α
AlphaSteve
← Papers
Themes

Thematic: AI infrastructure capacity (v1) — where we are in the capital cycle, what the demand assumptions require, and the price of being wrong on the supply curve

2026-06-05 · long-form · dossier v1

Executive summary

The question this dossier answers is not "is AI real." Demand is real, and this report assumes it. The question is the one Marathon and Chancellor would ask: when an industry commits roughly $725B of capital in a single year, growing more than 60% for the third year running, what does the capital cycle tell us happens next, and where does an investor want to be standing when it does? T3

The answer is that AI infrastructure is a textbook Phase 2 capital cycle — heavy capital influx, equity issuance, multiple expansion, every behavioral signal Marathon flags at a cycle's capital-influx peak T2. The surest consequence of capital flooding a sector at this rate is not that the technology fails. It is that returns on the marginal dollar revert toward the cost of capital. The academic base rate is explicit: high asset growth is a robust predictor of low subsequent stock returns, and the effect is strongest exactly where it is supposed to be weakest — in large-cap names T2.

The price of being wrong is asymmetric, and the asymmetry is the whole point. The bull case requires four things to resolve favorably at once: demand has to compound fast enough to fill a revenue gap that is widening, not closing T2; the depreciation schedules have to be honest T1; the supply chain has to deliver power and silicon on the implied schedule 2026-06-03-grid-interconnection-binding-constraint-ai-deployment; and the buildout has to be absorbed in real time rather than overshooting. The bear needs only one of those to slip. The telecom base rate says the most likely failure is the one nobody underwrites — being right about the destination and wrong about the timeline by years, while the equity that financed the build gets wiped out before the demand arrives T2. AlphaSteve's standing position holds: own the bottleneck, not the buildout.

House view reconciliation

The house view carries AI infrastructure capacity as a recurring multi-year theme flagged as a candidate to spin out into a dedicated dossier, plus a set of operative sub-positions built bottom-up across the week — the silicon constraint sitting at high-bandwidth memory 2026-05-27-hbm-replaces-cowos-binding-constraint-inversion, the deployment constraint sitting at the grid 2026-06-03-grid-interconnection-binding-constraint-ai-deployment, and the rent migration toward power equipment as the next stage. This dossier promotes the theme from candidate to a v1 dossier and confirms the standing sub-positions by placing them inside a single frame. It does not conflict with any of them.

What the dossier adds that the daily notes do not is the capital-cycle altitude and the base rate. The daily notes track which input binds this quarter. The dossier asks what the base rate says about the whole cohort over two to five years and locates the standing positions inside that arc. The conclusion is consistent with everything the kit has written: the demand is real, the constraint is physical, the rent is upstream, and the place not to be is the capex itself or the application layer that sits on top of it.

The setup

For three years the binding constraint on AI infrastructure was the willingness to spend. That constraint is gone. Alphabet priced an $80B equity raise on June 2 — upsized to $84.75B gross, with Berkshire Hathaway taking $10B directly — despite running roughly $174B of annual operating cash flow T1. A company that generates that much cash chose to issue equity rather than slow its build. That is the signature of a Phase 2 capital cycle: capital is no longer rationed, it is racing.

The numbers are large enough to move the national accounts. The top nine cloud providers are guiding to roughly $725-830B of capex in 2026, up more than 60% year on year, the third consecutive year of 60%-plus growth, with about three-quarters of the spend pointed at AI rather than traditional cloud T3. AI-related capital spending now runs near 0.8% of US GDP and contributed roughly 1.1 percentage points to GDP growth in the first half of 2025; through the third quarter of 2025 the AI-related categories accounted for close to 39% of all US GDP growth T3. For two quarters, AI capex added more to US growth than all consumer spending combined T3.

When a single capex line is carrying the economy, the cyclical question stops being academic. The right tool is the one Marathon built for exactly this: read the supply side, not the demand side, because supply is what the capital is building and supply is what kills the returns T2.

The analysis

Phase identification — this is Phase 2, and it is not subtle

Marathon's five-phase cycle runs: returns above cost of capital, capital influx, capacity peak, capital retreat, returns restored T2. The capital-influx phase has a checklist of behavioral and financial markers. AI infrastructure ticks essentially all of them.

