Five years after his last appearance, Jay Kannan, CFA, a research analyst on William Blair’s global equity team, returns to The Active Share to explore how technology has fundamentally transformed from software and services to infrastructure and hardware. Where Hugo and Jay’s 2021 conversation centered on e-commerce and cloud adoption, their discussion this time around focuses on the physical constraints powering AI—from semiconductor bottlenecks and substrate shortages to power grids and the cascading supply chain effects. Together, Hugo and Jay examine why they believe demand for compute far exceeds supply, where value is captured across the stack, and what's breaking under the weight of explosive token growth.
Comments are edited excerpts from our podcast, which you can listen to in full below.
Last time you were on the show, we discussed intangible and weightless assets. Today, I suspect our conversation is going to be heavier. Would you agree?
Jay Kannan: The contrast is stark. In 2021, at the peak of COVID-driven business models, everything was about software, SaaS, work from home, and e-commerce. Even our own discussion was centered around connected commerce. But today, technology carries a lot more weight. The same dollar of tech revenue now pulls in more physical, capital expenditure (capex)-heavy infrastructure than it ever has before.
We also discussed the future and what you thought was on the horizon. What did you get right, and what did you get wrong?
Jay: Let’s start with what panned out. First, digital adoption, and the virality of it, both at the consumer and enterprise level. It was structural, not just a COVID pull-forward. Second, the long growth runway for the cloud, as data infrastructure continued migrating from the backroom server rack to the cloud.
But there were other things I underestimated: the speed at which compute became a binding constraint; AI, which is now central to every tech discussion and investing framework today; and the physicality of technology. The conversation is now about power, substrate, packaging, and the infrastructure behind it all.
Where do you see growth over the next few years?
Jay: The cloud has come back to Earth. In 2021, the metrics that defined growth were subscriptions, ad impressions, and active users. Today, the most important metric in my mind is tokens.
At its core, a token is a unit of cognitive work. But tokens aren’t replacing attention; they’re replacing intelligence and thinking, which is an entirely different category, with different total addressable market (TAM) ceilings, different winners and losers, and a different revenue pool.
I think that even the most aggressive TAM estimate for AI is probably still too low. Global software spend today is just over a trillion dollars, while 5% of the global labor pool is around $1 trillion to $1.5 trillion. If AI begins displacing even a portion of that, we’re looking at not hundreds of billions of dollars, but a likely a few trillion dollars, of revenue growth.
So, what you’re saying is the demand for tokens, or the demand for compute power, is seemingly so strong that it puts to rest any question of whether this works.
Jay: Yes. We’re at the 3G moment for tokens. I think back to the ramp-up of mobile technology 10 to 15 years ago. With 2G, data was scarce. With 3G, it was abundant but expensive. By 4G and 5G, it became a utility. We consumed data the way we drink water, and that’s what’s happening with tokens.
For example, in the past month, I’ve hit my token limit on our AI work tool three times, just weeks after it was rolled out across the organization. Now, it’s not just tokens helping me do my job better. But that can give you a sense of where this is going.
Today, the most important metric in my mind is tokens.
Demand for tokens is beginning to exceed supply. Why can’t supply keep up, and what are the constraints?
Jay: Tokens are like barrels of oil, except the production capacity has an 18-month lead time, and the refinery is in Hsinchu, Taiwan.
What’s different from prior tech cycles is that bottlenecks have migrated year by year. In 2022, it was the graphics processing units (GPUs). Since then, it’s moved to memory, then to packaging, and then on to the advanced techniques needed to bring it all together.
Today, the constraint is the upstream physical components (such as substrate, glass, and copper) and increasingly, the supporting infrastructure around it (such as power, water, permitting, and factory technicians). These issues are harder to solve quickly.
History suggests the best cure for high prices is high prices—industries commit capex, supply eventually overwhelms demand, and prices eventually fall. But in your view, this tightness can persist for quite some time. Is this the mother of all supply squeezes?
Jay: For the last 15 years, every conversation about physical and upstream components was framed around price erosion, but this is the first time in two or three decades that we’re talking about the persistence, magnitude, and duration of price premiums.
For example, grid interconnect lead times for new power supply are 5 to 7 years in parts of the United States, while an advanced lithography tool for leading-edge chip fabrication is an 18-month wait. Even gas turbines have a wait time of more than 5 years. Supply will eventually catch up, but not anytime soon.
Recently, I visited a large semiconductor fabrication facility in Arizona. When I stood there looking at more than a thousand acres, where concrete was still curing, transformers were being craned in, and miles of cable were being run, I viscerally felt that supply could still not keep up. This is no longer a “global supply meets global demand” model. Rather, it’s evolved into a “local supply needs to serve local demand” model.
Is this a persistent tightness, where those well positioned around bottlenecks stay more profitable for longer?
