The Capex Mirror: How the AI Capex Cycle Inflates S&P 500 Earnings by 15%

Tue, 14 Jul 2026 00:00:00 GMT

The Capex Mirror

When a hyperscaler spends $1.00 on AI hardware, the seller recognizes the revenue immediately. The buyer records only a small part of it as first-year depreciation. This timing asymmetry is a standard feature of accrual accounting and has existed for decades. Summed across the market, it means an investment boom creates reported earnings while it runs, and removes them later.

The 6.8¢ the buyer charges in the spend year is not a waiver. It is a schedule. The other 93.2¢ is still owed to the income statement, and Exhibit 2 shows when it arrives.

Three implications follow.

First, the mechanism is the firm-level expression of the Kalecki–Levy profits identity:

Corporate profits ≡ investment + dividends − household saving + government deficit + net exports

Investment raises aggregate profits because one company’s capex is another company’s revenue, while the spender books an asset instead of an expense. A capex boom can therefore lift measured corporate profits before the new capacity generates any end demand.

Second, the effect should reverse over the life of the asset if the investment earns an adequate return. The roughly 77¢ deferred at purchase is recognized later through depreciation, and must be covered by the revenue the asset generates. The boom-time earnings are not fictitious, but they are pulled forward from the future. If the asset underperforms, the adjustment appears through weaker margins, accelerated depreciation, or impairment.

Third, the size of the distortion depends on three factors: the growth rate of capex, the seller’s profit margin, and whether both buyer and seller sit inside the same index. Once capex stabilizes, depreciation catches up and the wedge narrows. In the current AI cycle, all three factors are unusually large. Section 4 measures each.

Tracing the 2026 flows

Where does the money come from, where does it go, and what does each ledger record?

The mechanism is not new, and it has always been negligible: capex was small, and the companies selling the equipment were low-margin industrials. What changed is scale. Big-5 capex has grown eightfold since 2019, the sellers are now the highest-margin businesses in the index, and both sides sit at the top of the same benchmark. The question is how much.

How much are index earnings inflated?

Three sources are elevating index earnings: boom-dependent seller profits, private-stake mark-ups, and understated depreciation. Together we estimate ~$326B, about 15% of the $2.24T the S&P 500 reports on a trailing basis, and roughly 23% of the NASDAQ-100’s. In a normal year all three round to zero.

Two independent checks support the estimate.

Goldman’s attribution. AI-infrastructure beneficiaries account for roughly half of the S&P 500’s expected +24% EPS growth in 2026, about $290B of aggregate operating earnings growth in one year from the supply side of the build-out. That is consistent with a boom-linked level in the $250–400B range.

The macro check. AI investment rose roughly $550B above its 2019–23 trend. The profits identity says aggregate profits rise nearly dollar-for-dollar with investment, before household-saving offsets. NIPA corporate profits did rise ~$500B over the same window, while the non-AI economy decelerated. The boom is arithmetically most of current U.S. profit growth.

The $149B of seller profit should be read as an upper bound. Some of it would survive a capex plateau, because fleets still have to be refreshed, so not all of it would disappear. Section 6 models the reversal on that basis.

How anomalous is this?

The asymmetry is old. The magnitude is not. Every investment boom inflates concurrent profits, and what separates episodes is the product of three multipliers: capex growth, seller margin, and index co-residence. 2026 is the first episode to max out all three at once.

Four things separate 2026 from 2000, in order of importance.

Seller margins. Nvidia converts ~52¢ of every revenue dollar into net profit. Lucent converted ~4¢. The same capex dollar now generates an order of magnitude more concurrent reported profit than in any previous build-out. This, more than the scale of the spending, is what makes the current cycle unique.

Index co-residence, at the top. In 2000 the buyers were mid-index carriers and the sellers were concentrated in one giant, Cisco, at ~4% of the index. Today the buyers are ~16% of the S&P 500 and the sellers ~13%. Nvidia is the largest weight in the index at 7.3%, and Micron has just displaced Berkshire from the top ten. The index books both sides of the same internal transaction at maximum weight, and passive flows, now over 54% of assets under management, price neither side.

Scale. Big-4 capex alone is 2.3% of GDP, roughly double the carriers’ 2000 peak.

The mark-up loop, which 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 up through public income statements.

One offset is also new. Unlike the 20-year fibre lives of 2000, today’s 5.5-year assumptions are at least the right order of magnitude, and Amazon has already begun shortening. The accounting debate is happening in daylight. Against a steady-state economy the calibration is still stark: the wedge would be near zero and boom-dependent profits near zero. Against 2000, the previous record, today’s effect is roughly 2x on the wedge channel, 4x on the seller channel, and 5x per dollar spent.

Why we doubt the pace can hold

Everything above establishes that earnings are elevated while capex grows. The conclusion therefore rests on one question: can the growth continue? The tightest constraint we see is not power, chips, or demand. It is that the funding requirement compounds faster than any source of funds we can identify.

Two measured exponentials are in collision. Buyer operating cash flow is growing ~23% a year while capex grows ~70% a year, and the curves crossed this quarter. Project even a decelerated pace of +50% a year, consistent with Morgan Stanley’s $1.16T estimate for 2027, hold operating cash flow at +23%, and keep the buyers’ ~$200B a year of dividend and buyback commitments. The external funding requirement then explodes: $316B in 2027, $665B in 2028, and $1.26T in 2029.

