Alex Sacerdote on S-Curves, the AI Boom, and Finding Technology Winners

Alex Sacerdote

Yeah. Well, when the gun went off with OpenAI and ChatGPT in November 2022, we immediately took the firm and did a massive deep dive with our ten-person team. Anytime you have a new compute paradigm, there's a new stack, and that creates new winners and losers on the old stack. In this stack — Jensen talks a lot about it now — it's power at the bottom, chips at the bottom, then the clouds, then the foundational models, then the applications on top. At that time, early 2023, we said we wanted to be in the chips and the infrastructure first. Not only do they get the demand first, but we knew who the winners were. And no matter who won above — which we weren't sure of at the time — we knew we were going to need tremendous amounts of compute.

We did a deep dive into that, which we can talk about later, but over the next two or three years we started to get more clarity on how the foundational-model layer would evolve. Two or three years ago there were sixty different companies going after it, and OpenAI was in the lead. We did a webinar in April 2023 where we said: look, this might be winner-take-all. It might be a total commodity because there are open-source players. It might be a race to zero, or it might be an oligopoly with three or four leading players. What we saw over the following three years was that almost all the startups fell away and died, and then some of the largest companies in the world got involved too, including Amazon and Meta. Amazon never really showed up. Meta came in strong, then their effort faltered and they had to do a total reboot.

In the meantime, Anthropic was this dark-horse candidate, the startup that focused purely on the enterprise, while OpenAI had won the consumer market — and Gemini can never be counted out; we love Google too, it's one of our largest positions. So it really started to look like a three-horse race, somewhat like an oligopoly, similar to how the cloud market evolved, where three companies underpin the entire SaaS cloud world and have excellent businesses. We were also aware of the open-source risk from China, and we got comfortable that the quality of the tokens from the leading edge was superior — if you're 80% of the way to the top of the benchmarks, going from 80 to 85 is a huge unlock, and the open-source players don't have as much compute, so they can come close to the leading edge but they can't leapfrog it, and then they falter. Meanwhile the scaling laws and other means of improving the models — the feedback loops and so on — meant there was a very strong runway, and everyone close to the industry believed the scaling laws would continue. So we developed the thesis that it would be a three-horse race.

Then the big kicker was code. This is the true unlock of AI. In the first few years we knew AI would be big, but we were skeptical — we made large investments because we knew the training spend would be there, but we weren't sure how much revenue would come, or whether it could truly replace labour, because the early versions of the models were good but there was a lot of negative feedback from corporates about whether they could be truly agentic. In 2025 the first Claude Code and the coding tools really began to explode. The first generation was like Microsoft Copilot, which was $20 a month and could improve your grammar of coding, maybe find a bug, maybe write a paragraph-sized block of code. Then Anthropic came out in the middle of the year and it could do so much more — it started to run agentically. The coding market exploded, and we started hearing that people who could use it unfettered — even within Anthropic — were spending $100 a day on tokens, which comes out to $20,000 or $30,000 a year. If you think about how many coders there are in the world, twenty million, you've got a half-a-trillion-dollar market from coding alone. And that was on seven-, eight-, nine-month-old technology. We could see, just on the coding market alone, that Anthropic had a tremendous opportunity ahead of it.

So — this is pretty funny — we wrote in our letter, when we made the investment at the $180 billion valuation, that we thought they were hoping to get to nine billion in revenue.

Alex Sacerdote

And then the numbers were like nothing we'd ever seen before — a hundred million to a billion on the way to nine billion. When we did it in August 2025, nobody had any idea what 2026 could be. The second big unlock lately is that Claude Code has gone to almost completely agentic. Andrej Karpathy and Linus Torvalds — two of the smartest people in coding — completely flip-flopped on this last year. Karpathy said last year's code tools could write 20% of code and 80% would be hand-written; that flipped when the latest model came out, and now he says he hasn't written a line of code except in English. Not to mention the pure unlock for people who never knew how to code in the first place. So coding alone has completely taken off, and Anthropic has been able to stay ahead in it.

One difference between the clouds — GCP, AWS — and the AI companies is that the clouds are generally commodity: they're selling you servers and storage. There's a lot of software on top and some stickiness, but everyone thought the AI models would be pure commodity too. There's actually tremendous differentiation within them — different training methods, different skills they're good at. A lot of people build routers that switch between models, which makes it sound like commodity, but Anthropic is very good for anything to do with private equity and finance, Google is very good for ingesting PDFs — there's a lot of differentiation, critical IP, which is a great competitive advantage. Many companies have come after the coding franchise, and Anthropic has kept ahead.

The other thing that's good about the foundational models, and Anthropic specifically, is that it isn't just the API or the model. They're building a whole ecosystem of products around the API — the SDK, Claude Code, the orchestration layer, and all the tools, what they call the "harness": the software around the API that gets the most out of the model. This is exactly what we saw with AWS very early on, in 2013 — people thought it was a commodity server in a warehouse, big deal, but they saw this was a new way of doing computing, and they invented all these products, ahead of everyone else, that slowly built in lock-in.

