Trade the AI Capital Cycle, Not the AI Theme
Our thesis is simple and unfashionable: there is no longer such a thing as "the AI trade" as a single expression, and continuing to trade it as one is the primary source of risk in the book. What began as a narrative-driven re-rating — buy the obvious names, size up, hold — has matured into something structurally different: a capital cycle, with all the machinery that implies. Order books, lead times, depreciation schedules, financing mix, capacity additions, and eventually overcapacity. Capital cycles are tradable with far more precision than themes are, because they emit signals in a known sequence. The correct way to trade AI now is to identify which physical constraint is currently binding, own the holder of that constraint while its pricing power is inelastic, and separately underwrite who is financing the spend and with what. Everything else — thematic baskets, "AI exposure" screens, broad index proxies — is undifferentiated beta dressed as a view.
The phase shift: from multiple expansion to capex accounting
The first leg of this trade was paid out in multiples. Expectations re-rated faster than earnings, and almost any exposure worked because the dominant variable was narrative adoption, not delivered cash flow. That regime rewards breadth and punishes selectivity — which is exactly why so many books are still constructed for it.
The current leg is being paid out, or not paid out, in accounting. When a buildout of this scale runs through income statements, the relevant questions become mechanical rather than visionary: How much of the spend is funded from operating cash flow versus debt, vendor financing, leases, or off-balance-sheet vehicles? What useful life is being assumed on compute assets, and is that assumption being extended as pressure on reported earnings grows? How much of a supplier's revenue is effectively circular — funded by an equity stake, a prepayment, or a guarantee from the same counterparty booking the sale?
These are not abstract governance concerns. They change the identity of the instrument you are holding. A company funding capex out of internally generated cash is an equity story with a return-on-invested-capital question attached. A company funding capex with layered debt and long-dated offtake commitments is a credit story with an equity wrapper — its stock will trade off funding spreads and refinancing windows long before it trades off inference demand. Those two assets do not belong in the same sleeve at the same weight, and conflating them is how thematic books discover their real factor exposure at precisely the wrong moment.
Own the binding constraint, and rotate when it moves
The most durable structure in a capital cycle is the choke point: the input where demand is inelastic, substitution is slow, and supply cannot be added inside the investment horizon of the buyers. Pricing power concentrates there and nowhere else. The discipline is to keep asking which constraint is binding today, and — separately — which one the market has already paid for.
The constraint migrates in a recognizable order. It starts at the most specialized silicon, then moves upstream into packaging and memory integration, then out of the fab entirely and into the physical world: interconnection queues, transformers and switchgear, turbine and generation capacity, transmission, water and thermal management, and finally the skilled trades who assemble and commission all of it. Each handoff is visible in the same telltales — quoted lead times stretching, prepayments and take-or-pay contracts appearing, customers signing multi-year commitments at prices they would have rejected a year earlier, and incumbents suddenly able to raise price without losing volume.
Three operating rules follow:
- Prefer the constraint with the longest physical build time. Anything that can be debottlenecked by a shift pattern or a purchase order gets debottlenecked. Anything requiring permitting, heavy fabrication, or a grid interconnection cannot, and holds margin longer.
- Prefer the constraint where the buyer's alternative is idle capital. Elasticity collapses when the counterfactual is a half-built facility earning nothing. That asymmetry is the source of the pricing power, and it decays the moment the queue clears.
- Exit on supply response, not on valuation. Choke-point assets almost always look expensive while the constraint binds and cheap as it releases. The sell signal is capacity coming online and lead times compressing, not a multiple.
Your "diversified" AI basket is one position with several tickers
The structural error in most AI exposure is the assumption that spreading across the stack diversifies the bet. It does not, because the layers are contractually and financially interlocked. A hyperscaler's capex is a supplier's revenue; that supplier's equity stake or prepayment is partly what funds the customer; a neocloud's debt is collateralized by the very hardware whose residual value depends on the pace of the next product cycle. Hold all of them and you hold one exposure — the durability of a single capex stream — with the false comfort of a position count.
Two practical consequences. First, measure exposure in terms of the underlying driver, not the sector label: how much of the book's P&L moves if aggregate AI capex guidance is revised down, regardless of which layer the names sit in. Second, treat realized correlation among these names as a position in itself. When a theme is crowded and financially circular, correlation spikes in drawdowns — the diversification you were relying on disappears exactly when it is needed, and net exposure turns out to be far higher than gross-adjusted risk models suggested.
The better expression is relative value and dispersion, not direction
If the dominant feature of this phase is that the layers of the stack are being repriced against each other — cash-funded versus debt-funded, constrained versus commoditized, pricing power versus volume growth — then the natural expression is relative value inside the stack rather than directional exposure to the theme.
That argues for pairing: long the holder of the binding constraint against a basket of the same theme's beta; long the layer with pricing power against the layer absorbing the depreciation; long the input that cannot be manufactured faster against the one that can. Pairs neutralize the part of the return you have no edge on — whether the theme is in favor this quarter — and isolate the part you actually analyzed.
It also argues for taking the view in volatility terms. Single-name implied volatility across this complex is generally rich, and skew on the crowded leaders is habitually bid; selling that convexity naked is a poor trade because the tail is real. The more defensible structures are the ones that are long dispersion — that is, long the idea that individual names will diverge sharply while the index moves less — because divergence is precisely what a maturing capital cycle produces as winners and funders separate. Structures that finance long optionality by selling something correlated but less rich, and calendars that respect the fact that this complex now repriced around a small number of scheduled catalysts, are more honest ways to hold the view than buying outright upside after a run.
The hedge is not "short AI" — it's short the funding-dependent second derivative
Shorting the theme outright has been an expensive way to be eventually right. Secular demand is real, the leaders generate cash, and negative carry compounds while you wait. The efficient short is not the technology; it is the part of the structure that requires continuous access to capital markets to survive. Entities whose capex exceeds operating cash flow, whose asset base depreciates faster than their contracts amortize, and whose equity value is a thin residual on top of secured debt are levered to funding conditions rather than to adoption. They fall on a widening credit spread, not on a disappointing model release.