That money has to come from somewhere, and three constraints stand in the way.

We doubt the bond market can absorb it. Total U.S. investment-grade gross issuance runs ~$1.65T a year, across every bank, utility, industrial and pharmaceutical company combined. The ~$570B projected for AI in 2026 already claims a third of it, and Morgan Stanley is warning the supply will weigh on credit performance. Holding the pace would take ~40% of the entire market by 2028 and ~75% by 2029, at spreads that would not resemble today’s.

The interest bill would exceed the industry’s revenue. The financing stack needed to hold the pace reaches ~$2.5T by 2029. At 5.5%, generous with the 10-year at 4.5%, that is ~$139B of interest a year against AI end-revenue of $90–120B — a figure that is itself flattered by contracted-revenue accounting. The coupon has to be paid out of a revenue line that does not yet cover it.

We see nothing else large enough to fill the gap. Private credit and sovereign wealth offer hundreds of billions, not trillions. Vendor financing is already counted, and is the circularity documented in the companion note. Equity issuance, including the labs’ ~$3T IPO ask, meets the same constraint.

The conclusion needs no view on AI’s merits: we think the capex growth that manufactures today’s earnings is unlikely to be sustained, on funding arithmetic alone, beyond a 2027–28 horizon.

How it unwinds

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

The two paths differ in who pays, and how fast. Neither assumes the AI revenue never arrives. If it does, it offsets the depreciation and the drag shrinks. The asymmetry is in the certainty: the depreciation is already committed, and the revenue is not.

On a plateau, the hyperscalers absorb it, slowly. Amazon, Microsoft, Alphabet, Meta and Oracle have already paid cash for these assets, so nothing changes at the bank. But ~$330B of deferred cost reaches their income statements between 2027 and 2030, taking 3–4 points a year off index EPS growth.

On a decline, the sellers absorb it, quickly. Roughly $75B of net income disappears from Nvidia, Micron, Broadcom and the equipment makers within two years. The mark-up line at Alphabet and Amazon turns from an ~$85B tailwind into a headwind. And the hyperscalers’ depreciation keeps rising regardless, because those assets are already on the balance sheet. The combined ~$220B, about 10% of index earnings, lands inside two years, before any impairments.

Conclusion: the equity math is challenging

Earnings are overstated, and the valuation is at a record. On reported earnings the S&P 500 trades at 30x, against a long-run average near 16x. But $326B of those earnings are manufactured by the build-out. Strip them out and the same $67.19T of market value sits on $1.91T — a P/E of ~35x, above the ~30x of the 2000 peak. Measured on earnings that survive the capex cycle, the index has never been more expensive.

Breaking even would require revenue the industry does not have — even on the friendliest path. Note what a plateau does not do: it does not take the seller profits away. Capex holds at ~$775B, the chipmakers keep shipping, and the $149B stays. What unwinds is the deferred cost. Depreciation catches up with the spending: annual D&A rises from ~$262B in 2026 to ~$596B by 2030, so the accrual tailwind becomes a headwind that compounds — about $78B of extra expense in 2027, and ~$334B a year by 2030.

Offsetting it would take that much in incremental profit. At a 30% net margin, that means ~$260B of new revenue by 2027, rising to ~$1.1T a year by 2030. And it has to be truly new: over and above the $90–120B the AI industry earns today, and over and above the growth already priced into OpenAI’s and Anthropic’s valuations.

The math for equities is therefore very challenging. A record multiple, on earnings inflated by 15%, with a 3–4 point annual EPS drag arriving in 2027–28 — or ~10% at once if capex declines — landing on the ~29% of index weight that the buyers and sellers hold between them, with no offsetting sector to absorb it.

Appendix A: Notes to Exhibits 1–3

The accounting assumptions behind the three exhibits that carry the mechanism. Every other exhibit’s note sits with the chart itself.

Appendix B: Inputs and sources

Appendix C: Glossary

The accounting terms used throughout this note, in the order they matter.

Research commentary, not investment advice. Figures as retrieved July 2–8, 2026. All stated assumptions are in the exhibits so that they can be attacked line by line.

The Practice Room: Guitar Theory Drills

Mon, 13 Jul 2026 00:00:00 GMT

Theory isn't knowledge, it's latency. Everyone knows the vi chord is minor; almost nobody can name it in E♭ before the bar has gone past. That gap wants a drill, not a paragraph — so this page is the drill. It runs in the browser, keeps nothing, and works on a phone.

Two to start, chosen because fluency in either changes what you can play tonight.

Notes on the bass strings

Strings 6 and 5 only, frets 0 to 12. These are the two strings barre chords are rooted on, so knowing them turns one shape into all twelve chords. The rest of the fretboard can wait.

Numbers in a key

The seven chords of a major key, by number, in both directions — chord to numeral and back. This is what lets someone call I–V–vi–IV in D and have your hands already moving, and it's the same fact that makes transposing and capo arithmetic possible later.

Interval shapes from a root, key identification, transposition and capo. Sections here rather than posts of their own.

0xjeffery