The Capex Mirror: Concurrent Earnings Inflation in the AI Build-Out

The mechanism, stated precisely

One transaction contains the whole phenomenon. A hyperscaler pays $1.00 for AI hardware. The seller’s income statement recognizes the dollar now; the buyer’s recognizes almost none of it now. Nothing about this is improper — it is exactly what accrual accounting prescribes — but summed across the market, it means an investment boom creates reported aggregate earnings contemporaneously, and removes them later.

Three formal points sharpen it. First, this is the micro anatomy of a macro identity: in the Kalecki–Levy profits equation, aggregate corporate profits ≡ investment + dividends − household saving + government deficit + net exports. Investment is a source of aggregate profits precisely because one firm’s capex is another firm’s revenue while the spender books an asset instead of an expense. A capex boom therefore raises measured aggregate profits mechanically, before a single unit of end-demand materializes. Second, the effect nets to zero over the asset’s life if the asset earns its keep: the deferred ~77¢ arrives as depreciation in 2027–31 and must be covered by revenue the asset generates. The boom-time earnings are thus not fictitious; they are borrowed against the asset’s future utility. If the utility disappoints, the borrowing is repaid as margin compression or impairment. Third, the aggregate distortion scales with three factors — the capex growth rate (steady-state capex creates no wedge because D&A catches up), the seller margin (how much of each dollar becomes immediate profit), and the index co-residence of both counterparties (whether the same benchmark books both sides). Section 4 shows all three are at historical extremes simultaneously.

Tracing the 2026 flows

Question one: where does the money come from, where does it go, and what does each ledger record?

The tracing answers the first question with a number worth memorizing: in 2026 the system books roughly $250–265B of immediate seller profit against roughly $54B of buyer expense recognition on the same spend — a five-to-one contemporaneous asymmetry — while the funding for that spend increasingly comes not from the buyers’ own free cash flow (now ~zero) but from bond markets and recycled venture equity. Aggregate reported earnings rise on both sides of the transaction at once: the seller’s P&L records the profit, and the buyer’s P&L is barely touched.

How much index-level inflation, today

Question two requires defining “inflation” carefully, because three distinct things are true at once and they should not be added carelessly. Each is quantified below against the S&P 500’s measured aggregates: $67.19T of market value (June 30) and ~$2.24T of trailing GAAP earnings (trailing P/E ~30, divisor 8.98B).

Three cross-checks discipline the estimate. Goldman’s own attribution: AI-infrastructure beneficiaries account for roughly half of the S&P 500’s expected +24% 2026 EPS growth — about $290B of aggregate operating earnings growth in one year from the build-out’s supply side, consistent with a boom-linked level in the $250–400B range. The Kalecki macro check: AI investment rose roughly $550B above its 2019–23 trend; the profits equation says aggregate profits rise nearly dollar-for-dollar with investment before household-saving offsets, and NIPA corporate profits indeed rose ~$500B over the same window while the non-AI economy decelerated — the boom is arithmetically most of current U.S. profit growth. The market’s own behavior: H1’s pattern of record earnings met with flat-to-negative reactions (Meta −9% on a raise; Microsoft −2.5% and Meta −6% after April beats) is what one expects if investors partially see through the accrual component. On the NASDAQ-100, where the AI complex is a larger share and financials/energy dilute nothing, the same items are roughly ~23% of index earnings. A final necessary honesty: part of definition B would survive a capex plateau (fleets must be refreshed — replacement demand is not zero), so “boom-dependent” overstates the fragility somewhat; the reversal scenarios in §5 handle this properly rather than assuming profits go to zero.

How anomalous is this? The asymmetry is old; the magnitude is not

Question three. Capitalization-versus-expensing asymmetry has existed since accrual accounting; every investment boom inflates concurrent aggregate profits. What distinguishes episodes is the product of three multipliers — capex growth, seller margin, and index co-residence — and 2026 is the first episode in the data to max out all three.