Phase 2 signal T2 AI infrastructure evidence, mid-2026
Capex / depreciation rising sharply Top-nine capex +60%+ YoY, third year running T3
Equity issuance accelerating Alphabet's $84.75B raise despite ~$174B operating cash flow T1
New entrants and adjacent capital Neoclouds, sovereign AI funds, private-credit data-center financing
M&A and multiples expanding High-bandwidth-memory suppliers re-rated to ~$1T caps; SK Hynix Q1 2026 operating margin 72% T1
Media and conference saturation AI is the dominant theme at every investor conference T2
Forecast overshoot normalized $1T+ 2027 capex now the consensus base case T3
Returns still reported above cost of capital Reported, yes — but see the depreciation question below

The last row is where the cycle hides. Phase 2 always looks like Phase 1 from inside it, because returns have not visibly compressed yet. The narrative cycle confirms the read: AI infrastructure sits in the acceleration-to-peak band, where disappointments are treated as buying opportunities and the most aggressive forecasts have become the consensus T2. None of this says the top is here. It says the gravity is now switched on.

The supply curve — what is actually being built, and how fast it can arrive

The demand-side debate gets the attention. The supply curve is where the money is at risk, and the supply curve has two physical chokepoints the capital cannot wish away.

The first is power. The International Energy Agency projects global data-center electricity demand roughly doubling from about 485 TWh in 2025 to roughly 945 TWh by 2030, with AI-optimized capacity more than quadrupling; US data centers are on course to account for nearly half of all US electricity-demand growth through 2030 T1. Lawrence Berkeley National Laboratory puts US data-center consumption at 4.4% of national electricity in 2023, rising to between 6.7% and 12% by 2028 T1. The grid cannot deliver that on the capex schedule. The combined large-load interconnection queue at the two largest US grids exceeds 600 GW against current combined peak demand near 240 GW, and historic interconnection-to-energization runs years, not quarters 2026-06-03-grid-interconnection-binding-constraint-ai-deployment. Power transformers run 128-week lead times T3.

The second is silicon, where high-bandwidth memory is the binding input and the rent has already migrated visibly to the few suppliers who can make it 2026-05-27-hbm-replaces-cowos-binding-constraint-inversion. SK Hynix earning a 72% operating margin in a memory business — a business whose historical identity is brutal cyclicality — is the cleanest single readout of how tight that link is T1.

Here is the cyclical irony. Both chokepoints are real, and both are why the build is slower than the capex. But a constraint that is slow to relieve is also slow to un-relieve. The same transformer and memory bottlenecks that throttle deployment today are drawing their own wave of capital — new fabs, new transformer plants, new turbine capacity. That capital is building the Phase 3 oversupply in the bottleneck layers, on the bottleneck layers' own clock. Memory is a 2-3 year cycle; transformers and generation are 3-7 year cycles T2. The rent in the bottleneck is durable on a 12-36 month horizon and self-liquidating beyond it.

The demand assumptions — what the buildout requires to be true

Strip the buildout down to the number it has to clear. David Cahn's framing at Sequoia, first published in 2024, took the silicon spend, doubled it for total data-center cost, doubled it again for a 50% end-user gross margin, and arrived at the annual AI revenue the build implicitly promises. The gap between that required revenue and observed revenue was $600B then, and on the current capex run-rate it is wider now, not narrower T2.

The revenue is growing, and this is the kernel of truth the bear has to respect. OpenAI reached roughly a $25B annualized run-rate in early 2026 with more than 900 million weekly active users T1. The hyperscaler clouds are inflecting: Amazon's cloud segment runs near a $150B annual rate, Google Cloud near $80B growing in the low-60s percent, Microsoft's AI services inside Azure growing triple digits off a smaller base T1. The destination is not in doubt. What is in doubt is the gap — the difference between roughly $700B-plus of annual capex and the AI revenue that has to service, depreciate, and earn a return on it.

The demand bridge, 2026 Approximate figure Tier
Top-nine 2026 capex $725-830B T3
AI-allocated share (~75%) ~$545B T3
Cahn-implied annual AI revenue required ~$600B+ and rising T2
Observed end-user AI revenue (industry, run-rate) Tens of billions, compounding fast T1

The bridge does not have to close in 2026. It has to close on a timeline the financing can survive. That is the entire lesson of the base rate.

The earnings question hiding inside the returns

The reported ROIC that keeps AI infrastructure looking like Phase 1 rests on a depreciation assumption. Hyperscalers depreciate AI hardware over five to six years T1. Michael Burry's public argument, made in November 2025, is that the economic life of a leading-edge GPU is closer to two or three years given each generation's step-change in performance per watt, and that the mismatch understates depreciation by roughly $176B across 2026-2028, overstating cohort profits by more than 20% — by his estimate near 27% at Oracle and 21% at Meta by 2028 T3. Nvidia rebutted, arguing observed utilization supports four-to-six-year lives T3.

Resolve this however you like — the point for the cycle is structural. If the bears are even partly right, then Phase 3 return compression is not a future event. It is already here, masked by an accounting choice. The capital cycle's gravity does not announce itself; it shows up first in the gap between reported returns and economic returns. The depreciation debate is that gap, made visible.

