Building AI on borrowed money — a prisoner's dilemma

Special Report: Building AI on Borrowed Money — a Prisoner's Dilemma

2026-09-25 · Macro Analysis

> ⚠️ This is a macro analysis, not a forecast, and not investment advice. Sources and the full disclaimer are at the end.

---

Bottom line

Capital spending at the five big buyers now consumes nine-tenths of their operating cash flow. The buffer is essentially gone. Every dollar from here has to come from borrowing, from operating cash flow growing into it, or from spending slowing down.

The first of those is what is happening. Meta's total debt is up 126.6% in a year, Amazon's 57.5%, Oracle's 47.9% — while Microsoft's has fallen.

In the short run this genuinely lifts the index: borrowed money buys equipment, the equipment purchase is the seller's revenue and profit, and that profit lands in the index. Nvidia alone carries 8.19% of the S&P 500, and its own capex takes just 11.1% of its operating cash flow. It bears none of the construction burden; it collects.

The cost is booked on the buyers' balance sheets, and it has started to reach their income statements.

Whether this stays a sector problem or becomes a systemic one depends on where the leverage sits. The visible layer is high quality — investment-grade companies' own balance sheets. The real risk is in the layer no statement reaches: off-balance-sheet structures, and a private-credit market where borrowers pay interest by borrowing again.

How fast the cash buffer disappeared

Capex as a share of operating cash flow, on a consistent four-company basis (Alphabet, Meta, Microsoft, Oracle):

USD billions. 2025Q3 through 2026Q1 cover five companies; the rest cover four (Amazon has not reported its latest quarter, and Microsoft and Oracle run offset fiscal years). The year-on-year comparison is recomputed on the identical company set: 2026Q2's 90.3% sits against 79.0% for 2026Q1.

What is left after capex fell from more than fifty billion dollars a quarter to thirteen, in four quarters.

By company, latest reported quarter:

USD billions. Nvidia is here as the contrast: it is the side being paid.

Not everyone is borrowing

Pricing "the AI giants" as one sector misses the most important split in this table.

Microsoft is carrying it with its own cash flow and paying debt down; the other three are levering up. One story, three very different balance-sheet outcomes.

Oracle has taken the less conspicuous route: capital leases went from $14.1bn to $30.6bn in a year, up 117%, the fastest of the three. But over the last two quarters all three have shifted from leases toward straight debt — Amazon's non-lease borrowings are up 123% in a year, Meta's 189%. What is buried is growing; what is in plain sight is growing faster.

The income statement has started paying

The change in interest expense says more than the debt balance does.

Oracle's quarterly interest went from $1.06bn to $1.43bn, up 35% in a year, while operating income fell from $7.4bn to $7.0bn. Interest coverage dropped from 7.0x to 4.9x. Its debt equals 3.32 times annualized operating cash flow — the only company in this group whose coverage has fallen into single digits.

Meta's absolute leverage is still low, but it is deteriorating fastest: quarterly interest went from $0.23bn to $0.78bn, up 239%, while operating income fell from $21.9bn to $19.5bn. Coverage went from 96x to 25x in four quarters.

Spending rising, interest rising, operating income falling — that combination is showing up on the way up the construction curve, not on the way back down it.

Why nobody stops first

The payoff structure settles it. If everyone slows together, they all keep their cash flow and the capability gap is unchanged. If one slows while the others do not, the one that slowed is priced as falling behind immediately. So regardless of what anyone else does, continuing to spend is each company's dominant move.

The data fits that shape. Over the same stretch in which capex went from 68.7% to 90.3% of operating cash flow, Oracle's coverage fell from 7.0x to 4.9x and Meta's from 96x to 25x. Marginal conditions are deteriorating while spending accelerates. That is not what disciplined investment looks like.

The harder layer: the company least able to stop is the one that most needs to. Oracle's valuation rests on a growth story — slow down and the story breaks first. Microsoft has the largest cash engine and can outlast the race without ever conceding. The one that could stop does not want to; the one that wants to cannot afford to be seen doing it.

Which implies: no board will call this race off, so the end has to come from outside — from the financing market.

The competing explanation deserves stating. They may be seeing returns that outside statements cannot. Public data cannot falsify that. But if the returns were that certain, the company with the most cash and the least need to borrow — Microsoft — should be spending hardest, not most cautiously.

Financing costs are already moving; credit has not repriced

Over the last twenty sessions the 10-year Treasury yield rose 52 basis points to 5.18%, a z-score of 3.17; the 2-year rose 68 basis points to 4.87%. Over the same window the high-yield spread sits at 2.8% and widened 17 basis points, while the investment-grade spread is 0.79% and has not moved at all.

Two things follow. First, the rising cost of this debt comes from the risk-free rate, not from a credit premium. Second, credit markets have already begun to separate: the weak end is widening while the investment-grade end, where these companies sit, has not budged. The market is pricing weak credit. It is not yet pricing this buildout.

