Hyperscaler Capex Sponsors

The question

The right question for this book is not whether hyperscaler capex is large. It is whether the market's imputed probability of sustained high returns on that capex is reasonable, and whether the companies writing the checks are priced as though the returns will hold. Aggregate capex at the top four US hyperscalers has moved from roughly $90B in 2019 to an expected $340B+ in 2026. The implied reference class is something closer to telecom 1999 than to industrial compounders of the 1990s, and that comparison is the consensus worry. Our view is narrower: the sponsors with pre-existing cashflow engines can absorb a wide range of return outcomes without impairing intrinsic value, while a selected group of suppliers offers asymmetric payoff if utilization holds, and option value if it does not.

Base rates

The base rate for multi-year capex super-cycles is not encouraging. Studies of capital-intensive buildouts (rail, fiber, shale, wind) show median incremental ROIC compressing roughly 300-500bps from peak to trough, with a five-to-seven year lag between spend and return normalization. The base rate for large-cap platform businesses that already earn mid-twenties operating margins is materially better: companies entering a capex cycle from a position of high free cash flow conversion have historically preserved 70-80% of pre-cycle ROIC, because the incremental asset base is funded from retained earnings rather than debt or equity issuance. That distinction — who funds the build — is the crux.

Why the consensus view is wrong (or incomplete)

The prevailing narrative treats hyperscaler capex as homogeneous. It is not. Alphabet and Amazon are funding buildout with roughly 100% of spend covered by operating cash flow; several mid-tier compute providers are funding via a mix of preferred equity, vendor financing, and receivable-backed debt. The market is pricing both cohorts off a shared multiple framework, which creates dispersion we can exploit. Second, the consensus view assumes token-economics will mean-revert quickly. The evidence from 2024-2025 is that inference unit costs are falling faster than demand elasticity is absorbing, which lengthens the window of positive unit economics for capacity owners. Third, the consensus treats suppliers as commodities. A handful of ODMs and rack integrators have locked in multi-year, cost-plus frameworks that look less like cyclical hardware and more like contract manufacturing with escalators.

Position construction

The book is organized into three sub-books.

Sponsor balance sheets (40%) — GOOGL and AMZN at 20% each. These are the companies writing the largest checks with the least dilution risk. We own them for the combination of (a) cashflow optionality to slow the build if returns disappoint and (b) the embedded cloud franchises whose unit economics improve as capacity comes online. GOOG Class C is held at zero weight and is listed for ticker consistency with composite feeds.

Application-layer beneficiaries (20%) — APP at 20%. The thesis here is that the companies monetizing inference at the ad-tech and consumer-discovery layer capture a disproportionate share of the economic surplus, because their cost-of-revenue falls as inference prices fall while pricing power is anchored to advertiser willingness-to-pay. APP is the cleanest expression we found.

Suppliers and neoclouds (40%) — CRWV 16.2%, CLS 7.3%, HPE 6.9%, NBIS 5.8%, IREN 3.8%. This is the asymmetric sleeve. CRWV and NBIS are pure-play compute providers with contracted backlogs that extend past the near-term capex peak; CLS and HPE are rack and system integrators with book-to-bill above 1.2x; IREN is a lower-conviction option on power-adjacent compute. Weights decline with visibility of cashflow; the tail is small on purpose.

Asymmetric payoff

Our rough expected-value frame assumes three states. In a base case (50% probability), hyperscaler capex plateaus in 2027 and the sponsor cohort delivers 10-12% annualized; the supplier sleeve delivers 15-20%. In an upside case (25%), utilization holds through 2028 and the supplier sleeve delivers 40%+. In a downside case (25%), capex is cut 30% and the supplier sleeve draws down 35-45% while the sponsor cohort protects to a 10-15% drawdown. Probability-weighted, the book offers a mid-teens expected return with roughly 2:1 upside-to-downside asymmetry, driven more by the sponsor cashflow floor than by the supplier beta.

Three things that would change our mind

First, a sustained inflection in cloud gross margins lower rather than flat; that would indicate the incremental capex is not earning a competitive return. Second, a shift in hyperscaler financing mix toward off-balance-sheet structures or vendor-financed supplier receivables; that would signal return-on-capital pressure is being masked. Third, a credible entrant at the model layer that commoditizes inference pricing faster than capacity providers can amortize their builds.

What we are explicitly NOT betting on

We are not betting on a particular model architecture winning. We are not betting on semiconductor market share shifts at the accelerator layer. We are not betting on sovereign AI spend filling any capex gap left by the hyperscalers. And we are not betting that the suppliers can hold current margins through the next down-cycle; we are betting the backlog carries them through the window we care about.