The Hyperscaler Debt Machine

For twenty years the companies now building the world's AI infrastructure paid for their expansion the enviable way: out of their own operating cash flow. In 2025 that quietly stopped being true. Meta sold one of the largest corporate bonds in history and parked tens of billions more in an off-balance-sheet vehicle; Oracle's free cash flow turned negative under the weight of its commitments; a layer of GPU-landlord “neoclouds” borrowed against the chips themselves. This piece maps the orders of magnitude of that borrowing, the circular way much of it is arranged, the dependency it has created in the wider economy, and four scenarios for how it resolves. It builds on the capex supercycles and infrastructure booms pieces. Editorial estimates; methodology in the appendix.

The change nobody announced

The defining financial fact about Microsoft, Alphabet, Amazon and Meta for most of the last decade was that they did not need anyone's money. They generated more cash than they could reinvest, which is why they paid dividends, bought back stock, and sat on cash piles measured in the hundreds of billions. Capex — even large capex — came out of that river of operating cash flow without touching the debt market in any way that mattered.

The AI buildout has broken that arrangement. Combined hyperscaler capital expenditure ran on the order of $350–400bn in 2025 and the guidance for 2026 points higher still — toward and past half a trillion dollars a year. That is more than the cash flows, enormous as they are, comfortably cover once you also want to keep buying back stock; and it is being spent on assets — racks of GPUs — that depreciate far faster than the buildings around them. So the money is increasingly being borrowed. Alphabet, long almost debt-free, returned to the bond market at scale. Amazon issued multi-billion-dollar bonds and piled up finance-lease obligations. Meta sold a bond on the order of $30bn — one of the largest single corporate issues ever — and, separately, arranged a roughly $27bn financing vehicle for a Louisiana data centre that is structured to sit largely off its own balance sheet, with private-credit funds holding the paper. Oracle, having committed to build against a backlog that includes an enormous multi-year compute deal, watched its free cash flow go negative and its total debt climb past the $100bn range.

The orders of magnitude

Set against the aggregates, hyperscaler borrowing is still small. US non-financial corporate debt is on the order of $14tn; total US public debt is past $37tn; global debt across all sectors is north of $300tn. A few hundred billion dollars of AI-infrastructure bonds and loans does not move those needles. The danger in the number is not its size relative to the world — it is its concentration (a handful of issuers and lenders), its speed (most of it raised inside a single year), and the fact that it is being borrowed against a projection of demand nobody can yet prove. That combination is exactly the one the infrastructure-booms pattern flags as dangerous, regardless of the headline total.

The circular part

What makes this buildout harder to read than the railways or the fiber glut is that a large share of the demand underwriting the debt is supplied by the same companies benefiting from it. The chip maker takes an equity stake in the model lab; the model lab signs a multi-year commitment to buy compute from a cloud provider; the cloud provider borrows to buy chips from the chip maker; and each leg of that loop is reported by someone as revenue, backlog, or a signed contract that helps justify the next round of borrowing. Vendor financing — a supplier lending its customer the money to buy the supplier's product — was a hallmark of the telecom bubble that ended in the fiber bankruptcies, and it is present here at a scale the 1990s never reached. None of the individual deals is improper; the concern is that in aggregate they can make a demand curve look more solid, and more independent, than it actually is.

Why the whole economy is now watching

The reason this has stopped being a story about a few big companies' balance sheets is that AI capex has become a load-bearing part of measured economic growth. Through 2025, data-centre and AI-related investment accounted for a strikingly large share of US GDP growth — by several credible estimates, something approaching or exceeding half of it in the strongest quarters. Construction, power equipment, chips and the buildings to house them are real output; while the spending accelerates, it flatters the growth figures. The corollary is uncomfortable and symmetrical: if the buildout slows sharply, it subtracts from growth just as visibly, and does so at the same moment any credit stress would be hitting. Layer on the equity market's concentration — a small number of AI-exposed names now make up an outsized share of the main US index, and therefore of the retirement and pension savings that track it — and a re-rating of AI infrastructure stops being a sector event and becomes a household-balance-sheet event.

