Alex Sacerdote — Why the AI Boom Is Just Getting Started
"When you get the right part of the S-curve, you get exponential unit growth… the enterprise application AI market is less than 1% penetrated — we call this an L-curve, just straight up."
One-line take: Whale Rock invests where three things line up — a big S-curve (technology-adoption curve), a durable competitive advantage, and underappreciated long-term earnings power — which lets them buy great companies at low P/Es (Nvidia at ~4×, Tesla ~5×, Amazon "for free"). AI is the biggest S-curve ever and the model layer has settled into a three-horse oligopoly — Anthropic (his highest-conviction position, bought Aug 2025 at the ~$180B round), plus Google and OpenAI — with coding the "true unlock" (~20M coders × ~$20–30k/yr ≈ a ~$0.5T market from coding alone). His second big idea is the "decommoditization" of AI hardware: AI workloads push every part of the rack to its limits, turning once-commodity suppliers into IP-rich pinch points (Celestica, Corning, power, PCB) — "one of the best ways to play AI." He's bearish application software (sold ~all, entered the year net short), and runs a private process (Anthropic, Stripe) plus a new large-cap Mega-Cap Tech Fund.
1. Stocks & names mentioned
Sacerdote is a top-down technology-adoption (S-curve) investor — stance below reflects how each name was framed in this interview, not a price target. Research legend: QT Qualtrim · SA Seeking Alpha · STK Stock Analysis. Privates (Anthropic, OpenAI, Stripe) have no ticker. Ordered Positive → Neutral → Negative.
| Ticker | Name | Research | View | What he said | At |
|---|
| Anthropic | Anthropic (private) | — | Positive | His highest-conviction position (invested Aug 2025 at the ~$180B round). Coding is "the true unlock" — ~20M coders × ~$20-30k/yr ≈ a ~$0.5T market from coding alone; critical IP, an enterprise brand (CIOs "say Claude first"), escape velocity/scale and recursive self-improvement. | 6:31 |
| GOOGL | Alphabet (Google) | QT · SA · STK · FA | Positive | "We love Google… one of our largest positions" — Gemini is the third horse in the foundational-model oligopoly, attached to a huge business. | 3:18 |
| OpenAI | OpenAI (private) | — | Positive | Won the consumer; now improving in enterprise/coding with accelerating growth — one of the two or three leaders likely to hold position. | 38:16 |
| NVDA | NVIDIA | QT · SA · STK · FA | Positive | Bought in 2023 at ~4× earnings — the biggest S-curve; the chip "renaissance"/decommoditization. | 20:57 |
| AMZN | Amazon | QT · SA · STK · FA | Positive | Pitched AWS in 2013 ("won the war before it started," a 7-year lead, "got it for free") — the canonical mega-S-curve, ecosystem + scale moat. | 34:43 |
| CLS | Celestica | QT · SA · STK · FA | Positive | Sole supplier of the Google TPU server + ~50-60% of the cloud-Ethernet-switch market + liquid-cooling lead; bought ~3 yrs ago at ~8× earnings — "decommoditized" critical infrastructure. | 51:05 |
| GLW | Corning | QT · SA · STK · FA | Positive | Dominant fiber share (thinner/bendable, higher margin, fastest-growing segment); "scale-up over fiber" could 2-3× its opportunity. | 54:38 |
| AEIS | Advanced Energy Industries | QT · SA · STK · FA | Positive | AI racks draw 50-125% more power — power-supply ASPs rising ~40%/yr for four years at higher margin. | 56:10 |
| TSM | Taiwan Semiconductor | QT · SA · STK · FA | Positive | "Really levered to it" — a core AI winner in his new Mega-Cap Tech Fund. | 1:14:32 |
| ASML | ASML Holding | QT · SA · STK · FA | Positive | Critical IP — "you can't make a chip without their lithography." | 33:58 |
| APP | AppLovin | QT · SA · STK · FA | Positive | A research home-run — his analysts cracked the ad-tech S-curve early (followed it private) and built conviction before the market. | 1:08:16 |
| ADYEY | Adyen | SA · STK | Positive | Owned next-gen cloud-payments name ("Coke and Pepsi" with Stripe) taking share from legacy processors like Worldpay. | 17:23 |
