mean free, right? You can’t take the Kimi K3 model and run it on your laptop, right? It’s a 2.7 or 8 trillion parameter model, I can’t remember, where they recommend that you use 64 different accelerators to run this properly, right?
Okay. In order to run a Kimi K3 model effectively at enterprise scale, if you wanna do so on-prem, you need to go invest in like billions of dollars of NVIDIA gear to make that work, right? And so what’s happening, the hyperscalers or some of these new AI companies are saying, “Hey, we’ll do that for you, and you just rent it from us,” right? That’s one business model.
The other business model is, “No, we’ll keep the closed weights. We’re gonna build the amazing harness on top, and then actually we’re gonna build different applications on top of it for you.” And that’s what Anthropic and OpenAI, I think, are trying to do over time. And I’m far more partial to the concept that they will be the best positioned to build magical products that delight the enterprise and go viral.
And I actually think, one analogy I’ve been playing around with in my head—and it’s not perfect yet, so if it sounds choppy, that’s okay—is what if AI manufacturing and outcome manufacturing is actually far more akin to semiconductor manufacturing, right? Structurally capital-intensive, increasing cost curve. And what if there’s a real synergy between the design of the harness and the manufacturing of the token itself through the model? And that would actually lend itself to a vertically integrated solution, just like it lent itself to Intel dominating the CPU market of the ‘90s, right?
And so that would lend itself to OpenAI and Anthropic running away with it, which I think is a real possibility, and I think that’s why everyone’s freaking out on the other side. Jensen’s lining up, Alex Karp is lining up, Satya’s lining up, ‘cause they all know if OpenAI and Anthropic run away with the game, it’s a really big issue for their core businesses, and they are eventually commoditized.
So why that matters, though, is if it’s truly vertically integrated token manufacturing, then the economics go to two places, right? They basically go to labs and to the chip infrastructure companies, right? Eventually all roads lead to TSMC and ASML and maybe Intel too, right? Eventually in that world.
In the world that NVIDIA, Microsoft and Palantir would prefer, there’s a lot of economic profit left for NVIDIA and the clouds, and a lot of those applications will be built by smaller application companies that sit on top of, say, the Microsoft Azure platform or the Amazon AWS platform. And that token manufacturing is actually just who can get the lowest cost per token in a pure commodity market, and then the economic profit can kind of flow differently.
First off, my gut is that this is just so big that they’re all gonna be pretty successful. And I don’t know if it’s 60/40 vertically integrated versus open or 70/30 or maybe 60/40 the other way. I’d say at the moment, as we stand on August 4th, 2026, I’m probably in the camp that it’ll be 60, 70 in favor of the frontier labs over the long run, but let’s see. I mean, I can change my mind right after this podcast. So it kinda depends.
What does that mean in the short term to medium term? I do think that means that chips lead us higher because whether it’s open-weight or closed-weight models, the investment in chips is immense. And then what type of chips, right? There’s memory chips, there’s logic chips, there’s optical chips, there’s semi capital equipment, there’s semiconductor manufacturing, there’s EDA for design software. There’s a million different places. Right now, I think the most attractive place is the logic semis, particularly the CPUs. But we’re also recording this right before they all report. So let’s get the incremental information of the reports. But right now, I think as memory prices normalize into next year, logic chips can accelerate. And when I say memory prices normalizing, I don’t mean that we see a decline, but just a deceleration from this incredible pricing growth that we’ve seen in memory.
Kai: So I wanna follow on. You mentioned the short term as opposed to long term, and you also brought up the hedge fund situational awareness in the beginning, which—well, we’ll see. Jury’s still out on whether or not the long-term view is correct. Many people believe it is, but obviously in the short run, he did not survive, or he probably will not survive.
So that gets me to the next question, which is around the financing of the build-out. You know, obviously, there was a period in time when building out AI data centers was a rounding error on Google’s free cash flow. And we’re now well past that, in which case we’re approaching the point where this needs to be funded not out of free cash flow, but out of equity, out of debt, right? And we’re increasingly seeing debt financing. Some people point out to circular financing in a derogatory way, but there’s clever ways that this is being financed.
Do you see that as a near-term risk? That maybe even if in the end game this build-out does occur successfully, and this technology is game changing, but that there’s a hiccup along the way, as we’ve seen so many times before, around the way that this is actually financed, because several trillion dollars is a pretty big number.