The other way we think about this is where we are on the S-curve. We think the infrastructure layer is something like 10% penetrated, and it's still one of the best ways to play AI — we'll talk about how that feeds back through. Even though 800 million or so people are using AI, they're mostly using "AI 1.0," which is like a search engine on steroids. But now, with new primitives — Claude on your computer, linking it in — you build skills, companies and individuals start building their own skills, then true AI bots, then big corporations build much larger systems. Sundar Pichai said it's 10 basis points of the world's knowledge workers. Anthropic has something like 14 or 15 million daily active users; probably a small portion of those are truly using AI the way you can. That 10 basis points is a classic S-curve: these are the tinkerers, and it moves to the early adopters, then the early majority. You're going to go from 10 basis points to 1–2%, to 3–5%, to 15% over the next four years. This year a light switch went off in the enterprise, where everybody realised they need to do this now, and fast.

Alex Sacerdote

And so we think the enterprise AI, or enterprise-application AI, market is less than 1% penetrated. We've never seen anything like it — we talk about S-curves; we call this an L-curve, just straight up. And then take this to the infrastructure side: we're at 10 basis points of people really using AI, and we're already sold out — there's not enough compute in the world. Anthropic has half of what it needs right now, and that's before this huge take-up. Marc Andreessen said that in the next four years the one thing he's sure of is there won't be enough compute.

Alex Sacerdote

In Anthropic's case, we got to know the company — one of our analysts knew people in the finance group there. We actually looked at the $60 billion round and didn't do it; we didn't know the company as well then, and the gross margins were negative, and frankly we hadn't yet seen coding explode the way it had. One thing about public markets is you get to know companies over a long period of time and can invest on your own schedule. I got a chance to spend some time with Dario Amodei — I'd obviously listened to him on podcasts — and I started to realise their management team was excellent: the focus, the dedication, almost no turnover, the quality of the code, and the business plan was really starting to play out. It's one thing to grow from 100 million to a billion, but it's another to do nine billion.

So we reached out to the company as much as we could, and they took a meeting with us. We did a ninety-page PowerPoint deck where we used Claude Code to scour the internet for all the feedback we could find about the coding market — what their products were good at, where they might need to improve — and we did our own overview of what the coding market would become. They welcomed us into the round, and we've stayed close with the CFO since. It's been great to build that relationship, and I think we punched above our weight in terms of allocation. That one was a total home run.

In the rest of the world, we're in a period where the unicorn market is bigger than most stock markets in Europe — maybe even combined. It's definitely bigger than Germany, definitely bigger than the UK. We do two to three thousand face-to-face meetings with management teams a year, and about 10–15% of those are with private companies; we focus in on the ones we really want to learn about and find ways to meet with them and get involved in their rounds.

Our first one was Stripe. We had a large investment, around 2017–2020, in Adyen, a fantastic next-generation cloud-payments company taking share from Worldpay — modern cloud payments were maybe 5% of an $80 trillion market. But you can't invest in Adyen unless you know Stripe like the back of your hand, so we did tremendous due diligence, talked to 200 customers about Adyen and about Stripe, and realised this was Coke and Pepsi. We said we've got to find a way to invest, and I finally got to meet the Collison brothers in 2019 — that was our first one.

We weren't really known for privates. I have a friend involved with a venture firm that held a large position, and I told him: let me know if you ever want to sell some. I got a call from him during Covid, in April 2020. We knew a lot about Stripe — we didn't have the full financials, but we knew enough that at that valuation, around $35 billion, and knowing they'd disclosed over half a trillion in TPV, and that Adyen's take rate was 25–30 basis points against Stripe's 40–50, and how many employees they had, we could get at the profitability. It turned out the take rate was higher, and Stripe had been modest about their TPV — it was much higher than the disclosed $550 billion, closer to a trillion. We underwrote it on our own assumptions and it was much better than we'd modelled, so we were able to upsize the position, buying a $100 million block from the seller. Sometimes sellers like that we're the kind of holder who'll own it in the public markets for a long time too, rather than a VC who's going to sell — which is what we did with Nubank as well; we've owned that in the public market for a long period.

Alex Sacerdote

We have an investment framework: it's the S-curve, competitive advantage, and then underappreciated earnings power — we'll dive into each one. When you get the right part of the S-curve, you get exponential unit growth. If you have a very strong business model — and in tech there are so many different types of moats — your earnings don't grow linearly, they grow exponentially. That's the last piece: invest when there's underappreciated long-term earnings power. Very often earnings can grow from $1 to $10, to $20 — it happens far more than people think, and it lets you buy some of the best companies in the world at extremely low P/Es. When we were buying Nvidia in 2023, we were paying four times earnings. When we bought Tesla in 2019 for the EV S-curve, we were paying five times earnings. When we owned Apple, we were paying four times earnings. When we bought Amazon for AWS, we were effectively getting it for free. The world doesn't think exponentially — everyone's focused on the next year, the next quarter. Very few people believe you can accurately predict two, three, four years out. But if you follow and understand the S-curve, and you know the moats, and you know how to model, you really can predict these great outcomes.