The structural echo

The off-balance-sheet vehicles are the detail a reader of financial history should linger on. Financing a data centre through a separate entity, capitalised by private-credit funds and designed so the debt does not consolidate onto the operating company's books, keeps reported leverage low and preserves the credit rating. It is also, structurally, the same manoeuvre that pyramided Samuel Insull's utility empire before 1932 and that sat behind Enron's special-purpose entities — moving the obligation a step away from the balance sheet that ultimately depends on it. That does not make it fraudulent; today's versions are disclosed and legal. It does mean the true, system-wide leverage of the AI buildout is harder to see than the individual companies' investment-grade balance sheets suggest, and that the private-credit layer — less liquid and less transparent than the public bond market — is where stress would first concentrate and last be noticed.

Risks, rewards, and what actually decides it

• The reward is real and large. If AI demand compounds the way its builders expect, this infrastructure becomes the backbone of the next several decades of growth — the railroads-and-fiber outcome where the capacity, once built, is used and re-used long after the financing that paid for it is forgotten. The hyperscalers doing the borrowing have among the strongest cash flows in corporate history to service it.
• The core risk is a fixed obligation against an uncertain revenue. Cash-flow-funded capex can simply stop in a downturn; debt-funded capex cannot. Bonds and GPU loans are contractual claims that must be paid whether or not the AI revenue arrives on schedule — which is precisely what turned earlier infrastructure over-builds from corrections into bankruptcies.
• Depreciation is the quiet accelerant. GPUs are not railway track. Their useful economic life is measured in a few years, not decades, so a demand shortfall does not leave a durable asset quietly waiting to be used — it leaves a rapidly ageing one against which debt was raised. The mismatch between short-lived collateral and multi-year debt is the buildout's sharpest financial edge.
• The circularity is the thing to watch. The more of the demand that turns out to be the industry financing itself, the faster the whole structure unwinds if any single large node — a model lab that misses its revenue, a neocloud that can't refinance — steps back.

Four scenarios

• Escape velocity. AI revenue grows faster than the capex needed to serve it, the debt is repaid out of cash flow, and the concentration that looks reckless in 2025 looks visionary in hindsight. This is the outcome the buildout is priced for. It is possible; it is not the base case history would pick.
• The fiber outcome. Demand is real but arrives more slowly than the buildout assumed. There is over-capacity, a round of write-downs, a pause in new orders, and the failure of the weakest, most leveraged players — while the strongest hyperscalers absorb it and the compute is eventually used. A serious correction, not a systemic collapse. This is the median historical analogue.
• The WorldCom outcome. Demand disappoints materially, the circular financing reverses, and a credit event in the neocloud or private-credit layer forces sales of depreciating collateral into a falling market. Because the equity exposure is concentrated in a few index-dominating names, the re-rating transmits directly into broad savings, tips sentiment, and raises genuine recession risk. Lower probability than the fiber outcome — but the tail that actually matters.
• The zombie boom. The least dramatic and perhaps most plausible: demand is always just enough to justify continuing and never enough to earn back the cost of capital. Debt is rolled rather than repaid, capex keeps propping up headline GDP, and the economy spends a decade quietly mis-pricing a very large pool of capital — no crash to clear it, no payoff to vindicate it.

The honest conclusion

The sober reading of the pattern is the same one the infrastructure-booms piece reaches by a different route: the network being built now is very likely to matter and to endure, and that is not the same question as whether the capital structure financing it survives intact. What is new, and what this piece is about, is that the firms doing the building have moved from paying with their own cash to borrowing against a forecast — and have arranged a meaningful part of that borrowing where it is hardest to see. That shift is what converts an ordinary over-build into something with the capacity to reach the wider economy. Being on the right side of technological history, as the railroad, utility and telecom financiers all discovered, has never been the same thing as being solvent at the end of it.