Hugo Scott-Gall: Today I'm delighted to have with me Jay Kannan. Jay is one of our tech analysts. He covers hardware, which is basically semis, which is chips, semi-cap equipment. So the machines that make chips, telcos, which is a telecom company, which is the least interesting area of your coverage, I would argue. And he's coming on the show because he was on the show in 2021. He enjoyed it so much. He's wanted to come back since then, but we haven't asked him until now. Jay, welcome to the show.
Jayesh Kannan: Thank you Hugo, it's wonderful to be back.
Hugo Scott-Gall: Great. Right. We've got lots to talk about. As I said, when you came on the show back in 2021 at the depths of COVID, our conversation was very different to what I imagine we're going to discuss today. We talked a lot about intangible weightless. Today, I think we're going to talk weighty topics, but literally about heavy things. Do you agree?
Jayesh Kannan: I would say Hugo, the contrast could not be more stark. In 2021, I remember when we last had this conversation, it was the peak of COVID-driven business models. Software, SaaS, work from home, e-commerce. In fact, our own discussion was centered around connected commerce. Today, arguably, as you put it, tech has a lot more weight. So the same dollar of tech revenue today, pulls materially more physical, capex infrastructure capital than it ever did. And what does that mean? Today, technology is about concrete, copper, water, gas turbines, substrates, even electricians. Essentially the discussion, the investing framework, the mental models have moved away from the consumer interface to the backend of the backend.
Hugo Scott-Gall: Great. So that is going to be the bulk of our conversation about what is going on now, because what's going on now really is quite spectacular. But before we get there, let's talk about, you didn't come on with a whole series of predictions. However, we did talk about the future and what you thought was going to happen. What did you get right? What did you get wrong?
Jayesh Kannan: Let's start with what panned out as we expected. I'd say the most critical part of that is digital adoption. So just digital tools, the virality of adoption both in society, at the consumer level, and at the enterprise level. This was structural, not just a COVID pull forward. The second was also this long growth runway for cloud. So as more and more data infrastructure moves from the server rack in the back room to the cloud.
Now I'll talk a little bit about what surprised me, what I underestimated, what was not in my thinking. There were three things here. One, I would say the speed, the pace at which compute itself became the binding constraint. In 2021, we were discussing attention, real estate, last mile capabilities. None of that is binding today. Today it is compute. Second is AI. We can go back and look at the transcript, I would think we barely discussed AI in 2021. Today, it is not just critical, it is central to every tech discussion and investing model. And third, I would say is how physical this answer has become. Back then, this was a software conversation. Today, this is about power, substrate, packaging, and the supporting infrastructure.
Hugo Scott-Gall: So that was a very honest and some ways quite humble framing of what we discussed. And look, you know, we both got stuff wrong. So the most important thing I think we could have said in 2021 is that there's going to be this huge moment in 2022 where Chat GPT is the proof case for essentially the next wave of AI, which had the direct results of demand for GPUs, graphic processing units, going technical term, nuts. And so that would have been the most important insight. So I guess the question then on Where is growth? is just as important today. So where do you think, give you a timeframe, next two years, where is growth?
Jayesh Kannan: Let's start here, quite literally, the cloud has now come back to earth and that's what's changed in these 5 years. In 2021, the metric for discussing or even framing growth was centered around subscriptions, ad impressions, active users. Today, the most important metric in my mind is tokens. And tokens at its primacy is the unit of cognitive work. And tokens aren't replacing attention, they're replacing intelligence and thinking. So this is a completely different category with different TAM ceilings. So even understanding how the revenue pool works, the winners and losers is completely different.
I would say this, the most aggressive TAM estimate on AI is likely underestimating the true potential of AI in my humble assessment. It's bigger than we believe and there are many ways to frame it. Here's a simple example. Global software spent today is a little over a trillion dollars. 5% of global labour, the labour pool in the world is one and a half trillion dollars. If we now believe that some of this gets replaced by AI, we're already discussing a few trillion, not many hundreds of billions of dollars and that is where growth is.