Dom: Yeah. So look, on Leopold specifically and the situational awareness blowup, I would just say, I think the paper that they wrote was really quite prescient. I think the podcast he did two years ago with Dwarkesh really laid out almost to a T what happened the next two years fundamentally. And then there was clearly the portfolio construction leverage mistake, right? Which resulted in—who knows if it’s full. We don’t know the details of the situation, right? But clearly substantial capital destruction. So, you know, leverage helps on the way up, and it really hurts more on the way down, right? As so many people have learned before.
In terms of equity and debt financing of this build-out, my job was easier last year, ‘cause when people would bring up where we are in the bubble, I would always say, “AI has the potential to be the biggest productivity enhancer since electricity. Productivity-enhancing technologies come with speculative bubbles. My job is not to miss bubbles, but to navigate them responsibly for our clients by trying to capture upside and blunt downside.” I know that’s a mouthful, but I’ve practiced it a few times.
But last year my job was really easy ‘cause I would just say, “Oh, well, this is just being funded by the free cash flow of the most successful organizations of all time.” That changed this year, right? Google issued equity, eighty-five billion dollars of equity. Here it is, either the most profitable or the second most profitable or the third most profitable company in the world issuing equity for this capital build-out, right? So we’re clearly in a different stage, right? And when we talk about where are we, fourth inning, fifth inning, are we halfway through—we are in the capital cycle part of the build-out.
And so if you ask me what worried me in July, I would say the only thing that worried me was if the price correction could be so material that it could freeze the equity or the debt markets in terms of funding the build-out, right? Could the price correction cause a reflexivity shock to the negative side where people wouldn’t be willing to believe, right? CDS spreads, everything, talking about all these different companies, where are we at in terms of debt versus equity? Where should their bonds trade?
Look, the reality is if you do the math, it’s actually not that big for these companies, right? I know that it sounds like a big number when I’m about to say, “Oh, the funding gap’s a few hundred billion dollars each”—and each company’s different, to get to those higher end scenarios, even less. But it’s actually not, in the context of a three, four, or five trillion dollar market cap, right? In terms of the funding gap between how much they actually need versus their equity, and then it’s really not that big in the context of their net debt to EBITDA, right? Just in terms of their leverage.
I do think you can make a compelling argument that the amount that needs to be spent is so much bigger than what the investment grade market is used to, that maybe it overpowers the IG market. But I actually think that’s why you’ve seen yields go up. There’s a crowding out effect, right? Google is willing to pay you a real number now. And what’s the difference to Google of fifty basis points, right? Nothing, actually, in the grand context of things.
So I think we are clearly in the equity and debt financing portion of the build-out. I don’t think the spread is that wide between what they need to raise versus what they want to spend, or what I think they probably should spend is probably a better way to phrase that. And then as the ROIC kicks in, you start to see the inflection on the operating cash flow that can continue to fund a lot of this build-out. And then I think the fundamental question is, is AI structurally capital intensive forever? Or are they able to grow revenue materially faster than CapEx over time?
Jack: To your last point, is there a case for more and more self-funding here? I know you listened to it as well. This podcast with Gavin Baker just recently came out, and he was talking about this idea that the cost of compute is going way up, and as these contracts kind of roll, it’s gonna go up and up, and he was making more and more a case that more of this could be internally funded versus needing external funding. I mean, do you think that’s fair?
Dom: I think it’s gonna be a combination of both external and internal. Look, I mean, if you see continued acceleration under the hyperscalers, your ability to borrow is easier too, right? Because the market is willing to lend you more money if your core business is accelerating, or your ability to issue equity at more reasonable rates is easier too.
I do think there is a self-funding mechanism, which is simply the operating cash flow, the OCF acceleration, right? As that accelerates, as the ROIC kicks in, there’s just more dollars to go spend. And then do I believe in the world where we see material price increases? The part of the podcast—which was a great podcast you’re talking to—is when they’re talking about pricing increasing for older GPUs, as the returns to frontier intelligence continue to go up, and then there’s a scarcity of GPUs. I like the idea a lot. It would go against everything I know about GPU pricing historically. But let me put it like this: I think it’s a distinct possibility that that’s the case.