| Stripe | Stripe (private) | — | Positive | His first private (2020) — the "Coke" of modern payments; underwrote it cheap at a ~$35B valuation and upsized a $100M block. | 17:58 |
| META | Meta Platforms | QT · SA · STK · FA | Neutral | Came in strong on foundational models, then faltered and had to reboot — a swing factor for AI compute (it took the big deal Oracle cancelled). | 2:51 |
| AVGO | Broadcom | QT · SA · STK · FA | Neutral | Celestica's close partner on the open-source SONiC switching software — central to AI networking (referenced, not a standalone call here). | 53:15 |
| AAPL | Apple | QT · SA · STK · FA | Neutral | His S-curve case study — a huge winner from 0-50% US smartphone penetration; he sold in 2012 once it hit ~50% (illustrative, not a current call). | 25:34 |
| TSLA | Tesla | QT · SA · STK · FA | Neutral | The EV-S-curve example — bought in 2019 at ~5× earnings as price/range barriers fell and the "tornado of demand" hit (illustrative). | 22:33 |
| ORCL | Oracle | QT · SA · STK · FA | Neutral | His "industry standard" database-moat example; separately cancelled a big AI compute deal that Meta then took. | 33:13 |
| CRM | Salesforce | QT · SA · STK · FA | Negative | Emblem of software under pressure — AI is only ~1-2% of its ~$40B sales, with budget/seat headwinds and "headless"/relegated-to-a-database risk. Whale Rock sold ~all application software and entered the year net short it. | 45:20 |
2. Talking points
1:16 The new compute stack — chips first
- When ChatGPT "fired the gun" in Nov 2022, the 10-person team did a massive deep dive. Any new compute paradigm creates a new stack — power, chips, clouds, foundational models, then applications — and new winners/losers.
- In early 2023 they decided to be in chips and infrastructure first: they get demand first, you know who the winners are, and no matter who wins above, the world needs tremendous compute.
2:25 60 contenders winnowed to a three-horse race
- Three years ago ~60 companies were chasing the foundational model. An April 2023 webinar framed it as either winner-take-all, a commodity race to zero, or an oligopoly of 3-4.
- Almost all startups died; even Amazon (never showed) and Meta (came in strong, faltered, had to reboot) stumbled. Anthropic was the enterprise dark horse, OpenAI won consumer, Gemini "can never be counted out." It became an oligopoly like the cloud's three-company structure.
4:03 Open-source-from-China risk — and why the leading edge holds
- Got comfortable that leading-edge token quality is superior: going from 80% to 85% of benchmark is a huge unlock. Open-source players lack compute, can come close but can't leapfrog, then falter.
- Scaling laws and feedback loops still have a strong runway per everyone close to the industry.
4:33 Code — "the true unlock of AI"
- First-gen tools (Microsoft Copilot, ~$20/mo) just improved grammar/found bugs. Mid-2025 Anthropic's coding could run agentically and the market exploded.
- Inside Anthropic people spent ~$100/day on tokens (~$20-30k/yr); ×~20M coders ≈ a ~$0.5T market from coding alone — on 7-9-month-old tech.
6:31 The Anthropic investment — Aug 2025, ~$180B
- Invested at the $180B round; numbers were "like nothing we'd ever seen" — $100M to $1B on the way to $9B revenue. Nobody knew what 2026 could be.
- The second unlock: Claude Code going almost fully agentic. Karpathy and Torvalds flipped — Karpathy now writes no code "except in English."
8:15 Models aren't commodities — critical IP + the "harness"
- Unlike commodity cloud, models differ by training methods and skills: Anthropic strong on finance/PE, Google on ingesting PDFs — differentiation = competitive advantage.
- Anthropic builds an ecosystem around the API (SDK, orchestration, the "harness" software that gets the most out of the model) — echoing how AWS built lock-in products from 2013.
9:48 The infrastructure S-curve — ~10% penetrated, not enough compute
- Infrastructure-layer S-curve is ~10% penetrated and "still one of the best ways to play AI." Most of the 800M users are on "AI 1.0" (search on steroids); new primitives (Claude on your computer, skills, bots) are barely started.