Hugo Scott-Gall: You would say, I'm putting words in your mouth here, but we're well beyond the kind of will AI work? Are there real meaningful applications? We're way beyond that. From your point of view, you talked about tokens, which is really demand for compute power, demand for compute intelligence. That demand is so strong that it puts to rest the idea that it's a will it work? Will it work? Word it work? We're way beyond that. This thing works. Demand is super strong. Am I misrepresenting you there?
Jayesh Kannan: I would go a step further Hugo and say we are at the 3G moment for tokens today. So this takes me back 10 or 15 years to when we had new cycles of mobile communication technology. With 2G it was about scarcity, with 3G it was about expensive abundance. By the time we came to 4G and 5G, it became a utility. So we started using data similar to drinking water. It was not a need to have it was a must have. And we're seeing that happen today with tokens.
Here's the framing around it. I give you my own sense. In the past month, I've busted my token limit on the AI tool that we use at work three times. And this is just a few weeks after we've rolled it out across the organization. So that can give you a sense of where this is going. Now, it's not just tokens helping me do my job better. The discussion we're having as a team is how do we think about token consumption, the costs associated with it and the balance that we have with headcount addition and new headcount growth.
Hugo Scott-Gall: So you would say you are living proof of this works. You talked about you busting your token limit. You as an individual going about your job, your demand to use this exceeds supply. You think, you know, scale it up across hundreds of thousands of people that certainly in the professional services, knowledge economy, you think demand is well in excess of supply.
Jayesh Kannan: Demand is well in excess of supply.
Hugo Scott-Gall: Because there are clear tangible outcomes. Wins.
Jayesh Kannan: Demand habit creation is addictive and there is diffusion both through formal channels as well as informally across the organization. So more and more companies, industries, applications embedding generative AI at the heart of how they provide their own service. So this is a genie that's now out of the bottle.
Hugo Scott-Gall: So we've got lots of demand for tokens more than we can meet because why can't we meet it with supply constraint? Let's talk about those supply constraints. What are they? How do they get resolved?
Jayesh Kannan: Let's perhaps start here Hugo, tokens in my mind are the new barrels of oil. It's just that production capacity today has an 18-month lead time and the refinery is in Sinchu in Taiwan. Now why can't supply keep up and what is different relative to prior tech cycles?
Let's start with where the bottlenecks are and how this is changing over time. So typically one might have said let's just build more fabs which are semiconductor factories and have more of these manufactured in quick time. That was the issue we had in 2022 because a large part of the bottleneck were the GPUs or the actual chips that are produced. Since then, year after year, the bottlenecks have migrated. So we move from chips to memory to other supporting infrastructure. So packaging and advanced techniques that are used to bring all of this together.
Today, the bottleneck is off silicon. So the bottleneck are these other upstream backend physical components. And that's substrate, that's glass, that's copper. Increasingly, that's also supporting infrastructure. Power, water, permitting, the technicians required to fire up and maintain the factories. That's where the supply constraint lies. And these are harder to overcome in quick time.
Hugo Scott-Gall: Yeah. So it's always illustrative, indeed instructive, to look at history for these cyclical industries. Usually, the best cure for high prices is high prices. These industries will commit a lot of capex to increase supply, eventually supply overwhelms demand, price goes down. Saying this time it's different, it's a cliche in our industry, things are never truly different.
But what you're saying and what I know your view is, is that this tightness can persist really for quite some time, which is why if you look at the earnings of the makers of memory chips, they are forecast to be at record levels well beyond anything seen before. Again, I always like to kind of ask questions this way. Am I misrepresenting? Is this your view? Because this is something, the ending may not be different, but the duration and quantum, the size of this is different. This is, is this the mother of all supply squeezes on the supply chain to make chips?
Jayesh Kannan: In the last 15 years or so, when we discussed pricing as it related to physical components and upstream components, the framing was always centered around the level of price erosion over time. This is the first time in nearly two or three decades that we are discussing price premium, the persistence and magnitude and duration of that price premium relative to price erosion.
To give you some examples, the grid interconnect lead times for new power supply is 5-7 years in some parts of America. To get an advanced lithography tool in order to fabricate semiconductor chips at the leading edge is nearly an 18-month lead time. Gas turbines have over 5-year lead times. So of course, as you rightly pointed out, supply would continue to add capacity. And we would see this compress at some point in time. But for now, they are here to stay.
