The sixth name on the list was wrong, and that turned out to be the most useful thing in the document.
I was doing a fact-check I had never planned to do. A reposted summary of an S&P Global note on AI infrastructure debt had landed in a Telegram group I lurk in — the kind of aggregator feed that recycles financial headlines for a crypto audience with no particular connection to the underlying subject. The list of companies carrying the exposure read: Amazon, Microsoft, Alphabet, Oracle, Meta, SpaceX.
Five of those are right. The sixth is not.
SpaceX is a private launch company. Its AI infrastructure capex is a rounding error against Starship and Starlink, and xAI — the entity actually burning billions on the Colossus cluster in Memphis — is a legally separate company with separate creditors and a separate capital structure. Whoever retyped that list did not know the difference. Whoever forwarded it did not check.
That error is not trivia. It is a diagnostic. The information chain carrying the AI-debt story to ordinary readers is degrading faster than the debt itself is compounding. And the people now being invited to take exposure to that debt through tokenized wrappers are the least equipped to notice the degradation. That is why I want to write about this not as a macro commentator, but as someone who has spent a decade watching what happens when financial complexity gets sold to people who were never handed the vocabulary to question it.
What the report actually says, and where the numbers came from
Strip away the relay layer and the substance is thinner than the headline but more interesting than the panic.
S&P's core claims, as reported, are four. Six companies account for a dominant share of AI infrastructure spending. That spending runs somewhere north of seven trillion dollars between 2025 and 2030. At least several of those companies will run negative free cash flow through next year. And the rating agency has not broadly downgraded anyone, because the biggest stated risk is not insolvency — it is that the profits fail to arrive after the investment is complete.
Three of those four survive verification.
Negative free cash flow is real and checkable. Amazon, Alphabet, and Oracle all saw free cash flow turn negative or compress sharply during 2025 as capex absorbed operating cash generation. The absence of broad downgrades is also real, though the framing omits something material: Oracle's outlook was cut to negative and its credit default swaps widened noticeably. That is not a downgrade, but it is the rating agency leaning forward in its chair.
The seven-trillion figure is where the reporting gets sloppy in a way that matters. That number is in the right magnitude for McKinsey's estimate of global data-center capital expenditure through 2030 — roughly 6.7 trillion, global, all industries, including sovereign and enterprise spending that has nothing to do with these six balance sheets. It is not a six-company number. It has been scope-laundered from an industry-wide forecast into a specific corporate liability, and that single substitution does more work in the headline than any of the actual analysis underneath it.
I have made a version of this mistake myself. In 2020 I led a community translation of Aave's whitepaper for about five thousand non-technical readers across Eastern Europe, and the hardest part was never the liquidation math. It was the caveats. Every simplification I made to keep people reading was a place where a number could drift away from its original denominator. I learned then that a figure without its scope attached is not information — it is a mood.
The gap is real, and it is an order of magnitude wide
Here is the arithmetic that the aggregator version buries, and it is the part that genuinely deserves attention.
Take the capital expenditure at face value. Seven trillion deployed over six to eight years, depreciated over a similar horizon, implies something on the order of nine hundred billion to 1.2 trillion dollars per year in depreciation and capital cost that has to be covered by gross profit. Now look at the revenue side. Global generative AI revenue in 2025 — the leading lab's annualized run rate, its nearest competitor, plus the AI-attributable portions of cloud revenue — lands in the tens of billions, generously a few hundred billion if you are loose about attribution.
The gap between those two numbers is roughly twenty to thirty times.
That ratio, not the debt load, is the actual story. A twenty-to-thirty-fold gap between annual capital cost and annual revenue does not require a bubble to be dangerous. It only requires time. And time is precisely what debt financing removes from the equation, because debt has a schedule and revenue does not.
What changed in 2025 is not the size of the ambition. It is the funding mechanism. Through the first half of the decade, hyperscale AI capex was largely self-funded out of extraordinary operating cash flow — a genuinely remarkable thing, and the reason the "AI bubble" chorus was early and wrong for years. In the second half of 2025 that flipped. Meta raised roughly thirty billion in debt. Alphabet around twenty-five. Oracle eighteen. Amazon fifteen. Over eight hundred billion in AI-adjacent issuance in a single stretch.
That is the structural break. The AI buildout crossed from "funded by profit" to "funded by promise," and almost nobody outside the credit desks marked the crossing.
The carriers of the hidden liability
The second mechanism is quieter and far more consequential, and it is the one the original report gestures at without naming properly.
Off-balance-sheet structures are doing an enormous amount of work here. Meta's joint venture with Blue Owl around the Hyperion data center, sized in the tens of billions. Oracle's sale-and-leaseback arrangements with developers like Vantage. Microsoft's large-scale pivot toward operating leases and third-party capacity purchased from CoreWeave and Nebius rather than built on its own books.