So, the S-curve. Every technology follows this pattern: it comes out, and — smartphones were around ten years before the iPhone. The internet was around twenty years before Netscape. AI was hidden inside companies for years before ChatGPT took it public and ignited what it was. Electric vehicles: Tesla went public fifteen years before 2019, when it went vertical, because there were so many barriers to adoption. The first smartphones were clunky, no touchscreen — not Apple's — no wireless data system, and they were too expensive, $500–$600. Steve Jobs got the price to $200; AT&T had a 3G network; it was touchscreen, easy enough that your grandmother could use it. Apple built an ecosystem and made it simple. All the barriers to adoption were eliminated, and then you rocket — that's the tornado of demand, when everybody in the world knows they need this right away. That's the flip that happens. It happened with electric vehicles: the price was too high, Elon got it to $40,000; range anxiety was there, he got the range to 300 miles; the supply chain was finally in place so he could churn out millions of them. That triggers the inflection.

The other nuance is that it's not just "oh, it's taken off now" — it's how tall this S-curve is, how big, so you know when to sell and how long to hold on, because we're underwriting two or three years out; we have to know what the growth looks like thereafter. And these S-curves can be dynamic. When Amazon had AWS, it was a hidden line item inside Amazon, covered by retail and internet analysts, not hardware or chip analysts — it was a new business model. But we realised the TAM was the largest in enterprise-IT history, because previously the TAM was routers, memory, storage, Dell, EMC doing it all in-house. We figured out they were addressing $600 billion of IT systems directly, and we thought it would be roughly 50% deflationary, so we were 1–2% penetrated. But over time we realised it actually wasn't deflationary — if you talk to anyone now, they'll say building it yourself costs about the same. So the TAM was much bigger than we thought. There are mega S-curves and sub-S-curves — we've been lucky to have internet 1.0, mobile, cloud, e-commerce, and now AI, which we can confidently say is the biggest, and these all build on one another.

With the electric-vehicle S-curve, you have to pay attention, because at the time we thought maybe 40–50% of cars would go electric, but it hit a wall at 10–15%. Usually S-curves go almost all the way, but in this case, for a variety of reasons, it didn't — so you have to adjust and stay on top of it. Generally, when something gets to 30–40% penetrated, you stop having exponential growth, which means the sell side catches up and there are no longer big beats.

Alex Sacerdote

Generally we like the high growth, and it was a mistake with Apple — in the first five or six years it was awesome, our largest position, up 50–70% a year, except in '08. We sold in 2012, when roughly 50% of the US had a smartphone. Apple maintained its leadership position; it had a couple of years of underperformance, then the multiple got low, and they added several ancillary things and got to play in the app ecosystem, taking 30% of app revenue. They were able to compound nicely after that, say 20%, but the big years were in the 0-to-50%-penetration part of the curve.

Alex Sacerdote

Andy Grove says that at strategic inflection points, you can't trust the data — strategic inflection points are about intuition, anecdotal evidence. I love this book called The Tao Jones Averages: A Guide to Whole-Brain Investing, about right-brain and left-brain investing. The best investors have the right-brain, creative side, where it's visual — connecting the dots. We invested in the mobile-video-game S-curve for a long time before it worked: the screens were small, the processing power wasn't good, so you had these casual games. Then I was in China and saw a twelve-year-old boy with a huge phone playing an awesome video game, and I thought: oh my god, it's coming to the phone. It's visual.

Enterprise is harder, because you can't see it the same way. We go to the Gartner IT Symposium — thirty thousand American CIOs go there — and we saw this with Splunk, which used to be an amazing database company; their room, when they were presenting, was standing room only. We saw the same with VMware, thirty years ago, when they virtualised the server — standing room only, and you could see the corporate demand just beginning. With AWS, we went and the grand ballroom was completely packed at nine o'clock, and still completely packed at ten, and at eleven. You could actually see the demand exploding before it happened. So we look for all kinds of clues; there's a whole pattern-recognition process. And by the way, it's okay to be late — it's okay to miss the first one, two, three years in a lot of cases, because if the top of the S-curve is half a trillion dollars, the growth can go on for a long time. You don't always have to be right there; it's okay to miss the first 100% of the move. Peter Lynch — I started at Fidelity and he loved to mentor the young analysts, so I got some time with him — said: write out the chart, it's all about the future.

What also helps with the S-curve is how long it goes on, and then there's the slope of the S-curve, which matters. A lot of people think, because we're in a modern world, everything moves fast, but there are a lot of factors that determine the pace of adoption. We commissioned a gentleman, Horace Dediu, who used to work with Clayton Christensen, to go look at history — we have the big S-curves of the last hundred years up on our wall. The radio S-curve is one of the fastest ever: it took seven years to reach nearly 100% penetration. The dishwasher S-curve is much slower, because it needs to be plugged into the plumbing.

