A companion to the model live at ardeshir.io/capex.
The one result worth the whole model
Run the AI build-out backward. Cut lab demand to 40% of base — the “winter” scenario, sentiment turned, the narrative that AI was overbuilt now consensus. System capex falls roughly 78%. Ask what happens to hyperscaler depreciation and the intuitive answer is wrong: it falls about 3%, from $109B to $106B. Roughly 84% of depreciation is locked. So hyperscaler profit falls despite the spending cut, because the chips and buildings already bought keep amortizing against a revenue base that shrank with the demand they were built to serve.
That is the single non-obvious thing the model exists to say, and it follows from one distinction the rest of the market keeps collapsing: the AI build is a stock, not a flow. A flow unwinds when sentiment does — you stop writing checks and the outflow stops. A stock does not. The data center is on the balance sheet whether or not anyone believes in it next quarter, and it depreciates on schedule regardless. Anyone modeling “AI winter” as symmetric with “AI boom” has the physics backward. The build is easy to stop and impossible to unwind, and that asymmetry is where the damage lives.
The model is a what-if engine, not a forecast. Live data sets the sliders’ default positions; you move any of them. The output is never “what will happen.” It is “if this driver moves, who is exposed, by how much, and through which channel.”
The chain, top to bottom
The solver runs one direction, top to bottom, each stage consuming the one above it. Money enters at the top as AI-lab demand and lands at the bottom as revenue and profit on twenty-six named public companies, across five physical supply tiers — silicon, power and grid equipment, utilities and independent power producers, oil-and-gas midstream, telecom and fiber, and land and construction.
Pricing. Before any dollar moves, everything levered is priced. Each entity carries a rating bucket (AAA through B) and a hand-set issuer spread in basis points. The all-in cost of debt is the Treasury yield plus that issuer spread scaled by a live spread multiple driven off ICE BofA option-adjusted spreads by rating bucket. Ten-, twenty-, and thirty-year Treasury yields come live from FRED. When credit conditions tighten, the multiple widens the whole curve, and the layer computes each entity’s change in financing cost versus base. Credit quality is priced first because everything downstream inherits it.
Demand and capex carriers. Lab demand — OpenAI, Anthropic, the rest — scales from 1.0 at base to 0.40 in winter. Six hyperscaler carriers translate that demand into capex, and stickiness matters: Amazon, Google, and Microsoft, sitting on committed multi-year builds, cut less than the headline number implies. Oracle and xAI, who financed the build with debt and now face wider spreads, cut hardest. The system capex that survives is split across the five tiers by a calibrated allocation; silicon takes the largest absolute cut because it is the largest share.
The power dollar is zero-sum. Inside the power allocation there is a sub-budget that reallocates rather than grows. A gas-share tilt moves the same procurement dollar between gas-fired midstream (Williams, Energy Transfer, Kinder Morgan, Cheniere, EQT) and regulated utilities (NextEra, Dominion, AEP). Push gas share from a base of 0.25 to 0.70 and the midstream names rise while the utilities fall by exactly the dollar the midstream names gained. Same money, different pipe. At base the tilt is exactly 1.0, which is why moving it changes distribution without inventing demand.
Supplier revenue and profit. For each of the twenty-six, the revenue change is base revenue times AI exposure times the tier flow, damped by a backlog cushion. AI exposure is high for NVIDIA and a rounding error for Caterpillar at roughly 5%. The cushion is what a multi-year backlog does: GE Vernova with a five-year turbine backlog barely moves in a winter year while Vertiv, Talen, and merchant IPPs priced on AI load growth see revenue collapse proportionally. Then operating leverage does its work — a −40% revenue hit becomes a −60%-or-worse profit hit for fixed-cost names, and the BB/B-rated ones (Vistra, Talen, Lumen) take an additional refinancing hit from the wider spread. A supplier can be struck three times in one scenario: demand shock, higher input costs, and refinancing. Margins clamp at −0.30 so the model cannot manufacture absurdities.
The depreciation wall. Hyperscaler profit is computed as a delta off a stated base, not a rebuilt income statement, and the depreciation term is the piece that carries the whole result. Depreciation is last year’s cohort still amortizing plus this year’s capex phased in as it enters service. Cut this year’s capex to nothing and last year’s chips and buildings keep depreciating anyway. That is why capex can fall 78% while depreciation falls 3%, and why a naive model that sets depreciation equal to current capex over life produces the economically backward result that hyperscaler profits rise in an AI winter. The correct form is the one guarded by the model’s invariants.