And what happens is there’s a race between demand and supply, but the weight is so high on the demand side right now that every time you get supply increases, actually demand increases faster. And I don’t see that changing ‘cause I think we’re so early in AI. You know, Ben Horowitz said something like he thought we were 3% penetrated in AI, and I heard that number, and I thought it was right. I think that’s probably where we’re at.
Kai: You obviously have a semis background, and this is your bread and butter, so maybe I know what you’re gonna say. But some have argued that while historically chips have been a cyclical industry, we’re now entering this kind of golden age of a super cycle, right? Where do you fall on that? Obviously there’s significant near term demand, 3% penetrated. That, of course, stands to reason. But we live in a capitalist economy where profit margins attract competition, investment, capital cycle you already mentioned, and potentially innovations with regards to more efficient models that require less compute per unit of intelligence. Do you see downside risk in terms of chip stocks kind of regaining their cyclical characteristics, or is that not really something you’re worried about over the next several years?
Dom: Well, semiconductors are always cyclical and will always be cyclical. Let’s be very clear. We just happen to be in a great up cycle right now. I think this is why I love semiconductor investing, ‘cause both things can be true at the same time, which is we will have a correction, whether it is inventory driven or demand driven or supply shock driven. We will clearly have a correction at some point. How aggressive is the correction? Where does it show up in the supply chain? It’ll be a function of a lot of different factors at the end use case.
And most notably I’d say, AI is structurally different than software. AI is token manufacturing, and token manufacturing is capital intensive, right? That’s the big difference between software or internet. These companies could scale to tens of billions, hundreds of billions of dollars of revenue without material CapEx. Now, the software guys had to do it with a big sales and marketing budget, right? Which they used a lot of stock-based comp to get around. But the internet guys didn’t have to do that, right? The internet guys just hit two billion users, and they were able to monetize the world.
And that was like the beauty of Ben Thompson’s aggregation theory, right? That it really was a zero marginal cost industry, and then it was who had demand. In the end, eyeballs ended up being right, right? If you had two billion eyeballs, you were able to monetize incredibly effectively through either advertising or—I mean, advertising obviously being the main one on the internet.
So okay, AI is different. And why is AI different? AI is different because of the scaling laws. Roughly speaking, if you throw 10X more compute at a problem, you get 2X more intelligence, right? And so naturally people say, “Well, isn’t there a limit to intelligence? What’s the difference?” You’ve probably talked to someone with a 130 IQ and probably someone with a 140 IQ, and they may have both been a little boring. You know? What’s the difference between a 130 and a 140 IQ person, versus a 150 IQ person, right?
I would actually say don’t think about it as IQ, think about it as task completion. And once you start thinking about it as task completion, a million use cases open up your mind of what the models can’t do today versus what they can. So I’m a big user of Codex internally. I have all these agents running around making me charts, giving me data, all this stuff. I can think of a million things that it still can’t do that I want it to do, right?
And it’s funny, my Codex wasn’t working the other day. I had to use traditional ChatGPT, and I felt like I was going back in time. I’m living in this agentic world. I had to go to a chat-based world, not an agent world, and I couldn’t ask it to do tasks on my behalf, and it was really frustrating. And so I actually think what ChatGPT is doing, stuffing Codex into the ChatGPT user interface—I mean, it’s a big bet, but I think it will end up playing out really well for them.
So is AI structurally capital intensive? I think so. I think so because of the scaling laws, and there’s just been no evidence that the scaling laws are ending. Doesn’t mean that at some point we can’t see a blip in the scaling laws, doesn’t mean at some point we can’t see a supply shock out of China, doesn’t mean at some point that we may not have enough physical space to put the chips. All of those are real things that could result in inventory correction. And, you know, semiconductors are at the end of the bullwhip, and Google sneezes and a small cap optical stock catches not just a cold—I mean, potentially almost bankruptcy historically, right?
So that’s the part that’s hard for people, but that’s the part I love about semiconductors too. So I don’t think semiconductors are no longer cyclical, but I do think you’ve gone from a world of software and internet being the driver via PCs and smartphones to a world of frontier intelligence being the driver, and frontier intelligence being structurally capital intensive.