Price is one half; quantity is the other. The Fed's balance sheet stands at $6.75trn, up $16.8bn over four weeks, and bank reserves at $2.93trn, up $5.3bn. Money is not draining; the level is still edging up.

But the cushion is gone. Overnight reverse repo balances are down to $0.6bn. For three years that account absorbed money-market tightness first, leaving reserves untouched. With it near zero, the next squeeze lands directly on reserves, with nothing in between to take it.

For a buildout that runs on continuous issuance, this matters more than the size of the balance sheet: how open the financing window stays depends on how much cushion is left, not on how large the balance is.

The system's own risk read is "rates/inflation-driven risk-off (bonds sell too, not a haven)", with a systemic-stress score of 6.8/100 — not a crisis shape, a rates shape. And a rates shape is precisely the one a debt-financed buildout has most to fear.

Where it breaks first

The order matters more than the conclusion.

First, financing conditions change before demand does. Spreads widen, or an issue gets repriced. The tell is in investment-grade spreads, not high-yield — these are investment-grade issuers.

Second, the company least able to stop is tested first. Oracle's type, and the smaller compute providers whose borrowings rest on leases and pre-sold contracts. Microsoft is untouched; it never borrowed.

Third, the slowdown becomes the seller's revenue gap directly. The chain runs: capex guidance cut → chip orders → memory spot prices → power and data centers.

The tail of that chain is already falling: the data-center ETF (DTCR) has seen 30-day net outflows worth 2.62% of its assets, and the uranium ETF (URA) trades 5.46% below its 50-day average with a relative-strength reading of 40.3. The front of the chain is still hot: the semiconductor ETF (SOXX) took in 7.85% of assets over 30 days, CHPS 6.77%. Money is leaving the far end of this chain while pouring into the near end.

Fourth, collateral reflexivity. Where the debt is secured on long leases and pre-sold contracts, a slowdown erodes the collateral that justified the borrowing in the first place. Twenty-year data-center leases of exactly this kind already appear in the reports the system has ingested.

Who carries the consequence

Start with a fact that is easy to miss: of the S&P 500's ten largest holdings, five are buyers paying for this and three are sellers being paid.

A third of the index sits on the two sides of the same transaction. So a capex slowdown is, at index level, first a transfer rather than a net loss: what the sellers stop collecting is exactly what the buyers stop spending.

The asymmetry between the two sides is where the consequence actually comes from.

The seller side is the earnings engine. Nvidia converts nine-tenths of operating cash flow into free cash flow, carries none of the construction burden, and runs very high margins. Its revenue is someone else's spending, so a slowdown hits its profit close to one-for-one — multiplied by an 8.19% index weight.

The buyer side gets cash flow back, not growth. Free cash flow would rebound from thirteen billion a quarter to fifty-plus, and the statements would look better immediately. But the premium the market pays these companies rests on the AI story, not on their free-cash-flow level. Cash returns, story weakens, and the multiple may not hold.

The net effect is therefore not "how far the index falls" but a repricing inside it: earnings expectations move off the sellers, cash flow returns to the buyers, and both valuation cases get re-examined at once.

What it transmits into

Bonds: the direction reverses. Today's regime is rates/inflation-driven risk-off — stocks and bonds falling together, bonds not acting as a haven. Once a buildout slowdown reads as a growth risk, money buys Treasuries again and long yields fall. That is the opposite of the 10-year's 52-basis-point rise over twenty sessions, and it is the most likely directional reversal in this whole trade.

Credit: watch investment grade, not high yield. These are investment-grade issuers, so the pressure registers there first. High yield has already widened 17 basis points over twenty sessions; investment grade has not moved. The latter is the leading indicator for this buildout, and it has not started.

Power and compute real estate: already absorbing it. DTCR's 30-day net outflow is 2.62% of assets; URA sits 5.46% below its 50-day average at 40.3 relative strength. The far end of the chain is not waiting for a trigger.

Volatility: the mechanical bid amplifies rather than cushions. Twenty-day realized volatility on the S&P is 10.38%, in the 29.6th percentile of its own year. On a 10% volatility target, implied leverage of 0.96 sits near its cap. Double the volatility and mechanical exposure has to halve. None of that selling has an opinion.

The timescale is not one quarter. Capex takes three to five years to become depreciation in full, so the pressure on earnings arrives in installments rather than as a single shock. That means no clean collapse — and also that once it starts, it does not end quickly.

What would make this systemic

Between a sector stalling and a financial crisis sits one concrete question: whose balance sheet absorbs the loss.

The 2000 dot-com bust erased an enormous amount of equity value without becoming a banking crisis, because shareholders carried it. In 2008 a comparable decline in asset values did become systemic, because the losses landed on highly leveraged intermediaries. What decides the character is not the size of the fall but where the leverage sits.