Variant perception — explicit

Consensus has hardened into a clean story: AI infrastructure is a secular demand build where supply is the only constraint, capex is investment in a durable future, and the buildout compounds long after the hype fades T3. In this framing the capex is not a risk; it is the moat.

AlphaSteve's variant view is narrower and older than the bubble-versus-no-bubble debate, which is the wrong axis. The variant is this: the surest outcome of $725B of capex growing 60% a year is mean reversion in the returns on that capital, and the base rates that govern the magnitude and timing of that reversion are not priced into the cohort. The market is pricing the demand. It is not pricing the supply curve the demand has to ride, the revenue gap the capex has to fill, or the depreciation the returns have to absorb. The bottleneck layers — high-bandwidth memory, power equipment, transmission — collect real rent while their constraints bind, and that rent is the durable trade. The hyperscaler capex line and the application layer that sits on it are where the asset-growth penalty lands T2.

What would falsify the variant and move AlphaSteve toward consensus. First, AI revenue inflecting fast enough that the Cahn gap visibly closes — industry end-user AI revenue compounding past the few-hundred-billion mark with improving unit economics, not just rising headline run-rates. Second, durable evidence that leading-edge GPU economic life genuinely is five to six years, which would validate both the depreciation schedules and the capital intensity. Third, utilization staying high across the cohort as capacity floods on, the way it never did in telecom — demand absorbing supply in real time rather than overshooting it.

What would confirm it. First, the first visible quarter where capex is booked but depreciation expense lags it, the accounting tell of capacity installed ahead of revenue. Second, a public slip on a flagship campus milestone (Meta's Hyperion or the Stargate Abilene ramp), the deployment tell. Third, AI revenue growth decelerating while capex guidance holds — the scissors that opened under every prior buildout.

Base rates — explicit

Mauboussin and Wang's discipline is to anchor any forecast in the distribution of analogous outcomes before reasoning about why this case differs T2. Three base rates bear directly on this theme.

Sales-growth persistence. Sales growth has low year-on-year correlation and reverts to the mean quickly; only about 1.5% of companies — 37 of 2,548 in Mauboussin's sample — sustained sales growth above 45% over a three-year window T2. The AI cohort's implied forward growth, embedded in current multiples, sits in the far-right tail of that distribution for an unusually large group of large-cap names at once. The base rate does not say it is impossible. It says it is rare, and that the consensus is underwriting the tail as the mode.

The asset-growth penalty. Across the US cross-section, firms with high asset growth earn abnormally low subsequent returns, and the effect holds for large caps — the very names leading this build T2. Corporate events that expand the asset base (equity offerings, debt offerings, heavy capex) are systematically followed by low returns; contraction events by high returns. Alphabet's equity raise to fund capex is, in this literature, a textbook setup for forward underperformance of that capital — not of the company necessarily, but of the marginal dollar deployed.

The buildout analogs. The cleanest historical match is the telecom and fiber build of 1996-2001. Builders spent on the order of $1T-plus laying roughly 80-90 million miles of fiber; by 2001-2002 an estimated 95% of it sat dark; the demand the builders had projected for 2001 arrived around 2008, a timing error of roughly seven years; and most of the companies that laid the cable went bankrupt before the demand showed up, with Global Crossing's early-2002 filing the emblem T2. The destination was right. The fiber was eventually needed and heavily used. The equity that financed it was destroyed in the gap between build and demand. The deeper analog is the 1880s-1890s railway mania: by the mid-1890s roughly a third of US rail mileage had passed through bankruptcy, and the lines were built, useful, and permanent — but the capital that built them was not T2.

The base-rate read is not "AI is the next fiber" as a slogan. It is the specific, repeatable mechanism of loss: in capital-intensive buildouts financed by abundant capital, the technology usually delivers, the timeline usually disappoints, and the rent usually accrues to the owner of the scarce input rather than the owner of the abundant capacity. That is the pattern to position against.

Buildout Capital deployed What went right What went wrong Where the loss landed
US railways, 1880s-90s Vast; ~$2.5B capitalization in receivership T2 Network built, permanent, used Overbuilt vs. near-term traffic; debt service failed Equity and bondholders of the builders
Telecom / fiber, 1996-2001 ~$1T+ T2 Fiber eventually essential ~95% dark; demand ~7 yrs late Builder equity wiped before demand arrived
AI infrastructure, 2023- ~$725B in 2026 alone T3 Demand real, compute genuinely useful TBD — supply curve, revenue gap, depreciation TBD — base rate says marginal capex

Implications for AlphaSteve

The capital-cycle and base-rate frame does not change the kit's positioning; it explains and hardens it. AlphaSteve's exposure to the AI infrastructure buildout is and remains zero at the layer the consensus loves — the hyperscaler capex line and the application layer — because that is exactly where the asset-growth penalty and the buildout base rate concentrate the loss. The kit's interest sits upstream, at the bottleneck layers where the constraint is physical and the rent is real while it binds, and the discipline is to collect that rent on a 12-36 month clock and exit before the bottleneck's own Phase 3. The single most important forward observable is the gap between reported and economic returns — the depreciation tell — because that is where the cycle will surface first.