Kai: So what do you think is then the role for software, right? So SaaS and software was at one point the growth darling of the stock market. Obviously, they’ve been punished the past year, two years, stocks down fifty to eighty percent. And you make the correct point that AI is structurally different, unit economics are different than software. But you kind of make the case that through maybe vertical integration and harnesses, a lot of what it sounds like traditional software would have done can be done or should be done through agents and through OpenAI, Anthropic, and folks like that. So what’s left over, I guess, for the software companies? Or is that maybe too bearish a framing of your view here?
Dom: I think if you’re an established system of record, you get to be a dumb data pipe into OpenAI or Anthropic in the world I’m describing, right? The Salesforces of the world, the ServiceNows of the world, the Workdays of the world get to put their data into OpenAI, Anthropic, and then the intelligence can work on top of it.
I think you’re already seeing that, right? This whole push by Salesforce to go headless is a concept that they know the direction of travel is that the user interface is gonna migrate away from salesforce.com, the actual app, to either agents or to a chat interface, right?
And so I think traditional application software is in trouble. And the last thing I’d say is I haven’t seen a single compelling AI version of traditional enterprise software from a traditional enterprise software company, right? I could name dozens of fast-growing startups that are AI native. I don’t look at the products that the major software companies have put out and say, “Wow, that’s really compelling,” and it’s not just bundling to get them to those reported ARR numbers, right?
So I think that may have something to do with designing AI native software actually being quite different than traditional software, and understanding how the models play with the harness. Designing to an ever-increasingly smarter model is actually quite different than the world of traditional application software.
I do think that there’s different types of enterprise software companies. So as an example, SAP has a massive data gravity that’s a really interesting asset that they have through their ERP. There’s consumption software like Snowflake—”Hey, dump all your data of your organization into Snowflake, and then let the models run on top, or use our own coding agent to kinda make things better within Snowflake.”
Don’t get me wrong, I’m not saying that there aren’t areas. They’re probably primarily in infrastructure software, where you can feel more comfortable rather than application software, ‘cause the entire application interface has changed from humans to agents. And so that’s gonna be a tough transition, and I think we’re already seeing it.
And then there’s only so many dollars, right? If the IT budget is gonna start going towards intelligence tokens—I do think IT budgets are gonna grow, but what percentage of your corporate budget is gonna be intelligence tokens in five years is a really interesting, important question. And what percentage of your IT budget is that? Like, what if 100% of your 2026 IT budget is tokens? I think that’s a possibility in some medium time horizon.
Jack: How about the other side of the coin from the productivity, which is, I guess the right way to call it is labor displacement instead of job loss. But if we think about a continuum here in terms of a technology like this—I mean, I think we all think AI is gonna grow dramatically.
Dom: Yep.
Jack: And then the question is, if we’re way on the productivity side, then we kinda get the world of abundance people are talking about. If we’re way on the job loss side, we get the Citrini piece thing, and then maybe in reality, we get somewhere in the middle. I’m just wondering if you have any thoughts on that balance of those two things.
Dom: Well, I think it’s dishonest to say that AI won’t displace some jobs. I think AI will clearly displace some jobs, just like all technology innovation has always displaced jobs, right? How many bank tellers versus ATMs, right? The role of the typist pool versus email and Word, right? There’s clear jobs that change because the economy changes and technology changes. And I think to say that AI won’t displace jobs is dishonest, so people shouldn’t say that.
On net, I think we are gonna grow a lot of jobs because of AI, and I think that we’ve actually seen that happen so far, particularly in blue collar jobs, right? Electricians and plumbing for data center construction. But, you know, we were kind of joking about, ask your average software engineer, are they working more or less right now. I think they’re clearly working more right now, ‘cause I think there’s so much to do.
And so I do lean more towards this age of abundance camp, which is productivity results in revenue growth, revenue growth is good for the economy and good for the whole. GDP growth should accelerate in the US, and if GDP growth accelerates, that results in more jobs, right?
And I always think of Keynes saying that, you know, whatever his prediction was, 50 years from now, we’d only be working two hours a day for two days or whatever the exact prediction was. I don’t remember. But obviously we’re not doing that ‘cause humans are status creatures, and we love building, and there’s labor’s a love, and every startup’s a labor of love of the founder, and we build things, and it’s exciting, and we’re project based.