By that standard, the visible layer of leverage here is high quality: Microsoft's debt is 0.31 times annualized operating cash flow, Meta's 0.86, Amazon's 1.41, all on investment-grade companies' own balance sheets. That kind of loss absorbs well and does not force selling in the short run.

The risk is in the layer that is not visible.

First, off-balance-sheet structure: operating leases, purchase commitments and unconsolidated special-purpose vehicles, none of which any balance sheet reaches — and which is exactly what the reporting is about.

Second, private credit. The business-development-company sector the system tracks holds roughly $175bn, with payment-in-kind interest at 12.4% and rising, non-accrual loans at 3.1% and rising, and net asset value down 2.8% year on year. Rising payment-in-kind means borrowers are paying interest by borrowing again. That is the same disease one layer down — above, debt pays for construction; below, debt pays for interest.

Third, who holds the paper. This investment-grade debt ends up with insurers, pension funds and bond funds, and this report has no data on that distribution. It can only be named as a blind spot.

What panic would look like

If it comes, the sequence is unlikely to be "AI stocks fall 30%". It is four steps.

1. Correlation jumps. Twenty-day average pairwise correlation across sectors is 0.208, in the 40th percentile of its own year — still on the side where rotation holds the index up — but it has risen 0.083 in twenty sessions. Once it breaks higher, diversification stops working the same day. 2. Mechanical de-risking. Volatility-target strategies carry implied leverage of 0.96, near their cap; a doubling of volatility halves the exposure. That selling reads neither news nor valuation. 3. Collateral repricing. Borrowings secured on long leases and pre-sold contracts weaken as demand slows — the reason for the loan and the source of repayment are the same thing. 4. The financing window narrows. Only at this step is it systemic. The first three can happen in days; the fourth usually needs a specific event — an issue repriced, or a rating outlook cut.

The system currently reads rates/inflation-driven risk-off with a systemic score of 6.8/100, and neither crisis discriminator — credit widening sharply, bonds returning as a haven — has triggered. That is "not now", not "not going to".

Three tells are readable daily: investment-grade spreads starting to widen (currently unchanged over twenty sessions), bonds behaving as a haven again (currently selling with stocks), and sector correlation's z-score breaking higher (currently −0.47).

The market is pricing none of this

Read the last row against the first: valuations are not stretched and volatility is low, but participation is narrowing.

And one mechanical amplifier: twenty-day realized volatility of 10.38% sits in the 29.6th percentile of its own year. On a 10% volatility target that implies leverage of 0.96, near the cap. Double the volatility and mechanical exposure falls to half of today's — selling with no opinion attached.

Three limits on this piece

First, this is a lower bound, not the whole picture. Capital leases are already inside total debt (of Amazon's $209.9bn, $90.8bn is leases), but operating leases, purchase commitments and unconsolidated special-purpose vehicles appear only in annual-report footnotes, where no balance sheet reaches. What is invisible can only make these numbers worse.

Second, a single data source. Financial figures come from quarterly statements via yfinance; interest may be reported net. The direction is trustworthy, the last decimal is not. The quarterly series runs only five quarters — enough for direction, not for comparison across cycles.

Third, the system contradicts itself here, and this piece does not pick a side. As of 25 September the system reads technology (XLK) as leaning positive at 0.70 confidence and semiconductors (SMH) at 0.64, while two bearish opportunities opened on 19 September remain open. It also has to be said that the financing angle is arithmetic this piece did from the statements, not a reading the system produced — across 17,558 analyses in the last 90 days, not one used it. That is exactly why the capex_funding_snapshot probe was added this week.

Disclaimer

This piece was produced and assembled by an automated research system. It is general information and educational discussion only. It is not investment advice and not a recommendation to buy or sell any security, and it has not been tailored to anyone's financial situation, investment objectives or risk tolerance.

All figures come from public sources — company quarterly filings, the St. Louis Fed's economic database, fund issuers' published holdings files — and may contain collection, parsing or definitional errors. Past data does not indicate future performance. The "lower bound" caveat repeated throughout means some of the debt structure sits outside what these sources can see.

The operator of this site is not registered with the Ontario Securities Commission or any securities regulator. Investing involves risk and decisions are your own; consult a licensed professional adviser before acting.

---

Data as of 2026-09-25. Financial figures are each company's latest disclosed quarter (Microsoft, Alphabet and Meta 2026Q2; Amazon 2026Q1; Oracle and Nvidia their 2026Q3 fiscal quarters). Rates and spreads from FRED: DGS10, DGS2, BAMLH0A0HYM2, BAMLC0A0CM (through 2026-09-24); liquidity from WALCL, WRESBAL, RRPONTSYD; private credit from business-development-company sector disclosures for 2026-Q1; index weights from State Street's SPY holdings file dated 2026-09-24.