The other point I'll make here is two weeks ago, I had the chance to visit a large fabrication facility that makes semiconductor chips in Arizona. Hugo, the scale, the ambition, and the enormity of that project is very impressive, and it's highly incredible. I would say it's staggering. Like you understand viscerally when you look at the cluster of over a thousand acres of land where concrete is still curing, transformers are being craned in, miles of cables are being run, and you can see that still supply cannot keep up with demand. So they're building, they're acquiring new land, and they're doing all of this because the last point I'll add here is geographic lock-in is now becoming near permanent. So it's not a global demand global supply model that we spoke about when we had the age of globalization 20 years ago. Today it's local supply that meets local demand.
Hugo Scott-Gall: So what does that mean therefore when you think about, again, we don't know how strong demand really is because we need to have as much supply as you want and also there's price as well. But how quickly do you think supply can respond to ensure that demand is met?
Is this a kind of it feels like a long time, but it's all going to get done within a year. Some of the numbers you're throwing around around these physical challenges are going to take a lot longer. It's difficult to double the amount of gas turbines being built in the world in a year, being made in a year.
So is this a persistent tightness that means that those who are well positioned around bottlenecks can be much more profitable than historically? And the cure says bottlenecks aren't coming anytime soon which is another way of saying this is a different cycle, it really is stronger and it is for longer.
Jayesh Kannan: Combine everything you said on the supply side and add to that, that demand growth is explosive. So assume demand is linear, then we can have all of this framing around supply catching up. But token growth is at the magnitude of 10 to 15 times today. And it's increasing month by month, as we can see in some of the disclosures from the frontier labs.
So if you combine that explosive growth, which is not linear but rather exponential, along with the long lead time it takes not just for intent of supply but for actual supply to start meeting that demand, I would argue that the cycle time hasn't compressed it's likely expanded and this is going to likely persist especially in some chronic pockets for multiple number of years
Hugo Scott-Gall: Look, as investors, we're always looking for scarcity. Scarcity becoming abundant is something we want to avoid, but unless we are involved with the problem solvers of scarcity. So is that how you see it at the moment? There's elegant dance or this elegant tension within the equity markets that beneficiaries of tightness versus problem solvers to tightness. You can make money on both sides, but when one goes from the other, as in once something that was tight, suddenly really opens up, that's not a good time to be certainly in the history of investing in your in your sector.
Jayesh Kannan: Since you like history, Hugo, here's a good lesson that I was thinking about as it relates to the history of investing and the supply-demand tightness. In every prior tech cycle, it's the people who sold the jeans that outlasted most of the miners. It takes me back to the 1849 gold rush. Today, the jeans, if I may, are made in Taiwan, the shovels are made in the Netherlands, and the picks come from Japan.
What I'm trying to say here is that if we think about where is value being captured today, which is another version of the question you asked, I think about three layers. I think about software, where I'll include the applications, the large language models. Secondly, or below that is the hyperscalers. And then the final layer are the enablers. So, semiconductors, equipment, so on and so forth.
Software is where the headlines are today, the venture money is today. It likely may have the most upside, but it has the most fragile economics. So it's incredibly hard, if not impossible from my seat and ours to pick winners ex ante.
Let's go to the next layer, which is hyperscalers. Here's where the capital is today. The capital is being deployed. The scale advantages are real. But the way I see it, unit economics are still emerging. So there is likelihood that there will be differentiation here at the top, but workflow gravity will win. So they'll get commoditized towards the bottom within that layer.
Finally, we come to the infrastructure layer. And as I said in my model, infrastructure is likely the most reliable long-term value creator. Think about railroads, electrification, internet, mobile. It's effectively the how much problem that's more underwritable than the application problem. So, it’s a less glamorous business models but more cash generative, easier to underwrite.
So, where does value go from here or where do we go? It likely goes down the stack in the near term and like you pointed out, eventually goes up the stack but we have a while to get there.
Hugo Scott-Gall: Have you thought much about the LLMs, the model providers? So this is the likes of OpenAI, Anthropic, obviously Gemini from Google. How important is it to have the best model? How expensive is it to create the best model, operate the best model? If everyone has a model, is there real genuine value capture? And I think we need to differentiate between models that are facing consumers and models that are facing enterprises. You would have said in the late 1990s B2C and B2B.