And I think there’s so many ideas that we haven’t thought of. And then you say, “Dom, well, what are those amazing jobs that you haven’t thought of?” And I say, “I have no clue.” And that’s not a great answer. But if I had told my great-grandfather that, “Hey, I know you just got off the boat at Ellis Island, and one day your great-grandson is gonna have a personal trainer, and his job is gonna be sitting behind a computer screen”—and describe my day—I mean, it wouldn’t even resonate, right?
And I think we’ve seen immense technology innovation and productivity since then, and I think that will continue. It doesn’t mean it’s not gonna come without political strife, economic strife, all these things that are real, and there are real concerns, and we should take them very seriously. So I’m neither in the dismiss-it-and-age-of-abundance camp, but I’m not a jobs doomer either. I think we just have to be realistic that we’re probably going through a period of high churn in the economy where the net job number will probably go up, and revenue growth will probably accelerate, and GDP growth will probably accelerate. But people are gonna have to potentially do different jobs, and I don’t know what those are, but they always seem to pop up historically.
Jack: It’s funny, one of our guests pointed out the computer was a human job at one point. And if you told those people that this machine is gonna do the stuff you do, they’d probably say, “That’ll never happen.” So it’s hard to know.
Dom: It’s really hard to know, and it’s not any more comforting to anybody whose job could be displaced by AI, to say, “Oh, well, what are you gonna do?” I don’t know. But I am confident that this is gonna result... Confident is a strong word. I think this should result in major productivity growth for the economy, which should result in GDP growth acceleration, and historically, GDP growth acceleration is great for everybody. It’s when you don’t grow, that’s dangerous, right? ‘Cause then you start thinking about not how do we grow the pie for everybody, but how do we divide up what we have, right? And it doesn’t mean that you can’t think about that already. But growing the pie is clearly the best way for everybody.
Jack: One of the things I’m jealous of you about is that I’m a quant investor, so I’m running models all day, and you’re getting to look at one of the biggest technological revolutions in the world and decide how to build a portfolio with that. And so my question, I guess, is I wonder at a high level, how do you think about that? I mean, there’s so many different things you could be investing in. How do you think about creating a portfolio for the fund at a high level? What are you looking for? That’s a very broad question, but just in general, how are you thinking about constructing?
Dom: Well, look, I actually think that historically, quantitative portfolio managers have done a better job of portfolio construction than traditional portfolio managers. So we have an amazing quant team internally, and I try to learn from them on portfolio construction, risk management, right? My beta exposures, my momentum exposures, all my different factors. How do I build a portfolio?
Jack: And where we struggle is trying to invest during these productivity revolutions because we’re looking at the past fundamentals, and it’s like they’re not gonna tell us what’s gonna happen in the future.
Dom: I mean, some of my best ideas screen terribly on my little quant Codex agents, right? And they’re my best ideas. And so that’s where alpha lies, I think, for stock selection.
Look, how do I think about portfolio construction? The global technology strategy was one of the top-performing strategies in the country from 2010 to 2020, right? Had a terrible ‘22—interest rates went from zero percent to four percent over that period and had way too much software, and we saw a meaningful correction.
And so one thing that I wanted to do was make sure that my strategy was really the easy button in tech, right? And what do I mean by that? We’re gonna think about semiconductors versus software, we’re gonna think about internet versus cloud, we’re gonna think versus fintech. We’re gonna take all that into account. We’re gonna put together a strategy that is an easy buy-in, and it’s not an AI strategy. It’s not a semiconductor strategy. It’s a strategy that has a framework where I look for stocks with linchpin technologies innovating in secular growth markets with improving fundamentals at reasonable valuations, to find bottom-up stock selection ideas, surface them to the top, and then think about overall portfolio construction on top of that, right?
And so how much semis to have versus internet? That’s a tough question. That’s probably the thing I struggle with right now, and then if you look back the past few years, I’ve had a lot of semis, and I’ve had a lot less software, and the software that I have had has primarily been in this infrastructure software space.
And so you’ve gotta really think about how to structure it so that you could go through periods of July, or April of this year, or April of last year, and weather these downturns, right? And so that’s when you take into account your factors, right? Your beta, your momentum. You really try to understand these things.