Alex Sacerdote

B2B stuff can take a long time, because it needs to be plugged into existing systems — it's got to be put inside the house, like a dishwasher, whereas consumer adoption generally goes a lot faster.

Alex Sacerdote

I covered internet stocks at Fidelity — my first stock was Amazon, that's a whole other story. I also did B2B internet, and there was a whole bull case on it, but the underlying infrastructure wasn't in place, and it ultimately happened twenty years later, with SaaS. That's a risk with AI too: these big companies are very security-conscious, they can be slow to move, there's a lot of cultural friction — you need a few evangelists to push it through, and top management needs to push it too, while IT is saying it's risky. That happened with cloud as well; everyone was afraid it was unsecure to put data in the cloud, then we saw the CIA do it, we saw Capital One do it — we talked to the Capital One CIO, who said it's actually more secure in the cloud — and then it really started to take off. But those took a while, maybe because SaaS is like the dishwasher, it's got to be plugged in — it was growing, but at more like a 30–40%, maybe 50% growth rate. What's amazing about AI is that, at least for consumers, and even for business, you just open the browser and it's there. That's why we're getting this straight-up curve, and I think there's enough runway in the near term, going from 10 basis points of people really using it to 2–5% or whatever, to keep it going straight up. We call it a backwards L-curve. It's pretty exciting.

Alex Sacerdote

Well, we're definitely — you look for the S-curve, then do an exhaustive study of everybody with exposure to that area and try to find the one with a very powerful competitive advantage. A lot of people didn't like tech — Warren Buffett didn't like tech, because he couldn't predict the future fast enough — and the S-curve is our map for looking into the future. A lot of people were worried about tech because they thought there was so much disruption you could never trust a company to be a long-lived asset. What we've found over the years is that some of the competitive advantages within the digital world are as powerful, if not more powerful, than in the offline world.

There's the network effect, so powerful for LinkedIn, Facebook, Alibaba — you name it. You can become an industry standard: Oracle and Bloomberg are the industry standard. Oracle charges a lot, and there are free and open-source alternatives, but they had all the database administrators, all the software tuned to work with them — they had a chokehold on the relational-database market for decades. You can get to scale very quickly, because these S-curves grow fast — all of a sudden Anthropic is doing $30 billion in sales, or Amazon got so much scale so quickly: a Walmart-size scale advantage in five years, versus forty years for Walmart. So you can have network effects, scale, become an industry standard, become a platform people build on top of, have critical intellectual property — which is what Qualcomm had, you couldn't make a phone without paying them, or ASML, you can't make a chip without their lithography. And you can have brand: Google and Amazon never had to advertise to grow; Elon's never had to advertise for anything. And then there's cost-to-acquire versus lifetime value — the whole business model. Almost all the companies I've mentioned have several of these rolled into one.

Sometimes we can notice these things before the rest of the world. One of our high points was pitching Amazon for AWS at the Robin Hood Investors Conference in 2013 — we said the bulls have no idea what they're sitting on, Amazon had won the war before it even started; at the time we said there's Coke and there's no Pepsi. It turned out there was a Pepsi, but it was big enough to last regardless. We could see they had a seven-year lead — first-mover advantage matters. Then they became a whole ecosystem and platform, then they got scale — they were ten times the size of everybody else, and nobody could invest enough in R&D to catch them.

But you're right that if you don't have a competitive advantage, you can be in the best S-curve of all time and still lose out. If your name was RIM, Palm, Nokia, HTC, LG, Motorola — I could go on forever — zero, negative, negative, negative, negative. That's what we saw at the foundational-model layer, where there were something like fifty companies trying to do this, and they've nearly all fallen away, and two or three have emerged at the top, and there's a lot of reason to think they'll continue to hold their position.

Alex Sacerdote

Of all the S-curves we've done, AI is by far the most complex and the fastest-changing. We have to keep in mind that there are risks, but the rewards are also the highest, because we're talking about a market in the trillions — we used to say cloud might be $800 billion; this might be $3–5 trillion. Higher risk, higher reward. But take Anthropic: they have what looks like critical intellectual property, they've generally maintained their high market share in code. Second, they've built a strong brand for the enterprise — go and ask any CIO, and the first thing they'll say is Claude. They're getting to escape velocity and scale. What was scary for OpenAI and Anthropic, fighting these big companies like Google with their huge cash cows, was whether they could work in these hugely capital-intensive industries and find ways to raise capital — and to both management teams' credit, they've been able to. Certainly with Anthropic, given their roughly 10x sales growth and their fundraising ability, it looks like they've reached escape velocity, and now they have scale. The other thing Anthropic — and OpenAI — have is that, now that Anthropic is leading in code, they feed that code back into their own model: recursive improvement. And if you look at the pace of their innovation, it's accelerating, so maybe they get this liftoff stage.