Jay: If demand were linear, the supply catch-up framing might hold. But token growth is running 10x to 15x today and accelerating month by month. Combine that exponential demand with the long lead times required for supply to come online, and the cycle time hasn't compressed but likely expanded. I think this tightness is likely going to persist, in some pockets chronically, for multiple years.
As investors, we’re always looking for scarcity. But there seems to be tension in the markets right now between the beneficiaries of tightness and the solvers of tightness.
Jay: There’s an investing principle often attributed to Mark Twain I think about when it comes to supply and demand: during the 1849 California Gold Rush, while prospectors rushed to find gold, the people who became wealthy were often those who sold supplies to the miners—picks, shovels, jeans, food, and equipment. The actual miners frequently went broke or found little gold.
Today, the jeans are made in Taiwan, the shovels in the Netherlands, and the picks in Japan. Where is value actually being captured? I like to think about this question in three layers.
The first layer is software, which includes applications and large language models (LLMs). That’s where the headlines are, where the venture capital money is, and where the highest potential for upside is. But software also has the most fragile economics, and from where I sit, picking winners in advance can be difficult.
The second layer is the hyperscalers, which is where capital is being deployed today. The scale advantages are real, but unit economics are still emerging. I think there'll be differentiation at the top, but workflow gravity will likely push commoditization toward the bottom of that layer.
The third layer is infrastructure, which includes semiconductors and equipment. In my view, this is the most reliable long-term value creator (think railroads, electrification, the internet, and mobile). It's a “how much?” problem rather than a “which application wins?” problem. It might be less glamorous, but infrastructure tends to be more cash-generative and easier to underwrite.
In my view, infrastructure is the most reliable long-term value creator.
Does having the best LLM matter? If every major player has one, where does real value get captured?
Jay: There are a multitude of LLMs today, and you can organize them by capability, market, and geography. What's clear across all LLMs is the aggregate virality of adoption. LLMs are no doubt being used, but they’re also expensive to train. The newest, most capable LLM is typically the hardest and most expensive to build, but that's where revenue is being generated.
In addition, the usual competitive advantages apply to these LLMs: installed base, route to market, deep embedding in enterprise applications and everyday tools, and whether they’re focused on a particular use-case or going broad. All of that will determine who wins and who loses.
While consolidation is likely to occur, it won’t be winner take all: there will be multiple survivors, just not as many as there are today.
As the value of software shifts from legacy incumbents to new entrants, do these new entrants eventually start to look like the giants they're displacing?
Jay: The answer probably lies somewhere in between. Many of these new companies already have deep connections, if not formal linkages, with legacy incumbents. And one thing the incumbents have learned over the last 15 years is not to make it us versus them. Instead, they get involved early as partners, shareholders, customers, or suppliers. Every one of the major LLM providers has multiple connections across the ecosystem. They’re not beholden to any single company and often have agreements with, and shareholders among, several of them.
So, what we’re moving toward is a world of much greater complexity. There are more companies tied to more incumbents throughout the ecosystem. It’s not a one-to-one mapping; it’s a web.
What isn't working in technology right now?
Jay: There are two axes of bifurcation. The first axis is along the hardware and semiconductor line. The dynamic here is that data center demand is crowding out the rest of hardware infrastructure, and the cost for consumer devices such as smartphones and PCs is increasing to the point of demand destruction.
The second axis is services. For the last 25 years, outsourcing technology services have been built on one premise: geographic labor arbitrage, or paying less for the same work by moving it to a lower-cost offshore location.
LLMs blew up that model entirely. We’re no longer arbitraging geography. We’re arbitraging humans against LLMs. And LLMs are getting cheaper by roughly 60% every year, while the cost of human labor isn’t. This isn’t a cyclical problem but an existential one because the entire revenue base was built on an arbitrage that no longer exists.
We’re no longer arbitraging geography. We’re arbitraging humans against LLMs.
Do you think we’re in an AI bubble?
Jay: Honestly, I don’t know. There are pockets of the equity market where demand clearly outstrips supply, and fundamentals look extremely strong. But there are also pockets where near-term multiples are extremely extended and show early signs of froth.
What I can say is most bubbles are built on stories and narratives without cash flows. The market demand for AI, however, has begun to generate meaningful cash flows, with more to come.
I also used to think the unit economics of AI felt a little tenuous. But today, my confidence in the monetization pathway, in model capability improvement, and in the compounding economic value of adoption is growing faster than what I believe the market is pricing in.
We've talked mostly about digital AI, but some argue physical AI will ultimately be the bigger opportunity. Could robotics be the next chapter?
Jay: A smartphone has roughly $50 worth of components; a humanoid robot has around $5,000 worth of the same types of components. That's two orders of magnitude more, with a supply chain to build those components already paid for. It was first built out for consumer devices, then expanded for AI data centers. Scaling it further for robotics is an extension of that same buildout. But as we said earlier, that will likely only deepen an already chronic supply-demand gap.
