In Dwarkesh Patel’s own analysis, three factors multiply to roughly 3× annual compute growth — but the largest of the three moves existing capacity rather than creating new capacity, which means it carries a hard arithmetic ceiling.
In an essay titled Why Compute Might Get 10x More Expensive, Dwarkesh Patel — working from his own analysis and projection rather than from published industry data or consensus — breaks annual compute capacity growth of roughly 3× into three multiplicative components: 1.4× from Moore’s Law, 1.2× from new fab construction, and 1.8× from AI taking a larger share of leading-edge wafer allocation away from other devices such as smartphones and PCs. Notably, 1.4 × 1.2 × 1.8 = 3.024, meaning the stated parts reproduce the stated whole — an auditing property that most published growth figures do not offer.
The reallocation argument has a specific shape in his telling. He projects — and this is his forecast, not a scheduled industry event — that AI’s share of N3 leading-edge wafers moves from roughly 60% to 86% by the end of 2027, at which point the reallocation gains “will actually hit a wall within a year or two.” Once AI holds the near-entirety of the most advanced node allocation, there is nothing left to reallocate. The arithmetic stops there regardless of demand.
As colour for the demand side, Patel notes — using the word “reportedly” himself, treating it as second-hand — that Google is reportedly paying $900 million per month for 110,000 GPUs blending GB200s and GB300s, roughly twice spot price per hour, with spot itself running about 40% above its February level. That figure is his reported colour, not a verified data point, and nothing here is built on it.
The key insight: The largest single contributor to Patel’s compound growth figure is the one contributor that is bounded by arithmetic. Moore’s Law and new fab construction can, in principle, run indefinitely — constrained by physics and capital but not by a share that must stay below 100%. Wafer reallocation cannot exceed total wafer supply. The ceiling is not a risk; it is a definitional property of the term, and on Patel’s own projection it arrives around the end of 2027.
Two of these make capacity that did not exist. The highlighted one moves capacity that already did, and it runs out.
The Structural Read
The three factors are not the same kind of quantity, and that asymmetry matters more than any individual number. Moore’s Law and new fab construction both expand the total supply of leading-edge compute: more transistors per die, more physical plant producing wafers. Wafer reallocation adds nothing to total supply. It redistributes an existing stock — what was serving consumer electronics now serves AI training clusters instead.
From outside the system, a year in which AI received 1.8× more leading-edge wafers looks identical to a year in which 1.8× more leading-edge wafers were manufactured. The output series is the same. Only the mechanism differs — and mechanism is precisely the information a trend line discards. A series driven by reallocation has a terminal value that a series driven by production does not, and nothing in the historical data distinguishes the two. You have to know where the capacity came from, which is information the chart never contains.
This is also why the 60-to-86 figures are the most analytically useful element of the essay. They state the remaining headroom explicitly. Most growth stories never tell you how much room is left in the mechanism that is doing the most work. This one does the arithmetic in public, which makes it falsifiable in a way that a simple headline growth rate is not.
Product Overhang Doctrine — Applied
Reallocation Looks Like Growth Until It Doesn’t
The Product Overhang pattern describes capability that accumulates invisibly until it surfaces all at once. The inverse applies here: a constraint that accumulates invisibly until it surfaces all at once. When AI’s share of leading-edge wafers approaches its ceiling, the 1.8× reallocation term collapses toward 1.0× — not gradually, but as a step-change in the growth equation. The remaining two terms (1.4× and 1.2×) multiply to 1.68×, not 3×. Anyone extrapolating the compound figure without disaggregating its components will not see that transition coming in the data.
Three Implications
IMPLICATION 1 — THE AUDITABLE DECOMPOSITION
Because 1.4 × 1.2 × 1.8 = 3.02, Patel’s estimate is auditable in a way that a single headline growth figure is not. You can disagree with any one term — say you think Moore’s Law delivers 1.2× rather than 1.4× — and immediately see what that does to the total. That is the opposite of a black-box projection. The form of the claim carries analytical value independent of whether its specific numbers prove right.
IMPLICATION 2 — THE LARGEST TERM IS THE BOUNDED ONE
The 1.8× reallocation term is the dominant contributor to the compound figure — larger than Moore’s Law (1.4×) and larger than new fab construction (1.2×). It is also the only term with a hard ceiling defined by arithmetic rather than by physics or capital markets. That structural fact holds regardless of whether Patel’s specific share percentages or timing are correct. Any growth decomposition where the largest term is a share transfer deserves scrutiny about how much headroom remains.
IMPLICATION 3 — THE WALL IS PATEL’S FORECAST, NOT A SCHEDULED EVENT
The projection that AI moves from 60% to 86% of N3 wafer allocation by end-2027, and that the wall arrives “within a year or two,” is Patel’s own forecast. This publication makes no forecast of its own about compute prices, capacity, or timing. The value of the framework is structural: it tells you which term to watch and why a share-transfer term behaves differently from a production-expansion term — not when exactly the transition occurs.
Patel’s Projected AI Share of N3 Leading-Edge Wafers — His Forecast
Current share (Patel’s estimate)~60%
Projected share by end-2027 (Patel’s forecast)~86%
Remaining headroom from ~60% to 100%: ~40 percentage points. From ~86% to 100%: ~14 points. The reallocation term exhausts as the share approaches its ceiling. Patel’s analysis and projection only.
Business Engineer Framework
The Map of AI — Understanding Where Compute Sits in the Stack
Compute is Layer 0 of the AI stack — the physical substrate everything else depends on. The Map of AI traces all nine layers, from silicon to application, and shows how constraint at any one layer propagates upward. Understanding that the dominant growth term in current compute expansion is a reallocation rather than a production event changes how you read every layer above it.
Dwarkesh Patel’s decomposition — his own analysis and projection, not industry consensus — is most useful not for its specific numbers but for a structural property it makes visible: the term doing the most work in the compound growth figure is the term with a hard arithmetic ceiling, and most growth extrapolations will never tell you that. Moore’s Law and new fab construction can expand supply; wafer reallocation can only redistribute it. When the redistribution runs out of room, the 1.8× collapses — and on the decomposition as stated, you’re left with 1.68× where 3× used to be. That is not a prediction about when. It is a description of what kind of thing is happening, which is the prior question.
The decomposition above is Dwarkesh Patel’s own, taken from his essay, and it is his analysis and projection rather than measured fact or industry consensus. The figures were checked against the essay text rather than against the audio of him reading it. His path from 60 per cent to 86 per cent of leading-edge N3 by the end of 2027, and the wall he expects a year or two afterwards, are his forecast and not a scheduled event.The Google rental figure he cites is marked “reportedly” in his own words, making it second-hand even within the essay; it appears above once as colour and nothing is built on it.This publication makes no forecast of its own about compute prices, capacity or timing. TSMC’s actual node allocation and any company statement about it, ASML machine counts, lead times and order books, any chip or wafer price or contract, any company’s capacity, and any claim about what compute growth will actually turn out to be are not established and do not appear. Nothing above is investment advice.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