- Pichai: ~10 bips of knowledge workers truly use it; goes to 1-2-3-5-15% over four years. Enterprise hit a "light-switch" moment this year. There isn't enough compute — Anthropic has half what it needs; Andreessen is sure compute stays short for four years.
14:16 Getting private allocations as a public-markets investor
- It's a double opt-in. They passed on the $60B round (didn't know the company well enough, negative gross margins, hadn't seen coding explode yet).
- Spent time with Dario, judged management excellent (almost no turnover); built a 90-page deck using Claude Code, won an allocation, "punched above our weight" — a total home run.
16:39 The private process — 2-3,000 meetings a year
- Unicorn market is bigger than most European stock markets combined; you must know these companies. First private was Stripe (2020).
17:23 Stripe & Adyen — the "Coke and Pepsi" of modern payments
- Owned Adyen (next-gen cloud payments, taking share from Worldpay; cloud was ~5% of the ~$80T market). Couldn't underwrite Adyen without knowing Stripe cold — realized they were Coke and Pepsi.
- Met the Collison brothers in 2019; bought a $100M block in April 2020 at ~$35B. Knew ~$0.5T+ TPV (actually ~$1T) and the take rate (40-50 bips); underwrote it cheap and it proved better.
20:13 The framework — S-curve + competitive advantage + underappreciated earnings power
- The right part of an S-curve gives exponential unit growth; with a strong business model, earnings grow exponentially ($1 → $10 → $20). The world thinks linearly and short-term.
- This lets you buy great companies cheap: Nvidia ~4× (2023), Tesla ~5× (2019), Apple ~4×, Amazon "for free."
21:45 How S-curves inflect — barriers removed → "tornado of demand"
- Tech sits flat for years (smartphones 10 yrs before iPhone, internet 20 before Netscape, Tesla 15 before 2019) until barriers fall.
- iPhone: Jobs cut price to $200, 3G, touchscreen "your grandmother could do it." EV: price to $40k, range to 300 miles, supply chain ready. Then a "tornado of demand."
23:13 How tall is the curve / when to sell — Apple at ~50%
- You underwrite 2-3 years out, so you must know how tall the curve is. AWS's TAM was the biggest in enterprise IT ever (~$600B addressable, then bigger once it proved non-deflationary). Mega-S-curves vs sub-S-curves.
- Generally sell ~30-40% penetration when exponential growth ends and the sell-side catches up. Apple was a mistake — sold ~2012 at ~50% US smartphone penetration; the 0-50% part held the big years.
26:47 When to start buying — intuition, scuttlebutt, pattern recognition
- Grove: at strategic inflection points you can't trust the data — it's intuition and anecdote. Visual cues (a kid in China with a huge phone = mobile gaming inflecting).
- Enterprise is hard to see; the Gartner Symposium reveals demand (standing-room-only for Splunk, VMware, AWS). It's OK to be late and miss the first 100% if the top of the curve is huge (Peter Lynch: "white out the chart").
29:18 Pace of adoption — radio vs dishwasher; B2B is slow
- Radio reached ~100% in 7 years (one of the fastest); the dishwasher was slow because it must be "plugged into the back end." B2B is slow — it has to plug into existing systems (B2B internet only happened 20 yrs later with SaaS).
- AI's twist: with consumers and even business you "just open the browser and it's there," so it goes straight up — the "backwards-L curve."
33:13 The moats — the digital competitive advantages
- Network effect (LinkedIn, Facebook, Alibaba); industry standard (Oracle, Bloomberg — a chokehold on relational databases); scale (Anthropic to ~$30B sales fast; Amazon Walmart-scale in 5 yrs vs 40); platform; critical IP (Qualcomm, ASML — "can't make a chip without their lithography"); brand (Google, Amazon, Tesla never had to advertise).
- The 2013 AWS pitch at Robin Hood: "the bulls have no idea what they're sitting on… won the war before it started," a 7-year lead, then ecosystem + 10× scale. Best S-curve still loses without a moat (Rim, Palm, Nokia, Motorola).