OpenAI has been focused on so many different sectors, but they're starting to do better in enterprise, their coding tools are good, and they're seeing accelerating growth there too. And the consumer franchise — it looks like enterprise is much better right now, because you and I are willing to pay a lot for something that's replacing human labour, whereas with consumers maybe you get advertising, or maybe people would pay for a "Claude-bot" assistant if you could get it perfectly right for them, but they have a gazillion eyeballs there. Things do shift, but usually — we have a chart we use for almost all our pitches — on the internet, the leader goes bigger, faster, and wins. Shopify becomes the leader and keeps going. Amazon, the leader, keeps going. It compounds on itself. And another factor is scale: you need the compute, and you have to be able to pay for the compute, so there's only so many players who can do that. Those are some of the moats now showing up.

There are exceptions — with paradigm shifts, AOL going to dial-up and then to broadband, and AOL didn't make the change. Netscape came out early and didn't have as strong a business model. But if you talk to anyone in the Valley, or any startup, they'll tell you they're building on top of these three, and the world's a huge place, the economy's a huge place — they'll be able to differentiate within it.

Alex Sacerdote

Maybe five years ago, we had 40–50% of our portfolio in software. Early on, in our April 2023 seminar, we said to invest in chips first, but at the application layer we initially thought: these companies are huge, they have huge sales forces, they can take these AI APIs and build products, and they have the data — this is going to be amazing for software. Pretty quickly we realised their AI products weren't very good, they weren't moving the needle, nobody could charge for them. We sold almost all of our application software — we still have one or two small positions, but going into this year we were actually net short, and it really helped us in the first quarter.

The old way of software is like a pen and paper, or a horse and buggy. The new way is like a jet engine, or frankly the transporter from Star Trek — it's so revolutionary that it feels like it has to be disruptive, even if it isn't disruptive right away. The software companies have another problem: they've fallen down every CIO's priority list. Even if AI isn't disruptive yet, budget is going to Anthropic tokens instead, because there's faster ROI there — and that pushes on the software companies' own budgets, which hurts them. Third, a lot of software companies used to raise price every year; now they're probably nervous about doing that. Fourth — we'll see what happens with jobs, there are smart people on both sides of that — but we're seeing some companies really cut headcount or freeze hiring, which hurts on seats.

If you want to be optimistic about the software incumbents: we talked about how early the primitives of AI still are, so maybe they've just taken a while to get to something they can commercialise — but they might not have the right people, and it's a different selling motion, selling something that does human work rather than a fixed system; you need forward-deployed engineers, and they might not have those internally. Then there's the risk you can just build it yourself. The bulls will say companies are never going to build their own ERP system, and that's probably right — old tech is very sticky. Mobile video games didn't hurt console games; the tablet didn't hurt the PC; the smartphone didn't hurt the PC. There's a lot of integration and workflow baked into this software, and companies do prefer to buy rather than build. All that's true, but you can't rule out a world where, in one to five years, you get a brand-new AI-native company going after each of these strong incumbents, and their data advantage gets obviated, and it becomes easy to rip and replace with something AI-native.

What's good, if you like software, is that valuations are very high and everybody knows they're under pressure — some people are tempted to buy them, but the AI coding tools are getting better and better, so we'll have to wait and see. We're watching these companies closely for signs of revenue that can change the trajectory. It's hard, though — if you're Salesforce, with $40 billion in sales, and now $500–700 million of AI-related ARR, you've got this huge base and it takes a while to move the needle. In software there's the rule of 40 — growth rate plus operating margin; 20% and 20% is good. For AI we have a new version of that, really for chip investing: what percent of your sales are AI, say 30%, and what's your market share in that category, say 30% — that's 60, and that's a great place to look, because you've got exposure and a strong market position. The problem for software incumbents is their AI revenue is 1–2% at this stage, and it's a long way to go.

One thing we're picking up on now, though — and this is half-baked — is that AI could make some of these software platforms more important, because what's the first thing you do with Claude? You plug it into Slack. If that becomes a key repository, it makes Slack a permanent fixture in the organisation. Maybe the next wave of AI is agents that use tools, operating inside existing incumbent software the way a human being would.

Alex Sacerdote

It's still early in our thinking, but maybe Workday, or the HR systems, the big systems of record — the agents may end up running on top of them. CRM is going headless, or building a headless version, which is sort of the bear case too — you get relegated to being just a database. There's a human interface today; then they need to build the AI interface, which is really no interface — it's the agent going straight into the data, so you lose that customer interaction layer. But if the agents are going straight into CRM and doing the work inside it, that solidifies CRM — you won't have to worry it's going away.