Three scenarios, three different stories
The point of separating the channels is that the same-looking shock produces different casualty lists depending on where it originates.
Winter is a demand shock. Capex collapses, high-exposure suppliers with no cushion take the worst of it, and the depreciation wall means hyperscalers lose money while spending less.
Bondshock is a +100bp parallel shift with no change in demand at all. The dispersion here is driven by financing cost alone, so the ordering is pure credit quality: Oracle and xAI cut capex around 20%, Microsoft and Google around 12%. Nothing changed about the AI story — only the price of money — and everything downstream inherits that credit-quality ordering.
Gas pivot moves the gas share from 0.25 to 0.70 with total demand fixed. Midstream and gas-fired names rise; regulated utilities fall by the identical procurement dollar. No new demand is created; the money is reallocated. It is the cleanest illustration that “gas is winning” and “AI spending is growing” are different claims that the model refuses to conflate.
Why the numbers are trustworthy: the invariants
The model earns its output through six invariants it must satisfy before any result is read. These are the reason the thing is an analytical instrument rather than a slider toy.
The first says the base reconciles to zero — every supplier reads exactly 0.00 change at base, so any departure from zero is caused by something you moved and nothing else. There is no hidden bias to launder a narrative through. The second encodes the stock/flow distinction directly, which is what forces the depreciation wall to behave correctly instead of the backward way. The third makes credit quality dominate the risk-free rate, so who-cuts-first is decided by balance-sheet strength. The fourth makes the power procurement budget zero-sum, so tilts reallocate and never conjure demand. The last two are epistemic guards: the model cannot produce absurdities, and it must fail loudly rather than silently.
What is filed and what is a judgment call
The model is disciplined about the difference between a fact and an estimate, and it labels every field as one or the other.
Filed, hard figures: Treasury yields from FRED daily; credit-spread levels from ICE BofA OAS by rating bucket; hyperscaler and supplier fundamentals — capex, operating income, revenue, depreciation, property and equipment, interest expense, long-term debt — from SEC EDGAR XBRL company facts, pulled from annual 10-K frames with the resolving XBRL tag recorded per field. Fiscal years are carried through honestly rather than coerced into calendar years, so Oracle’s May, NVIDIA’s January, and Microsoft’s June year-ends stay themselves. Wholesale power and Henry Hub come from the EIA API, and planned generation capacity by fuel anchors the gas-share default. Realized 60-day rolling betas come from daily closes, placed next to the model’s own betas so the model can be wrong in public.
Estimates to challenge, flagged as such: AI exposure and backlog cushion for every supplier are judgment calls and will stay judgment calls. Lab spend from OpenAI, Anthropic, and others is press-reported commitment, never audited, and permanently flagged as estimate — neither lab files with the SEC, and nothing in the pipeline presents their numbers as anything firmer. Issuer-specific spreads are hand-set; only the bucket level is live. The capex split across the five tiers is calibrated from disclosures and industry estimates. Foreign private issuers with no US-GAAP XBRL — Schneider Electric, Siemens Energy — and private companies like SpaceX, xAI, OpenAI, and Anthropic are never presented as filed.
The model also states plainly what it does not do. There is no second-order feedback: its output does not move market prices that then move its inputs. There is no supplier-of-supplier depth. It is a one-period flow-of-funds model, not a general-equilibrium simulation, and it is more useful for being honest about that.
Use it
The model is live and interactive at ardeshir.io/capex. Move a slider, re-solve, and read the exposure map. The right way to use it is not to hunt for a prediction but to interrogate a channel: pick the shock you actually fear — demand, credit, fuel mix — and watch which of the twenty-six names it reaches, by how much, and through which balance sheet. The value is in the causal ordering, not the point estimate. When the market is pricing an AI winter, the question is never whether spending falls. It is who is holding the locked stock when it does.
Part of the Univrs research ecosystem:
- Research: metalearn.org · ardeshir.io · sepahsalar.org · univrs.io
- Creative: imagine.univrs.io
- Code: github.com/univrs · github.com/ardeshir