36:46 Why Anthropic & OpenAI resist erosion
- AI is the most complex, fastest-changing S-curve — higher risk, but the highest reward (a market in the trillions; they now think ~$3-5T vs cloud's ~$800B).
- Anthropic: critical IP, sustained high share in code, an enterprise brand (CIOs "say Claude first"), escape velocity/scale, and recursive self-improvement (feeding their code back into the model — innovation accelerating). OpenAI won consumer and is now accelerating in enterprise/coding. "The leader goes bigger, faster, and wins."
40:46 The software bear case — sold ~all, net short
- Once ~40-50% of the portfolio was software; thought incumbents (sales forces + data) would build great AI products. They didn't — products weren't good, couldn't charge. Sold almost all application software; entered the year net short (helped in Q1).
- Four headwinds: faster ROI is in Anthropic tokens (software drops down the CIO priority list); token spend pressures budgets; pricing power gone; and job cuts/hiring freezes hurt seat-based models. Plus "build-it-yourself" risk via AI-native startups.
45:20 Salesforce & the modified "rule of 40"
- Salesforce has ~$40B sales but maybe ~$0.5-0.7B of AI ARR — only ~1-2%, a long way to go. The traditional rule of 40 = growth + margin.
- His new chip-investing version: %AI of sales × market share in that category (e.g., 30% × 30% = 60) flags where you have both exposure and position. Software fails it today. A silver lining: CRM "going headless" — agents working inside the data could solidify it, and Slack as a key repository could become permanent.
48:07 Chips — the "decommoditization" of hardware
- For 40 years nothing changed in the data center (Intel x86; Moore's law kept pace; everything — PCB, memory, enclosures, networking — commoditized). Now AI workloads grow 10× a year, pushing every part to its physical limit.
- This "decommoditizes" hardware: tremendous unit growth plus heavy innovation across the server — a "renaissance of chips" where the names are public with powerful IP.
51:05 Celestica, Corning, power & PCB — the pinch points
- Celestica: sole supplier of the Google TPU server + ~50-60% of cloud-Ethernet switches + liquid-cooling lead (kept IBM-supercomputing talent); ~$200-300k liquid-cooled machines = critical infrastructure that never gets swapped; bought ~3 yrs ago at ~8× earnings; co-wrote the open-source SONiC layer with Broadcom.
- Corning: huge fiber share (thinner, bendable, higher margin, fastest-growing); scale-up over fiber could 2-3× the opportunity. PCB: AI servers need ~40-layer boards (Elite Material's copper-clad laminate); ASPs/margins rising. Power: every Nvidia rack draws 50-125% more power — Delta/Advanced Energy ASPs +40%/yr for four years. They're ~30% short DRAM/NAND/PCB already.
58:48 Why few get it right — and the rate-of-change edge
- "Why teach people blackjack?" — it's genuinely hard; needs decades of S-curve pattern recognition and conviction to hold scary up-charts (Nvidia "must be a bubble" every year). Rate of change of %AI × share matters more than the absolute.
- Many semi-analysts missed it because they didn't see what was happening at the foundational-model layer — the big-picture, holistic view is the edge.
1:00:59 The risks — regulation, models plateauing, a player faltering
- Public/government negativity on AI (a state banning data centers; only ~20% optimistic) and regulation risk — but "the genie is out of the bottle."
- If a leader hits a wall, open-source catches up → a race to the bottom (bad for model stocks, fine for chips, which don't care who wins tokens). If a player falters, its compute gets absorbed (Oracle cancelled a deal, Meta took it).
1:03:04 Why he avoids the application layer
- Apps always come later (iPhone apps took years), but the ecosystem is still unclear and "a little bit dangerous": where does the model end and the app begin, and can apps build a moat?
- Watching Brett Taylor's Sierra (not involved) as the test case for whether a durable AI-native app company emerges.
1:08:16 The research "learning machine" — scuttlebutt & AppLovin
- Uses Philip Fisher's scuttlebutt approach (Common Stocks and Uncommon Profits). The AppLovin call: two ad-tech analysts cracked the story before anyone, followed it private, knew the competitors and built a relationship with CEO Adam Foroughi — "I don't see AI doing that."