Alex Sacerdote

For the past forty years, nothing much changed in the data centre. Even through the cloud era, it was basically Intel x86 — that became the data-centre chip sometime in the '90s. Compute workloads grew 25–40% a year in the cloud era, but Moore's Law was improving at roughly that rate, so it didn't require tremendous innovation, and there was almost no growth in hardware for years and years. The whole industry basically commoditised — every part of the server, the printed circuit board, the memory, the enclosures, the networking. There was no innovation; you'd go from one gig to ten gig networking and that would take seven years, and even in the switch year it took some innovation and created a little cycle, then commoditised again.

Now you go to AI, and workloads are growing 10x a year, pushing every aspect of the hardware to its physical limits. So not only do you get tremendous unit growth, but the industry is decommoditising — we call it that. At every aspect of the server. Memory, which used to be pure commodity — this high-bandwidth memory is now stacked ten chips high, the I/O is 10x what it was, it took Samsung years to do it, and it's a critical piece that keeps upgrading, so they have to be working with Nvidia three or four generations in advance.

We had this with Celestica — a contract manufacturer, and this has been a disaster industry since 1999, it all went offshore to China, pure commodity. But they hung on, and their heritage was IBM supercomputing, and they kept that talent and skill. Then we noticed they were the sole supplier of the Google TPU server — this was about three years ago, the stock was trading at eight times earnings. They also had a whole business selling Ethernet white-box switches — code for commodity Ethernet — into the clouds. It turned out these are excellent businesses. Not only tremendous growth, but to build an AI server, it's liquid-cooled, running much hotter, a $200,000–$300,000 piece of machinery versus an old $5,000 server — if that breaks, you throw it away; if this breaks, the whole thing goes down. It becomes critical infrastructure, like a critical part on a plane — you'll never get swapped out. It turned out Celestica was quite good at liquid cooling, and a lot of others tried and failed, so they've retained that position. The Ethernet market also used to upgrade on a seven-year cycle, 100 gig to 400 to 800; now they're upgrading every year, which is hard to do. There's a whole software layer too — the open-source SONiC layer — and the Celestica engineers who wrote some of that open-source software work very closely with Broadcom. What we thought was just a great growth driver turned out to be a great competitive advantage — they have 50–60% share of the cloud Ethernet-switch market, which is crucial because AI is incredibly network-intensive.

Even the printed circuit board: a regular server needs ten layers, these AI servers need forty, and there are very few PCB suppliers who can make that. We also own Elite Materials, which makes the leading ingredient, copper-clad laminate, that goes into these boards. PCB units are growing, layer counts are rising, so you've got a 50–60% CAGR just in units, then ASPs rising, then gross profits rising, and visibility has gone from "we'll call you next week if we need you" to "we need you designing this roadmap with us for the next four years." So you've gone from 5% growth, low margin, to 35–50% top-line CAGR for the next four years, with rising margins — and on top of that, there are shortages of everything, so even the commodity parts are having a great cycle.

We see that up and down the supply chain. We own Corning — they make the fibre, and they have a ridiculously high share of it. I was reading about a Microsoft data centre they just built — there's enough fibre in that one building to circle the world four and a half times. Corning's fibre is thinner, more bendable, specially manufactured to exact specs, higher margin, and it's the fastest-growing part of their business. In networking, there's "scale out," connecting server racks together, and "scale across," connecting data centres together — when you can't get all the power in one place for training, you wire clusters together across sites, and the wire needs to be 10x thicker, which creates huge growth. The real kicker is "scale up" — connecting every GPU in the rack to every other one — that's currently done over copper, but eventually will move to fibre, and when that happens it two-or-three-x's Corning's opportunity. So at every layer of the rack, everyone's overwhelmed. Even the story in power supplies: every Nvidia chip or rack uses 50–125% more power generation over generation, and that literally drives the ASPs of companies like Delta and Advanced Energy. I can't believe some of these stories — ASPs going up 40% a year for the next four years, and higher margin. The broader picture is that if we're right about this L-curve in AI demand, we're already short — the DRAM market, the NAND market, the PCB market — we're already about 30% short on all of these, as things stand now.

Alex Sacerdote

It's a good question. We did a presentation in 2024 where we listed everybody's market share, and I asked Claude to plot it, and it actually didn't get it right, because it missed the rate of change. The rate of change is important — you go from 10% to 30% and your growth rate accelerates and your margins accelerate. So rate of change matters a great deal.

Alex Sacerdote

My mom said, "Why do you tell everyone your secret?" It's like — why does the casino teach people how to play blackjack? It's harder than it looks. You have to be comfortable investing this way. We've been doing tech for twenty years at Whale Rock, we've got a team that's covered many cycles, we know the differences. Very few people have paid attention to hardware and chips at all — you've got all these newbies coming into it now.

