The AI S-Curve Investing Thesis: Why a 44%-a-Year Fund Says We’re Only 0.1% In

In 2026, the consensus word for AI is “bubble.” Headlines warn of stretched multiples, circular financing, and a top that feels overdue.

Then there is Alex Sacerdote — and the AI S-curve investing thesis he champions.

He is the founder, CEO, and portfolio manager of Whale Rock Capital Management, the Boston firm he started in 2006. His fund has compounded at roughly 44% a year over three years — fourth-best among global hedge funds — and returned about +54% in 2024 alone. And his message is the opposite of the bubble crowd: we are only about 0.1% of the way through this.

That is the paradox this piece unpacks. The AI S-curve investing thesis is not “AI is going up.” It is a chain of reasoning about where on the adoption curve we actually sit, what the binding constraint is for the next four years, and who captures the value when that constraint bites. Follow that chain honestly — including where Sacerdote’s claims run ahead of the verified data — and it terminates somewhere unexpected for a U.S. hedge fund: the two companies that sit atop Korea’s KOSPI.

Where Are We on the Curve? The 0.1% Claim

The thesis starts with a number designed to feel wrong.

Sacerdote argues that knowledge workers actually using agentic AI — systems that take multi-step actions, not just answer prompts — amount to roughly 0.1% of the global workforce. He attributes that estimate to Google CEO Sundar Pichai, not to his own research. Separately, he puts enterprise AI application penetration at under 1%, and the “under 1%” figure is the one verifiable as his direct claim.

The S-curve is the frame that makes this matter. New technologies don’t diffuse in a straight line. They crawl through an early flat stretch, then bend into a near-vertical climb as adoption compounds, then flatten again at maturity. Think of it like a stadium filling for a concert — empty for an hour, then suddenly packed in fifteen minutes once the doors really open.

If agentic adoption is near 0.1% and enterprise penetration is under 1%, Sacerdote’s argument is that we are still in the flat early stretch — with the steep climb from a few percent toward double digits still to come over the next four years.

The investing failure mode here is linear thinking. A human brain extrapolates the last few quarters and assumes more of the same. An S-curve does the opposite of “more of the same” precisely at the inflection — which is why early adoption levels routinely get mistaken for the ceiling rather than the floor.

Context, not endorsement. The 0.1% figure is Sacerdote citing Pichai; the <1% enterprise number is his direct claim. Neither is an independently audited statistic. We carry them as a thesis premise, not as established fact.

What’s the Constraint? Compute, and the Data Agrees

Sacerdote’s second move is sharper, and here the external data does most of the arguing for him.

His claim: the single most certain thing about the next four years is that compute will be short. He says it is already sold out, and that even Anthropic has secured only about half the compute it needs. That second figure is his characterization — it is not an Anthropic disclosure, and Anthropic has separately signaled it deliberately avoids over-buying GPUs rather than failing to buy them.

But strip away the attribution debate and look at what’s independently verifiable, because the 2026 spending data validates the direction with or without Sacerdote.

Fourteen of the largest datacenter operators are on track to spend roughly $750 billion in capex in 2026, up from under $450 billion the year before (BloombergNEF). And the striking part is not the number — it’s that, at that spend, they still describe themselves as unable to buy enough. NVIDIA and Broadcom reportedly asked TSMC for additional capacity and were turned down. When the buyers with the deepest pockets on earth get told “no,” that is a supply constraint, not a demand wobble.

This is the part of the thesis that is self-proving from the outside. You do not need to trust a hedge fund manager’s “half” estimate to see that ~$750B of spending colliding with a capacity wall is exactly what a binding compute bottleneck looks like.

Reading the claims honestly: a side-by-side

The intellectually honest way to present a bullish thesis is to put the strong claims next to the verified data and let the reader see the gap. Two of Sacerdote’s headline assertions are estimates, not audited facts — and saying so is what separates analysis from stenography.

Why this table matters. The direction of both claims survives scrutiny: supply growth is below norm, and compute is genuinely constrained. The magnitudes — “30%,” “half” — are Sacerdote’s framing. A thesis that needs you to ignore that distinction is weaker than one that survives it. This one survives it.

The Hardware De-Commoditization, or How 40 Years of x86 Broke

Here the thesis turns from “when” to “what,” and it is the real investment idea underneath the 0.1% headline.

For roughly four decades, the interesting margins in computing lived in software and in the x86 CPU. Everything around the processor — memory, networking, the printed circuit board, the cooling — was a commodity. Interchangeable. Low-margin. The kind of business no one wrote a thesis about.