Jayesh Kannan: I’d say there are a multitude of models today and you classified it by both capability, by consumer end market, I would probably add geography as well. There are the Western models and then there are the Chinese models. What is clear across all of these models is the aggregate virality of adoption. So there's no doubt that models are being used.
Models are expensive to train especially at the frontier level. So the newest model is the hardest, the most expensive to train and that is where monetization frontier lies. So my view is that a lot of the legacy models, so think lower capability versions of models across providers in time will get commoditized. But where they will make most of their revenue dollars is at the leading edge. The newest model that they come out with. Of course, the other competitive advantages that we often think about and frame when we discuss software and services come to play here as well.
So whether it be the install base, whether it be the route to market and the way they get to their customer, embedded in enterprise applications, think about all the tools we use today in the ecosystem. And of course, about the go to market of these models as well. So whether they're focused on a particular type of use case will determine which of these win, lose.
The punchline in my view is there will likely be consolidation in the future of these models. There'll be fewer of them that will scale, but there will be multiple of them. It will not be winner take all.
Hugo Scott-Gall: People say software is dead. Anthropic, Open AI, Deep Seek in China, are showing you that software is not dead. It's just that the value is shifting away from legacy to new entrant software. Plus there's clearly, if you think about potential revenue projections for something like an Anthropic provides Claude, there's going to be some revenue gain from labor as in, it will replace humans.
Does that argue for these companies eventually becoming a bit like sort of hyperscalers? By that I mean, Google is quite a vertically integrated company. It owns a cloud business, has its own chip, it has its own LLM also has search, it also has YouTube. So do you think these new entry companies are going to up looking quite like existing legacy tech companies or will they do it their own way?
Jayesh Kannan: The answer probably lies somewhere in between. So a lot of these new companies in various shapes and forms already have connections, if not stronger linkages with the legacy incumbents. And if there's one thing that the legacy incumbents have learned over the last 15 years is not to make it us versus them, but to get involved, to partner, to become a shareholder, to become a customer, to become a supplier to these startup companies, if I may so refer to them, at the very early stage.
So all of these companies, in fact across and in each of the model providers today has multiple connections. They're not just beholden to one of the Mag 3, 4, 5, 6 companies. They often have agreements and as shareholders, many of them. So I would say that we go into a world where there's a lot more complexity and it's not one for one. So more companies are tied to more incumbents and there are more connections throughout this ecosystem. Some of them are formal under the same company and domicile and some of them more informal and diffused.
Hugo Scott-Gall: Let's flip it around. What really isn't working in tech? By what isn’t working? I mean business models that are under pressure, possibly terminal pressure.
Jayesh Kannan: Here are the two axis of bifurcation Hugo. One is along the hardware semiconductor line and my punchline there is that data center demand and components are crowding out the rest of hardware infrastructure. Namely, at the smartphone, the PC and at the consumer end. So you mentioned memory prices earlier. Now that's causing a lot of smartphone PC demand costs to rise to the extent that it's created demand destruction. So on one hand, it's the consumer device and the consumer end market that will suffer and that's here to stay.
The other axis is around services. For the last 25 years, outsourcing of technology services was the world's largest beneficiary of labour arbitrage. The premise was pay less for the same work because you can get the work done by moving it to a lower cost offshore location. Large language models and AI change that unit of arbitrage. We are not arbitraging geography anymore. We are arbitraging humans against models. And the model is getting cheaper by a magnitude of 60% every year, while the human is not. So this is not a cyclical problem for the industry. This is an existential one, because the entire revenue dollars was built on geographic labor arbitrage.
Hugo Scott-Gall: Here's the question that has been bubbling under the surface throughout all of this. I'm going to ask you, all right, do think we are in a bubble?
Jayesh Kannan: It's hard to say. I would say that I'll
Hugo Scott-Gall: That's the best answer you could.
Jayesh Kannan: Or let me say it with even more humility: I don't know. There are pockets which would suggest that demand outstrips supply. The fundamental characteristics that we evaluate in research appear extremely strong. But at the same time, there are pockets of the equity market where near-term multiples of securities are extremely extended, which suggests early signs of froth.