1:09:00 The Mega-Cap Tech Fund — alpha in large-cap
- Sees a structural underweight of the world's largest tech companies; much of Whale Rock's historic performance came from the biggest names (Apple, Amazon, Tesla).
- The Mega-Cap Tech Fund's universe is the top-30 global market caps, picking the best ~12-13 — these have "wonderful moats." His AI-levered winners: Nvidia, TSM, SK Hynix, ASML. (Closing personal story: his late father, ex-Goldman, was Whale Rock's chairman "gray hair" until 2011.)
3. In plain English
A jargon-free summary of the thesis behind each name — what it actually is and why he holds that view. (Plain-language companion to the table above; renders on each ticker's consolidated page.)
Anthropic — Anthropic (private) Positive
Anthropic is the private company behind the Claude AI models — Sacerdote's single biggest, highest-conviction bet, made in August 2025 at a roughly $180 billion valuation. A "foundational model" is the core AI engine that everything else is built on; companies pay per "token" (the chunks of text the model reads and writes), so heavy users rack up big bills.
His thesis: coding is the "true unlock." Engineers using Claude to write software were burning ~$100/day in tokens — about $20-30k a year each — and with ~20 million coders worldwide that's a ~$500 billion market from coding alone. Anthropic has stayed ahead in coding, has a "moat" (a durable edge) from critical know-how, and an enterprise brand so strong that CIOs "say Claude first." It has hit "escape velocity" — enough scale and fundraising muscle to keep pulling away — and is even feeding its own coding tool back into improving its models, so progress is accelerating.
GOOGL — Alphabet (Google) Positive
Google's parent is "one of our largest positions." In the race to build the best core AI models, Sacerdote sees an oligopoly — just three serious players — and Google's Gemini is the third horse alongside Anthropic and OpenAI.
The extra appeal is that Gemini is bolted onto an enormous, cash-rich existing business (search, ads, cloud), which funds the brutally expensive AI buildout. "Gemini can never be counted out."
OpenAI — OpenAI (private) Positive
OpenAI, the private maker of ChatGPT, is the second of his three model-layer winners. It already "won the consumer" — ordinary people's default AI app — and is now getting better in the enterprise and in coding, where growth is accelerating.
His broader point is that the leaders in a technology race tend to keep leading ("the leader goes bigger, faster, and wins"), so OpenAI is likely to hold its spot as one of the two or three survivors.
NVDA — NVIDIA Positive
Nvidia makes the chips that train and run AI. Sacerdote bought it in 2023 at about four times earnings — dirt cheap — because he saw the biggest "S-curve" ever beginning. An S-curve is the typical path a new technology takes: a long flat start, then an explosive vertical takeoff once the barriers fall, then a leveling off; catching the takeoff is where the money is.
He frames Nvidia as the centerpiece of a chip "renaissance," where AI is forcing real innovation back into hardware that had been a sleepy, commoditized business for decades. Every year people called it a bubble; every year it kept compounding because the demand was real.
AMZN — Amazon Positive
Amazon is his textbook "mega-S-curve" win. Back in 2013 he pitched it for AWS — its cloud-computing arm — when that business was a hidden line item the market ignored; he said "the bulls have no idea what they're sitting on," that Amazon had "won the war before it even started" with a seven-year head start, and that you were effectively getting AWS "for free" at the price.
The lasting moat is the combination of being first, becoming a platform others build on, and reaching such enormous scale (10× rivals) that no one could afford the R&D to catch up.
CLS — Celestica Positive
Celestica builds the physical guts of AI data centers. For decades it was a low-margin "contract manufacturer" (it assembles hardware for others) — a commodity business. Sacerdote bought it ~3 years ago at about eight times earnings after spotting it was the sole supplier of Google's TPU server and held ~50-60% of the market for cloud "Ethernet switches" (the gear that wires servers together).
This is his "decommoditization" idea: AI machines run so hot they need liquid cooling and cost $200-300k each (vs ~$5k for an old server), so if one breaks the whole system goes down. That turns Celestica's parts into critical infrastructure — like a critical part on a plane that never gets swapped out — giving it real pricing power and durable advantages where there used to be none. ASP = average selling price.