The structural novelties, ranked by importance. (1) Seller margins. Nvidia converts ~52¢ of each revenue dollar to net profit; Lucent converted ~4¢. The same capex dollar generates an order of magnitude more concurrent reported profit than in any prior buildout — this, more than capex scale, is what makes 2026 unique. (2) Index co-residence at the top. In 2000 the buyers (carriers) were mid-index and the sellers concentrated in one giant (Cisco, ~4%). Today the buyers are ~16% of the S&P and the sellers ~13% — Nvidia is the largest weight (7.3%) and Micron just displaced Berkshire from the top ten — so the index books both sides of the internal transaction at maximum weight, and passive flows (>54% of AUM) price neither side. (3) Scale. Big-4 capex alone is 2.3% of GDP, roughly double the carriers’ 2000 peak. (4) The mark-up loop is new in kind: in 2000, no carrier booked its vendor’s private-round appreciation as income; in 2026, $53B of one quarter’s platform “other income” is the private AI cohort marking itself through public P&Ls. (5) One partial offset is also new: unlike 2000’s 20-year fiber fictions, today’s 5.5-year lives are at least the right order of magnitude, and Amazon has begun shortening — the accounting debate is happening in daylight. Against normal times the calibration is stark: in a steady-state economy the wedge is ~0 and boom-dependent profits are ~0; against the 2000 episode — the previous record — today’s effect is roughly 2x as large on the wedge channel, 4x on the seller channel, and ~5x per dollar spent.

Why the pace cannot hold: the funding arithmetic

The preceding sections establish that current earnings are elevated while capex grows. The conclusion therefore rests on one question: can the growth continue? The decisive answer is not power, chips, or demand — it is that the funding requirement compounds faster than any available source of funds.

Two measured exponentials are in collision: buyer operating cash flow is growing ~23%/yr while their capex grows 70%/yr (Epoch AI, from filings) — the curves crossed this quarter. Project even a decelerated pace (+50%/yr, consistent with Morgan Stanley’s $1.16T 2027) against OCF at +23%, keep the buyers’ $200B/yr of dividend and buyback commitments, and the external funding requirement explodes superlinearly: **$316B in 2027, ~$665B in 2028, ~$1.26T in 2029**. There is no market large enough on the other side of that trade. Total U.S. investment-grade gross issuance — every bank, utility, industrial and pharma combined — is roughly $1.65T a year; 2026’s ~$570B of AI-related issuance already claims a third of it, and Morgan Stanley is already warning that this supply will weigh on credit performance. Holding the pace would require the AI complex to absorb ~40% of the entire IG market by 2028 and ~75% by 2029, at spreads that would no longer resemble today’s.

And debt cannot rescue it, for a reason internal to this note’s subject: debt requires debt service, and the end-market cannot yet cover the coupon. The cumulative financing stack needed to hold the pace reaches ~$2.5T by 2029; at 5.5% — generous, in a hiking-Fed regime with the 10-year at 4.5% — the interest bill alone is ~$139B a year, exceeding the entire current end-revenue of the AI industry (labs plus incremental AI cloud services: ~$90–120B, itself CARR-inflated). A financing structure whose annual interest exceeds the financed industry’s annual revenue is not a bridge to profitability; it is the textbook definition of Minsky Ponzi finance, and credit markets price that recognition abruptly, not gradually. The remaining escape routes are small against the need: private credit and sovereigns (hundreds of billions of capacity, not trillions), vendor financing (already counted — it is §7’s circularity in the companion note), and equity issuance (the ~$3T IPO ask is the labs’ version of the same constraint). The conclusion follows without any view on AI’s merits: the capex growth rate that currently manufactures the earnings must decelerate on funding arithmetic alone, on a 2027–28 horizon — and when it does, every channel quantified in §3 reverses by the mechanics of §6. Current index earnings are therefore not merely flattered; they are flattered by a force with a visible expiry.

What unwinds it, and on what schedule

The inflation is self-extinguishing by construction — the only question is which of two paths does the extinguishing.

The monitoring list follows from the mechanism, in causal order: (i) the capex second derivative — guidance revisions at the April/July/October prints, since the wedge lives on growth, not level; (ii) depreciation-life convergence — each company that follows Amazon’s shortening converts deferred cost to current expense with no cash change (Microsoft’s two-thirds short-lived disclosure makes it the bellwether); (iii) the OCF/capex crossings Epoch has dated (aggregate Q3 2026; Alphabet early 2027, Meta mid-2027, Microsoft late 2028) — each crossing forces the debt market, not the equity market, to underwrite the wedge; (iv) bond spreads on the ~$570B of projected 2026 AI issuance — the accrual game ends when creditors reprice it; (v) the mark-up line at Alphabet/Amazon, which turns from tailwind to headwind the first quarter private AI marks fall; and (vi) impairment language — the first hyperscaler 10-Q that writes down a GPU vintage converts Definition C from analyst estimate to reported fact.

Appendix: sources and assumptions

Research commentary, not investment advice. Figures as retrieved July 2–8, 2026. The framework is deliberately reductive where reduction clarifies: all stated assumptions are in the exhibits so they can be attacked line by line.