Each of these is a legitimate financing technique. Each of them also moves a fixed obligation somewhere other than the line item an analyst checks first. A sale-leaseback converts capex into a multi-year rent commitment. A joint venture converts a build into a minority equity stake plus a capacity contract. Third-party capacity converts owned infrastructure into a purchase obligation. In every case the economic exposure persists; only its address changes.
This is where my regulatory work becomes relevant, and I want to be careful not to overstate it. In 2025 I advised an EU task force drafting governance standards for decentralized protocols, alongside ten legal experts, focused on protecting retail participants in structures they cannot inspect. The single most contentious item in those sessions was never about crypto at all. It was about disclosure of contingent obligations — how you make a counterparty risk visible when the entity bearing it is not the entity signing the contract.
The AI buildout is now the largest live experiment in exactly that problem.
And here the risk topology splits sharply, in a way the flat six-company framing destroys. The four mega-cap hyperscalers generate well over five hundred billion in combined annual operating cash flow with net debt to EBITDA comfortably under one turn. For them, leverage is a second-order question — a financing preference, not a solvency threat. S&P's refusal to downgrade them is not negligence; their buffers are genuinely large.
For Oracle, the neoclouds, and the private AI labs, the same debt is a first-order question. They do not have the cash flow to absorb a demand pause, and they are the ones holding the multi-year capacity commitments.
The circularity nobody is pricing
There is a third layer, and it is the one that makes me genuinely uneasy rather than merely cautious.
Vendor financing has returned at scale. A major chip supplier has committed to invest up to a hundred billion dollars in a leading AI lab while simultaneously backing that lab's multi-gigawatt compute purchase — and has separately invested in the neocloud providers who buy its hardware. The money moves in a circle: supplier capital funds customer capex, customer capex purchases supplier product, supplier revenue justifies further investment.
I have read this pattern before, in the Lucent and Nortel literature from the telecom buildout. Equipment vendors financing their own customers' purchases is not fraud. It is a rational response to a land-grab. It is also, historically, one of the most reliable accelerants of a bust, because when demand softens the vendor's investment loss and the vendor's revenue loss arrive in the same quarter.
What worries me more than the circularity itself is the concentration hiding behind it. The genuinely enormous single exposure in this system is not any one balance sheet — it is the order of 1.4 trillion dollars in multi-year compute purchase commitments. Commitments, not debt. Contractual obligations that do not always appear as liabilities, that are frequently unsecured by parent guarantees, and whose realisable value depends entirely on whether the buyer's business materialises.
The system's largest vulnerability is a receivable, and receivables are valued by assumption.
The variable everyone skips: what is a GPU worth in year four
Now the contrarian turn, because I think both the bulls and the bears are arguing about the wrong number.
The bears point at debt-to-EBITDA. The bulls point at operating cash flow. Both are looking at the numerator of the problem and ignoring the denominator, which is the residual value of the asset the debt was raised against.
Accounting depreciation for AI accelerators is typically assumed over five to six years. The physical useful life of a chip running continuously in a high-thermal-load cluster is a contested and largely unmeasured question, and the economic useful life is a different question again — because a GPU does not lose value when it breaks, it loses value when a faster one ships. Every generation step resets the price of the previous one.
There is no liquid secondary market for these assets. There is no mark-to-market discipline. There is no counterparty posting variation margin against the declining resale value of a rack of accelerators. The entire collateral base of the AI credit boom is carried at an accountant's assumption, revised annually, disclosed in a footnote.
I have watched this exact failure mode at small scale, and it is why I stopped treating collateral as a number. When I translated Aave's liquidation mechanics for non-technical users, the question I got most often was not about interest rates — it was "what happens when the collateral falls faster than the liquidation can execute?" That is the right question, and it is the question the AI infrastructure market has not been asked, because nobody is forced to mark anything.
Here is the part my own industry should be uncomfortable about. Real-world asset credit is one of the fastest-growing categories in DeFi, and AI infrastructure — data centers, compute contracts, GPU leases — is precisely the kind of yield-bearing instrument that gets packaged for on-chain distribution. The wrapper will look like a token. The underlying will be a lease on an asset whose value is set by a depreciation schedule.
Build for humans, not just nodes — and humans cannot audit a footnote they were never shown.
Where this actually goes
I am not forecasting a crash. The hyperscalers can fund this cycle to completion out of cash flow if they choose to, and the demand for inference is real, growing, and not going away. The bear case that has been wrong for three years can stay wrong for three more.
What I am saying is narrower and harder to dismiss: the risk in this buildout has migrated from equity to credit, from the income statement to the footnote, and from the balance sheet to the contract. The four companies that can absorb it are not the ones the headline lists first. The entity carrying the most opaque exposure may not even be named correctly in the version most people read.
And the transmission channel into my own industry is already open. Tokenized credit does not care whether the underlying lessee has a parent guarantee. It cares whether the yield is attractive.
Education is the ultimate yield. So here is the question I would put to anyone being offered a double-digit return on AI infrastructure this cycle: when the collateral is revalued, who marks it — and will you be told before or after the position is closed?