And then the thing I talk about with my risk team all the time is, yes, that is the risk I’m taking, and that is the bet that I want to be making. So making sure that there’s no unintended bets in the portfolio, and that something’s not surfacing that I’m not aware of, right? I’m very happy to take bets. That’s how you generate alpha, is you take bets. But you wanna make sure you avoid the unintended bets.
You know, Rumsfeld got so much flack for the quote, “There’s the known knowns, the known unknowns, but then there’s the unknown unknowns, and those are the ones that get you.” It’s the same with risk factors. There’s the known knowns, there’s the known unknowns. There’s stuff you know you don’t know. But it’s the stuff that you don’t know that you don’t know when it comes to portfolio construction and risk management—those are the ones that get you. And so always try to be on the alert for those and get those into the second bucket at least.
Jack: So do you think more about the weighting of the buckets than the individual names? Do you think that’s more important in terms of construction, getting those things right?
Dom: I think a lot about factor construction, and I think a lot about sub-sector bets. If you had to boil it down, those are the two. Because on individual names—okay, so there’s different types of semiconductor beta. There’s memory beta, there’s logic beta, there’s high-quality beta. I mean, take the difference of an NVIDIA versus a SanDisk. They’re radically different stocks even though they’re both semiconductors, right? And so there’s different types of volatility in each individual name, and I think trying to understand that. And then, frankly, using performance as a guide to understand where your risk factors are is actually a really important part of that portfolio construction as well.
Jack: One of the interesting things about your fund is it’s a global fund, and those of us in the US, we tend to just think about the US companies. So how would you characterize the opportunity, or what you’re seeing outside of the US in tech right now?
Dom: Yeah. So both strategies, at any given point, may be 70 to 80% US and 20 to 30% OUS. I think there’s great—you know, my framework, linchpin technologies innovating in secular growth markets with improving fundamentals at reasonable valuations. What is a linchpin technology? It’s mission-critical to the success of its customers. I mean, ASML is a clear linchpin example, right? TSMC is a clear linchpin example.
Innovating in these secular growth markets—are you taking share in fast-growing end markets? Do you have improving fundamentals? Do you have revenue that’s accelerating, operating margins that are expanding, free cash flow conversion that’s improving? That’s the one that got the hyperscalers last year, right? Revenue was accelerating, but free cash flow conversion was going down. And then do you have a reasonable valuation?
And then often the market will dare you to buy stocks ‘cause it will have a few of the factors, but not all of them, right? And so memory is a great example. Fundamentals are decelerating. Revenue growth has to grow less from here, right? You can’t grow three hundred and fifty percent year over year forever. And they’re trading at low single digit or mid-single digit PEs, right? So that’s the market daring you to make a choice between valuation and fundamentals.
So put those all together, and that’s how you try to pick great stocks. And if those stocks happen to be in Japan or in the Netherlands or in the US, we’ll go find them. And that’s the team of twenty-plus people all around the world trying to find great names and surface great ideas.
Jack: Well, this has been awesome. Kai and I could go on for hours with this, but we value your time, so we won’t do that. But we do have two standard closing questions we ask all of our guests. The first is, what’s one thing you believe about investing that most of your peers would disagree with?
Dom: You know, that’s a great question. It’s a hard one ‘cause I do think some of my peers do believe in this, but I think the power of reflexive cycles is underestimated. And so that’s why I said in July I was worried about what does it mean for the capital markets. I think we may be going through one of the greatest reflexive cycles of all time. And so when people ask me what book to read in stock picking, you have to read The Alchemy of Finance, right? Where Soros lays out reflexivity.
Jack: And our last question is, based on your experience in markets, if you could teach one lesson to the average investor, what would it be?
Dom: You know that graph where it’s the 80/20 graph, and it’s—they call it the midwit graph, right? The dumb guy, the average person who thinks they’re smart, and then the really smart guy. And the dumb guy and the really smart guy think the same thing, and then the middle guy thinks some very complex thought. It would just be, buy accelerating fundamentals. Right? And I think really smart investors—and not dumb investors, but you know, if you’re simple about it, buy accelerating fundamentals is usually a great place to hunt.
Jack: Well, there’s hope for dumb guys like me out there.
Dom: You know, what did Buffett always say? “You don’t need the highest IQ to be in this business. You need a high enough IQ.” And I always liked that.
Jack: Well, Dom, thank you very much. This has been awesome. We appreciate your time.
Dom: Thank you, guys.