Alex Sacerdote

And Gavin's done a great job. People weren't comfortable with it, and it's harder to do than it seems. And the chart — a lot of these companies' charts are already up a lot, so it's scary: can I really buy here? You also have to have a holistic view, because without conviction — every single year with Nvidia over the last four years it's been: they had a great year, oh my god, it's got to be a bubble; they have another great year, six months of marking time, it's got to be a bubble, this is getting out of hand. The bear cases aren't totally without merit, but if you can see the whole picture and understand how these things are unfolding, and gain conviction in that — frankly, so many semiconductor analysts missed it because they didn't see what was really happening at the foundational-model layer. It helps to have the big picture, to have decades and scores of S-curves you're looking at, and to see where each one fits.

Alex Sacerdote

One thing that bothers me is there's a lot of negativity in the general population about AI, and in some parts of government — Maine just banned data centres, and only about 20% of people are optimistic about AI, which raises the potential for negative regulation. But I do think the genie is out of the bottle. Another risk is if AI slows down in its improvements. Jensen Huang said this years ago, talking about graphics chips: "if good enough is good enough, I won't have a business." Every year he made the graphics a little better, and people always wanted the best. In AI, if Anthropic or OpenAI hit a wall and stop improving, the open-source models will catch up, and it might become a race to the bottom — which wouldn't be good for those stocks, but could actually be good for the chip companies, because chip companies don't care —

Alex Sacerdote

Who wins. That's another positive, and Nvidia will benefit if open source takes off — Jensen kept mentioning that at his last GTC, so that could be a risk to the model companies but not necessarily to us. Another risk is if one of the players falters and loses its position — that could mean a lot of compute demand that doesn't materialise. Though if AI is big enough, somebody else sucks that capacity up — we saw that when Oracle cancelled a big deal and Meta went right in. But suppose Meta had decided not to keep investing in AI at all, decided they couldn't keep up and it was a waste of resources — we watch that carefully. In general, we see more companies truly going after this, even Microsoft trying to build its own models. Those are some of the key risks.

Alex Sacerdote

Well, we do have some — part of OpenAI is ChatGPT, which is an application. But we think the application layer always comes later — the first three or four years of the iPhone, the applications took time to really arrive, so maybe it's just starting here too. To date, we've found that layer pretty risky, because it's not clear where the foundational model ends and the application begins, and whether applications can build enough of a moat to fend off the model companies and build durable businesses. We thought we'd see it in some incumbents, like the CRM players, and maybe it's just a matter of time, but we haven't really seen it yet in the enterprise world. There are some very good startup application companies out there, but the ecosystem isn't clear the way the chips ecosystem was clear when we started, or the way the foundational-model ecosystem is clearer to us now. At the application layer it's still unclear, and a bit dangerous.

There will be great application companies built, though. We've been watching Brett Taylor at Sierra — he was CEO of Salesforce, he wrote Google Maps, he was CTO of Facebook — he's building this fantastic company called Sierra. We're not involved, but that's where the rubber hits the road: will he be able to turn it into a huge company? He's doing quite well; we'll see. It's a matter of timing whether these things prove sustainable — it usually doesn't happen in the first three or four years, it comes a bit later.

Alex Sacerdote

I'd like to say we're so advanced in our AI systems that it's already a huge change, but it isn't yet — it's helping us get up to speed, we have a handful of great internal tools, but it's not supplanting the analyst's job. So much of what we do is meeting with as many companies as humanly possible, developing relationships with the management teams we cover, talking to the competitors. The system is straight out of Common Stocks and Uncommon Profits, written by Philip Fisher in the 1950s — the "scuttlebutt" approach, growth investing: get out there and talk to suppliers, customers, competitors, look for the key characteristics of the leading companies, and really develop conviction in them. If it's a new, complicated area — ABF substrates, PCBs — we can get up to speed on the technical side quickly using AI, but it can't pick the stocks for you.

If you're an analyst good at the blocking and tackling, there's still a role for that, but you need real insight on top of it. We're now using AI to write notes, review the quarter, and those notes are much better, but there'd better be a really good paragraph at the top — the wisdom: what does this mean, how does it fit our thesis, what changed? Don't just be a reporter. AI can be a great reporter; it can't yet see into the future. Take the work our analysts did on AppLovin two years ago — I think we had two of the best ad-tech analysts around, and they convinced me to buy. I actually started my career nearby, at an internet-advertising startup, before banking, so I knew ad tech, historically a terrible industry. Michael and Sam figured out the AppLovin story before almost anybody. They followed it while it was still private, knew all the competitors, all the terminology — Sam went to the app-advertising conference in Las Vegas, we went to industry conferences, talked to scores of people. They did the modelling work and built a great relationship with Adam Foroughi, one of the best managers out there. I don't see AI doing that.

Alex Sacerdote

One of the great things is just the friendships I've built with so many smart investors. Philip Fisher said part of his process was to get to know ten or fifteen like-minded people around the country and share ideas — and it is a two-way street. Many of them have been on your podcast, and you develop real friendships, share ideas, talk ideas. I call it the tripod: when I like something, and my analyst likes it, and somebody I really respect also likes it — those three legs of the stool really help build conviction.

Alex Sacerdote

There are several ways an investor can give us money.