Sacerdote argues AI snaps that. Compute load per model is rising on the order of ~10x a year, and at that slope the “boring” parts of the stack stop being boring. Memory, networking, PCBs, and cooling get re-rated from commodity components into hard intellectual property — businesses with moats and pricing power again.

The evidence is concrete, and it shows up in companies most readers have never heard of:

When the laminate under a GPU board and the fiber between racks become spec-defining, supply-constrained inputs, the whole “commodity hardware” mental model is obsolete. That re-rating — commodity to IP — is the engine of the thesis. Sacerdote owns names such as Nvidia to express it (Nvidia is the one holding verifiable from 13F filings; other thesis names like TSMC, ASML, and the Korean memory makers are part of the argument, not confirmed positions).

The chain, made explicit, is the spine of the entire thesis:

1. Penetration under 1% — the S-curve’s flat early stretch.
2. Adoption bends vertical, dragging compute demand up with it.
3. Compute becomes a four-year bottleneck, sold out at $750B of spend.
4. The scarcity pushes value down the stack: commodity hardware re-rates into IP.
5. And the tightest knot in that re-rated stack is memory.

The Tightest Knot: HBM, and Why the Thesis Lands in Korea

Every chain has a weakest link that everything else waits on. In this one, Sacerdote argues it is memory — DRAM, NAND, and the PCBs they ride on — which he estimates is roughly 30% short of demand.

That 30% is his number, and it is not what the measured data says. But the direction is confirmed by sources that have no reason to flatter a hedge fund.

IDC describes a “permanent, strategic reallocation of silicon wafer capacity” and pegs 2026 supply growth at +16% for DRAM and +17% for NAND — below the 20-30% historical norm. Goldman Sachs raised its 2026 DRAM supply-demand gap forecast from 3.3% to 4.9%, calling it the worst shortage in 15 years. So the “30%” magnitude is unverified, but “severe structural shortage” is independently true.

Then comes the mechanism that tightens the knot further. HBM4 — the next high-bandwidth memory generation, stacking 16 dies versus HBM3E’s 12 — consumes roughly 33% more die per stack. As HBM4 ramps into volume around late-2026 and 2027, it eats additional standard-DRAM capacity. HBM already absorbs an estimated ~23% of DRAM wafers; the new generation pushes that higher. The shortage isn’t easing — it’s compounding.

This is the turn that makes a U.S. bullish thesis a Korean story. Samsung and SK Hynix together hold roughly 90% of the HBM market, with SK Hynix in the lead at an estimated 53-62% share. Samsung is planning a ~50% HBM capacity surge in 2026 — lifting from about 170,000 to 250,000 wafers a month by year-end — and SK Hynix’s order books point to HBM shortage persisting through 2028.

Follow Sacerdote’s logic all the way down — penetration under 1%, compute as the four-year bottleneck, hardware de-commoditizing into IP, memory as the binding knot — and you arrive at the two companies that sit at the top of Korea’s KOSPI by market cap. The endpoint of an American compute-scarcity thesis is, structurally, a Korean memory duopoly.

The Other Half: Sell Rules and Concentration Risk

A thesis that says “we’re only 0.1% in” is dangerous if it’s read as “buy and never think again.” Sacerdote’s own framework is more disciplined than that, and the discipline is the part bullish readers tend to skip.

His S-curve method is about position, not timing. The rule is to buy when a technology enters the steep-growth phase, and to sell when the adoption rate decelerates — typically once penetration passes the 30-40% zone and the curve starts to flatten. The exception is a durable moat: a company like Apple can be held past maturity because its moat outlives the adoption curve. And because the total addressable market can be enormous, missing the absolute start isn’t fatal — he points to AWS as a case where late entry still worked.

The discipline shows up in how he finds the inflection: not from models, but from “scuttlebutt” — direct interviews with customers, suppliers, and competitors to detect the bend before it shows in the numbers.

For a Korean investor, the second discipline matters even more: concentration risk. A market where two semiconductor names dominate the index is a market where the index is the semiconductor cycle wearing a national costume. When that cycle turns, the concentration that amplified the upside amplifies the downside just as hard — a dynamic that recently surfaced in a KOSPI circuit-breaker event tied to exactly this concentration. The thesis can be right about direction and still leave a portfolio dangerously single-threaded.

The balance, stated plainly. “Only 0.1% in” is a statement about the runway, not a license to ignore the exit. The same S-curve that says buy early also says sell on deceleration — and in Korea, it says watch your concentration.