GLW — Corning Positive
Corning makes the optical fiber that carries data around and between AI data centers — and it has a dominant share. Its fiber is thinner, more bendable, and can be made to exact specs, which makes it higher-margin and the fastest-growing part of its business (one Microsoft data center reportedly held enough fiber to circle the Earth 4.5 times).
The big upside ("kicker"): as AI clusters grow, the industry will start connecting the chips inside each rack ("scale-up") over fiber instead of copper — and Sacerdote says that shift alone could 2-3× Corning's opportunity.
AEIS — Advanced Energy Industries Positive
Advanced Energy makes the power supplies that feed AI server racks. Because each Nvidia chip or rack draws 50-125% more power than before, the price of each power unit it sells (its ASP, average selling price) is rising about 40% a year — and Sacerdote expects that to continue for four straight years, at higher margins.
It's another "decommoditization" pinch point: a once-boring component that AI's extreme power demands have turned into a high-growth, higher-margin product.
TSM — Taiwan Semiconductor Positive
TSMC is the world's dominant chip manufacturer — almost every advanced AI chip is physically made in its factories. Sacerdote calls it "really levered" to AI: as AI chip demand explodes, the demand flows straight through to TSMC.
It's a core holding in his new Mega-Cap Tech Fund, which picks the best of the world's largest tech companies — names with wide moats that he thinks are structurally underweighted by investors.
ASML — ASML Holding Positive
ASML is the Dutch company that makes the lithography machines used to print the circuitry on advanced chips. Sacerdote uses it as the classic example of "critical IP" — proprietary technology no one can work around: "you can't make a chip without their lithography."
That monopoly-like grip on an essential step makes it one of the strongest moats in the whole AI supply chain, and a levered way to play rising chip demand.
APP — AppLovin Positive
AppLovin is an advertising-technology ("ad-tech") company, and Sacerdote cites it as a research home-run rather than an AI story. Two of his analysts cracked the AppLovin story before the rest of the market, tracked it back when it was still private, learned all its competitors, and built a relationship with its CEO.
His point is about process: deep, old-fashioned "scuttlebutt" research — talking to everyone around a company — found the winner early, and "I don't see AI doing that."
ADYEY — Adyen Positive
Adyen is a next-generation, cloud-based payments processor (the plumbing that lets merchants accept card and online payments). Sacerdote owned it as the "Pepsi" to Stripe's "Coke" — two modern winners taking share from clunky legacy processors like Worldpay.
Modern cloud payments were only ~5% of a roughly $80 trillion market when he looked, so the runway to take share was enormous. The "take rate" — the small cut a processor keeps on each dollar that flows through it (its TPV, total payment volume) — is how these businesses make money, and Adyen's was attractive.
Stripe — Stripe (private) Positive
Stripe is the private payments giant — the "Coke" of modern online payments — and was Whale Rock's very first private investment, in April 2020. He'd met the founders (the Collison brothers) in 2019 and knew the business cold from researching Adyen.
He underwrote it cheaply at a ~$35 billion valuation: he could estimate its profitability from its disclosed TPV (total payment volume — the dollars flowing through it, over $0.5 trillion and really closer to $1 trillion) and its "take rate" (the cut kept per dollar, ~40-50 basis points). The numbers proved even better than he assumed, and he upsized into a $100 million block.
CRM — Salesforce Negative
Salesforce is the emblem of his bearish view on application software (the business programs companies buy, like CRM — customer-relationship-management tools). AI is only ~1-2% of its ~$40 billion in sales, so even if its AI products work, they're a tiny drop in a huge bucket and will take years to matter.
The headwinds stack up: customers would rather spend on faster-payback AI tokens, which squeezes software budgets; these vendors have lost the ability to keep raising prices; and AI-driven job cuts hurt "seat-based" pricing (charging per user). The scary scenario is software going "headless" — AI agents bypassing the human screen and working straight in the data — which risks relegating Salesforce to just being a database. Whale Rock sold nearly all its application software and entered the year betting against it (net short).
Summary & timestamps derived from the public YouTube video (transcript in transcript.txt) for personal study. Not investment advice. © Invest Like the Best / Colossus for source material.