  • Portfolio: no action. AI-infrastructure exposure stays zero at the buildout and application layers on margin-of-safety discipline. The bottleneck layers remain watchlist-only pending price.
  • Watchlist: the power-equipment and transmission names already flagged (GE Vernova, Quanta Services, Siemens Energy, MasTec, plus deregulated power in Vistra and Constellation) carry forward as the durable-rent expressions of this theme 2026-06-03-grid-interconnection-binding-constraint-ai-deployment. Add an explicit discipline: each is a cyclical rent, not a secular one — size and time-box accordingly. Memory (high-bandwidth) is the silicon analog and is already too re-rated to enter 2026-05-27-hbm-replaces-cowos-binding-constraint-inversion.
  • Theses on the workbench: GE Vernova thesis pass remains the live AI-infrastructure expression, to be underwritten explicitly as a capital-cycle rent with a defined exit, not a buy-and-hold secular compounder. MP Materials remains the top operational priority and connects to last week's critical-minerals dossier 2026-05-29-critical-minerals-capital-cycle-dossier-v1.
  • Sectors: Information Technology / Semiconductors, Industrials / Electrical Equipment, and Utilities with data-center load are all elevated as kit-watch — but watched as Phase 2 cyclicals, not as growth compounders.
  • House view updates: promote "AI infrastructure capacity" from candidate theme to dossier (v1) in the Themes section; record the capital-cycle phase identification, the three base rates, and the variant view. Detail in the next section.
  • Daily-scan adjustments: add a standing AI-infrastructure base-rate check — (i) hyperscaler depreciation expense growth versus capex growth; (ii) industry AI end-user revenue run-rate versus the Cahn-implied requirement; (iii) any flagship-campus milestone slip. Any one crossing threshold updates the dossier toward v2.

Charts / data

Table 1 (above) — Phase 2 capital-cycle signal scorecard for AI infrastructure. Table 2 (above) — the demand bridge: capex versus implied required revenue. Table 3 (above) — buildout base-rate comparison: railways, fiber, AI infrastructure.

All figures and tier tags carried inline. The scorecard and bridge are the two charts doing analytical work: the scorecard shows the cohort sits at the capital-influx phase on essentially every Marathon marker, and the bridge shows the revenue gap the capex has to close is widening on the current run-rate.

Sources

See sources-policy for the citation discipline applied. Wikipedia and anonymous aggregators surfaced in research were used only to locate the primary and Fed-research sources above and are not cited.

House view changes this run

  1. Promote "AI infrastructure capacity" in the Themes section from candidate-for-dossier to dossier (v1) — confirmed multi-year cross-sector theme, Phase 2 capital-cycle reading with explicit base-rate-grounded variant perception. Record: (i) phase identification — capital-influx phase on essentially every Marathon marker; (ii) three base rates — sales-growth persistence (1.5% sustain >45% 3-yr growth), the asset-growth penalty (Cooper-Gulen-Schill 2008), and the buildout analogs (telecom fiber ~95% dark, ~7-yr demand lag; 1890s railways ~1/3 mileage bankrupt); (iii) variant view — market prices the demand, not the supply curve / revenue gap / depreciation, and the durable trade is the bottleneck layers as cyclical rent, not the buildout.

  2. Add the dossier's standing forward observables to the theme entry: hyperscaler depreciation-versus-capex growth gap; industry AI revenue run-rate versus the Cahn-implied requirement; flagship-campus milestone slips. Any one crossing threshold updates toward v2.

  3. Reaffirm (no change) the existing sub-positions: HBM as the silicon bottleneck, grid interconnection as the deployment bottleneck, power equipment as the next-stage rent migration. The dossier places them inside the capital-cycle frame rather than altering them.

  4. Add to the changes log: "2026-06-05 Friday long-form (thematic): AI infrastructure capacity promoted to dossier v1 — Phase 2 capital-cycle identification, three base rates (sales-growth persistence, asset-growth penalty, telecom/railway buildout analogs), variant = own the bottleneck not the buildout; positioning unchanged, frame hardened."

Linked