The Moat Map and the Recursive-Improvement Edge

One more layer explains which AI companies Sacerdote thinks capture the value, not just which hardware does.

He maps six classic digital moats — network effects, standards, scale, platform, core IP, and brand — which is a textbook-accurate taxonomy. On top of it, he argues the foundation-model labs (he names Anthropic and OpenAI) have stacked core IP, brand, and scale simultaneously. And the strongest moat of all, in his view, is recursive self-improvement: AI that writes and improves its own code, compounding capability faster than competitors can copy.

Anthropic is, by his account, his highest-conviction position. That conviction is a claim about a private company’s edge, not an audited fact — but it explains why the thesis treats compute scarcity as a feature for incumbents rather than a bug: the labs with the deepest moats are also the ones first in line for scarce compute.

One honest caveat on moats. Recursive self-improvement is a compelling story and an unproven one. It is the most speculative link in the chain — strong if true, and not yet demonstrated at the scale the thesis assumes.

What It Means for Korean Readers and Working Professionals

Strip the thesis to its load-bearing sentence and it reads: the question isn’t whether AI is a bubble — it’s where we are on the S-curve, and the answer points at compute, then memory, then Korea.

For a Korean professional, three things follow. First, the instinct to stay out because “it’s a bubble” and the instinct to buy because “we’re only 0.1% in” are answering different questions — one is about price, the other about position — and conflating them is the actual error. Second, even if the position read is right, the sell rule and the concentration risk are not optional footnotes; a KOSPI dominated by two memory makers is a single-cycle bet wearing the costume of a diversified index. Third, this is not a foreign story. If a 44%-a-year American fund’s most certain four-year call is compute scarcity, the supply-side node that call runs through is, structurally, Korean.

Investment implication. The real debate was never “bubble or not.” It is “where on the S-curve” — and an honest reading of that question routes through compute, into memory, and onto the two names atop the KOSPI. Direction: confirmed by IDC, Goldman, and BloombergNEF. Magnitude: still Sacerdote’s estimate, and labeled as such.

Risk factor. The thesis breaks if adoption stalls below the inflection (the S-curve never bends vertical), if compute supply catches up faster than the four-year call assumes, or if the memory shortage eases as new capacity lands. And for Korean portfolios specifically, the concentration that magnifies the upside magnifies the drawdown — the sell discipline (deceleration past ~30-40% penetration) is not optional.

Frequently Asked Questions (FAQ)

Q. What is the AI S-curve investing thesis in one sentence?

A. It is Alex Sacerdote’s argument that AI adoption is still in the flat early phase of an S-curve (he cites agentic usage near 0.1% and enterprise penetration under 1%), which makes compute the binding constraint for the next four years and pushes investment value down the hardware stack into memory. The conclusion routes value toward the companies that control scarce inputs like HBM.

Q. Is the “AI bubble” view wrong, then?

A. The thesis doesn’t claim valuations can’t be stretched. It reframes the question: “bubble or not” is about price, while the S-curve view is about position on the adoption curve. Both can be partly true at once, which is why the thesis pairs early-stage optimism with a strict sell rule when adoption decelerates past roughly 30-40% penetration.

Q. Are the “30% shortage” and “Anthropic has half its compute” figures verified facts?

A. No. Both are Sacerdote’s own estimates or characterizations, and this analysis labels them as such. Independent data confirms the direction — IDC shows below-normal supply growth (+16% DRAM, +17% NAND) and Goldman calls it the worst shortage in 15 years — but the specific magnitudes of “30%” and “half” are not independently measured.

Q. Why does this thesis point to Samsung and SK Hynix?

A. Because the chain ends at memory. Compute scarcity pushes value into the hardware stack, and HBM is the tightest knot in that stack. Samsung and SK Hynix together hold roughly 90% of the HBM market (SK Hynix leading at an estimated 53-62%), so a U.S. compute-scarcity thesis structurally lands on Korea’s two largest companies.

Q. What’s the main risk to the thesis?

A. Three things: adoption stalling before the S-curve bends vertical, compute supply catching up faster than the four-year call assumes, or the memory shortage easing as new capacity comes online. For Korean investors there’s a fourth — concentration risk, since an index dominated by two memory makers behaves like a single bet on the semiconductor cycle.

Disclaimer: This article is for informational purposes only and does not constitute investment advice. Figures attributed to Alex Sacerdote are his stated estimates or characterizations, not independently audited facts, and are labeled as such throughout. Positions, valuations, and supply forecasts are subject to change. Conduct your own research and consult a licensed financial advisor before making investment decisions.