Picks & Shovels

The question

Are the vendors selling compute, lithography, power, and interconnect into the hyperscaler capex cycle priced for the run-rate of spending implied by current CFO guidance, or for something materially lower?

Base rates

The reference class here is capex super-cycles led by a small number of customers with balance-sheet capacity: the 1996-2000 telecom buildout, the 2003-2008 mining capex cycle, and the 2010-2014 shale cycle. In each case, the picks-and-shovels cohort compounded revenue 20-35% annually through the mid-cycle years, then derated sharply when order books rolled. The historical base rate for equipment and component suppliers achieving a 3-year forward return above the index during the mid-cycle phase is roughly the 65th-75th percentile of industrial subsectors. The base rate for holding into the derating phase is worse than holding the S&P -- roughly the 25th percentile. So the thesis is path-dependent: we are claiming the cycle is mid, not late.

The imputed probability embedded in consensus forward numbers for this cohort is that 2027-2028 accelerator and foundry spend is roughly 15-25% below the run-rate of signed 2026 orders. That is a reasonable prior for a normal cycle. It is likely too conservative if hyperscaler compute is, as their own disclosures suggest, supply-constrained rather than demand-constrained.

Why the consensus view is wrong (or incomplete)

The sell-side is forecasting AI infrastructure spend as if it were a normal capex cycle that mean-reverts. That framework has worked for forty years. The causal mechanism that breaks it here is that compute is an input to a product (inference) whose unit economics improve with scale and whose demand curve is still being discovered. In a normal capex cycle, the marginal dollar of capacity competes with existing capacity for a known market. In this cycle, the marginal dollar of capacity creates its own market by lowering the cost of a new category of software labor.

That does not mean the cycle cannot end. It means the end is more likely to come from a constraint -- power availability, HBM yields, advanced-node wafer starts -- than from demand exhaustion. The market is pricing demand risk; the real risk is supply-side. That asymmetry is the inefficiency.

Position construction

Three 20% core positions anchor the book: NVDA for the accelerator, TSM for the foundry, AVGO for the custom silicon and networking ASIC franchise. These are the choke points; if the buildout happens at all, the dollars flow through these three. ASML at ~10% is the sole supplier of EUV and the constraint behind every leading-node wafer -- a pure royalty on advanced-node capacity. LRCX, AMAT, KLAC together (~13%) are the deposition/etch/metrology oligopoly; margin profiles are 30%+ operating and the installed base generates a recurring services stream that damps cyclicality.

ARM (~4.5%) captures the ISA royalty on every custom hyperscaler CPU. CEG and ETN (~7% combined) are the power stack -- nuclear PPA optionality and electrical gear into data-center builds. VRT (~1.7%) is cooling; COHR and CRDO (~1.1% combined) are optical components and retimers where the unit-count story is more attractive than the unit-economics story, hence smaller weights. DELL (~2.6%) is the systems integrator with the thinnest moat in the book and is sized accordingly.

The book is deliberately concentrated in the top six names because the hard 1-year-vs-SPY screen disqualifies most of the cohort; we are buying the survivors of the rerating, not the hope trades.

Asymmetric payoff

If hyperscaler capex grows 15-20% through 2027 (roughly the midpoint of disclosed capex guidance) and multiples hold, the weighted book returns roughly 25-40% per year. If capex flat-lines in 2027 and multiples compress one turn, the book returns roughly -15 to -25%. If capex accelerates and a credible inference-demand signal forces sell-side to take 2028 numbers up, the right tail is 50-70% with multiple expansion.

At a 60% probability of the base case, 25% probability of the bear, and 15% probability of the bull, expected value is roughly +15 to +22% annualized versus an SPY base rate near +8%. The payoff is asymmetric because the bear is bounded by cash-generative installed bases and the bull is open-ended.

Three things that would change our mind

1. Two consecutive quarters of hyperscaler capex guide-downs aggregating more than 10% versus the prior cycle peak, with management language shifting from supply-constrained to demand-constrained.
2. HBM pricing declines of 20%+ quarter-over-quarter sustained for two quarters, indicating accelerator demand is softening ahead of reported revenue.
3. TSM advanced-node utilization (N3/N2) dropping below 85% on the quarterly call, which would signal that the wafer-start constraint has broken.

What we are explicitly NOT betting on

We are not buying model developers, foundation-model API businesses, or the vertical AI application layer. The unit economics of inference are still being negotiated, and the historical base rate for the application layer in a platform shift is that returns accrue to a small number of winners selected ex-post. The picks-and-shovels cohort does not require us to pick the winner of the application layer; it requires only that somebody spends the capex. That is a strictly weaker claim, and the portfolio is sized for it.