What I can say though is two things. One, typically a lot of bubbles are stories, narratives, call it what you may, without cash flows. This one has meaningful cash flows that are being generated today and will likely grow in the future.
The second thing I would say Hugo is my view on this has also changed in the last many months. I would say six months ago, I was thinking a fair bit about AI and my read was this is extraordinary technology, real engagement, but the unit economics appear a little bit hand wavy.
Today, my confidence around the monetization pathway on model capability on improvement in that capability and the economic value of adoption is compounding faster than what I believe the market is pricing in. And that I would say is, and we are at the tip of the iceberg, so a lot more here to go. And that's what's changed in the last few months.
Hugo Scott-Gall: It definitely seems that way, doesn't it? If you think about demand, it's much stronger than I think we as a team thought six, nine months ago. And why is that? More applications and certainly from one of the new entrants, better monetization, and also just these really, really tight bottlenecks that are resulting in almost parabolic pricing for some components, some parts of the supply chain. And that definitely as you said, that is generating real dollars and cashflows.
It's interesting to extrapolate this and think what does it mean for inflation? What does it mean if the price of these components in the economy go up, that will come through in end prices, but we can leave that for another day. One more thing before we go, I try and weave in robotics into every podcast. I do often fail, sometimes succeed, but we didn't really talk about physical AI. A lot of what we were talking about was digital AI.
There are some that argue the physical AI will end up being a larger TAM, total addressable market, than digital AI. Is that really a factor of the bubble question? Sure, things can get frothy and stocks can be overvalued, but the underlying kind of earnings power or economic value created can power on. Is there another, not leg, because it’s a lot more than a leg, another big chapter in this coming from physical A.I. and robots.
Jayesh Kannan: Yes, absolutely. The question is not about whether physical AI is real. The question is how large could this be?
Hugo Scott-Gall: Yeah, we don't have to talk about robots. I just wanted to keep saying robots.
Jayesh Kannan: And here's, this is compute math, this is not estimates, but if we believe that there are $50 of components that go into the smartphone, $5,000 of the same type of components go into a humanoid robot.
And the second thing I would say is that the supply chain here has already been paid for and is already building out capacity. They were doing it for consumer devices, they're now doing it for AI data centers. Doing it for humanites is an extension of that. And we don't have to be right or precise about when we will have how many robots.
So will it be 10 million robots by 2030, for example. Could be fewer, could be more. But the answer is that every robot that is created will need two orders of magnitude more of the physical components that we just discussed and to our discussion from earlier, this will create an even more chronic supply shortage and supply demand gap.
Hugo Scott-Gall: Yeah. Great. Jay, let's end it there.
Jayesh Kannan: Thank you.
Hugo Scott-Gall: Thanks for coming back on the show. Will you come back on in another five years time?
Jayesh Kannan: I hope it's sooner than that. We have many more things to discuss.
Hugo Scott-Gall: We do. We do. Thank you very much.
This content is for informational and educational purposes only, and is not intended as investment advice or a recommendation to buy or sell any security or to adopt any investment strategy. Investment advice and recommendations can be provided only after careful consideration of investors’ objectives, guidelines, and restrictions. The views and opinions expressed are those of the speakers as of the date of this recording. Are subject to change without notice as economic and market conditions dictate and may not reflect the views and opinions of other investment teams within William Blair Investment Management.
Factual information has been obtained from sources we believe to be reliable, but its accuracy, completeness, or interpretation cannot be guaranteed. Any discussion of particular topics is not meant to be comprehensive and may be subject to change. This material may include forecasts, estimates, outlooks, projections, and other forward-looking statements. Due to a variety of factors, actual events may differ significantly from those presented.
Past performance is not indicative of future results. Investing involves risk, including the possible loss of principal. Any investment or strategy mentioned herein may not be suitable for every investor. References to specific companies are for illustrative purposes only and should not be construed as investment advice or a recommendation to buy or sell any security. William Blair Investment Management may or may not own any securities of the companies referenced. It should not be assumed that any investment in the companies referenced was or will be profitable.