Alex Sacerdote

For the first fifteen years it was a long-short fund. You want to be focused, and if you defocus, that's hard, so we grew it to the scale we wanted. We're twenty years old now; about ten years in, people started asking for a long-only product, so in 2020 we launched the long-only fund — we're six years into that now, and it's larger than the long-short fund; the bulk of our assets are in these two. Around 2015 we formalised the idea that we might do private investing, and gave investors the option to opt in or out, at 15% or 25% of a portfolio, though we didn't actually break the seal on privates until 2020. In 2021 we launched a hybrid fund that could go up to 80% into privates, for people who wanted more exposure there. And very recently we launched the Whale Rock Mega Cap Tech Fund.

We think there's a huge structural underweight to the largest tech companies across the industry, partly because a lot of our own performance over the years came from some of the largest companies — Apple, Amazon, Tesla — and it's hard for allocators to overweight to the degree that would matter. A lot of our largest pools of capital — endowments and so on — realise they've been massively underweight the largest tech companies, because they hold a lot of private assets and not much public, and of what public exposure they have, maybe half is international; and of their public bucket, there's a belief that there's no alpha in large-cap, so they underweight it and lean on small- and mid-cap stock pickers instead, because it's intuitive that large-cap can't have alpha. Even in their hedge-fund allocations, even if long-biased, they're not going to have 15% in Nvidia. But this is a product of the digital economy — in tech the leader usually grows bigger, wins, and develops very high market share quickly, with great competitive advantages, selling around the globe — and that's going to keep producing massive profit pools and massive market caps into the future. Most endowments are effectively betting against that, because they're so underweight.

Finally, one of our clients came to us and asked which index to use, and I'm on the investment committee at Hamilton College, where we'd been wrestling with the same question — we kept hearing it, and finally said we'd build this. There's a lot of alpha to be had, and the Mag 7, or FAANG, or whatever it becomes next, moves differently over time — in 2022 they all rallied together, but last year they were very divergent, and this year some are down. So we created the Whale Rock Mega Cap Tech Fund, where the universe is the top thirty market caps globally, and we pick the twelve or thirteen we think are best. There's actually tremendous alpha available in the largest caps, because with a small-cap stock it just takes one person figuring out it's good to move the price, but with a mega-cap it takes a hundred diversified PMs realising Google isn't a loser, it's a winner — and can we figure that out before 95% of those generalist PMs do?

Alex Sacerdote

Yeah, we like our odds in that, and there is alpha to be had there. As an asset category it's great too, because these companies by definition have wonderful moats — maybe they're not always the super-S-curve name, but sometimes they are: Nvidia certainly is, TSMC is heavily levered to it, SK Hynix is extremely levered to it, ASML is levered to it. We're four months into that fund now.

The right way to think about Whale Rock, really, is that we've built a research machine for understanding the world through the lens of companies, and the thing we're constantly trying to improve is that research machine — the products are just different ways of expressing it. We call it the Whale Rock learning machine: a group of ten highly experienced people. Warren Buffett reads books, and we read books, and we read blogs, but we're also in tech, so we go out and talk to people — we do 2,500–3,000 face-to-face meetings with management teams a year. Munger and Buffett talk about compounding knowledge; we've been compounding that knowledge for twenty years. There's turnover on the team, but broadly a lot of consistency — Andrew and Michael have been with me nineteen and eighteen years, and the average experience level on the team is around ten years, including some of the newer people. That research engine supports all of our products; it's the same people doing the public and the private work. We're not going to scour the world and turn over every stone, but when we see something that fits our system, we're able to act on it.

Alex Sacerdote

I have to say it's definitely my father. I was super lucky — he graduated from Cornell in electrical engineering, pivoted to Wall Street, and had a great career at Goldman Sachs; he ran corporate finance in the '80s and then ran private equity as chairman in the '90s. He was whip-smart, but he also had such humility, such a great gentleman. When I started Whale Rock, friends and family — he was the first call — and he said: I've been at Goldman for 41 years, how about I come and join you? I'll be the grey hair, I'll be the oversight, I'll be the chairman — you do what you do, build the firm, build the team, run the money, and I'll help raise some money. We got to work together for six years, until he passed away in 2011. I feel so lucky to have worked with him — it's not easy running a fund, and we never once raised our voices at each other. He was an amazing mentor to so many people, and when he passed away I got so many letters from people who said your father was such an influence on me, such a gentleman, such a great mentor. I just feel so lucky to have worked with him, and if I could be half the person he was, I'd be completely winning.

Alex Sacerdote

I don't know — he was modest, he was whip-smart, he was wise. He was also known as a great investor, which isn't the most common thing at an investment bank. He sat on their commitments committee, which kept him out of a lot of tougher situations. He was very warm — people could go into his office with problems, personal or work, and he handled it with grace. He had this soft way about him, and a great sense of humour too.

Alex Sacerdote

Yeah, I'm so lucky.

Alex Sacerdote

Thanks so much.