The SpaceX Error: Autopsying the $7 Trillion AI Debt Warning Nobody Audited

On a Tuesday in late 2025, a credit note attributed to S&P Global began circulating through crypto news aggregators. It named six companies carrying AI infrastructure debt, running negative free cash flow "at least through next year." Five names were predictable: Amazon, Microsoft, Alphabet, Oracle, Meta. The sixth was SpaceX.

That sixth name is the tell. SpaceX is a private launch and satellite operator. Its AI capital expenditure is a rounding error against Starlink and its launch cadence. xAI — Elon Musk's separate AI venture — is a legally distinct entity whose Colossus cluster consumed billions in 2025, financed substantially through special-purpose vehicles and vendor-backed debt. The reasonable inference: the original report said xAI, and an aggregator rewrote it as SpaceX. A single line of logic can unravel a thousand lies — but here the lie is not malicious. It is structural. Whoever republished that note could not distinguish a rocket company from a GPU cluster. That failure of basic taxonomy is the same failure that runs through the entire AI debt narrative: people are pricing a story they cannot read at the code level.

Context: The Largest Industrial Bet in Financial History

The setup demands precision, because the numbers being thrown around are being used as if they were interchangeable. They are not.

Through 2025, the largest technology companies on earth entered a capital expenditure cycle without historical parallel. The frequently cited figure — roughly $7 trillion in AI-related capex between 2025 and 2030 — traces to McKinsey's global data-center buildout estimate, cross-checked against Morgan Stanley and other sell-side models. That figure is real. It is also global and industry-wide. It is not, as the aggregators implied, the aggregate spend of six named companies. This is the first口径 — the first measurement-basis — error in the chain, and it matters enormously, because the $7 trillion is being used rhetorically to imply a concentrated, nameable credit risk when in fact it describes an economy-wide infrastructure wave.

What S&P actually flagged was narrower and more defensible: that the debt embedded in AI infrastructure may exceed what these companies disclose, that free cash flow at several of them has turned negative or narrowed sharply, and that the deepest risk is not default but the failure of profit to materialize once the concrete is poured and the GPUs are racked. As of late 2025, S&P had not broadly downgraded the majors. Oracle's outlook, however, had already been shifted negative, and its credit default swaps had widened materially. Those two facts — no broad downgrade, but a negative watch on the most levered name — are the entire story compressed into a single line.

I have spent eleven years reading ledgers that were designed to be misread. When a credit agency says "disclosure may be insufficient," it is not making a prediction. It is telling you that the balance sheet you can see is smaller than the balance sheet that exists. That is an invitation to forensic work, and the crypto press corps answered it by miscounting the companies.

Core: Where the Debt Actually Lives

The core analytical failure in the circulating note is a category collapse: it treats capex as if it were debt. It is not. Of the capital deployed, the debt-financed share is plausibly 20 to 40 percent, with the remainder funded from operating cash flow, operating leases, and equity. This distinction cuts both ways. It understates the immediacy of credit risk — the big four are not about to trip covenants — but it dramatically understates the structural risk, because the debt that does exist is migrating into exactly the vehicles that disclosure regimes were never built to capture.

The most valuable fact in the whole chain is not the $7 trillion. It is the financing-mode switch that occurred in the second half of 2025. Meta issued roughly $30 billion in AI-related bonds. Alphabet, roughly $25 billion. Oracle, roughly $18 billion. Amazon, roughly $15 billion. Together, north of $80 billion in a single half-year window. That is the moment AI capex stopped being self-funded from surplus cash flow and became leveraged. Everything downstream — the widening spreads, the negative-watch on Oracle, the sudden analyst alarm — follows from that single transition. The market is not reacting to the size of the spend. It is reacting to the change in who bears the risk.

The Off-Balance-Sheet Layer

Here is where my forensic training earns its keep. The note's core accusation — that real liabilities exceed disclosed liabilities — is correct, and the evidence is hiding in plain sight in structures that never appear as "debt" on a headline balance sheet:

Meta's Hyperion data-center joint venture with Blue Owl, at roughly $27 billion, is a special-purpose vehicle. Oracle's sale-leaseback arrangements with developers like Vantage are, functionally, financing dressed as rent. Microsoft's large-scale pivot toward operating leases and third-party capacity procurement — CoreWeave, Nebius — moves capital intensity off its own books and onto counterparties whose credit is thinner and whose disclosures are lighter.

These are precisely the carriers of "contingent liabilities above disclosure." When a hyperscaler signs a multi-year capacity contract with a neocloud, the hyperscaler books an operating expense. The neocloud books a debt-financed buildout. The credit risk has not disappeared. It has been relocated to the least transparent node in the network. And the rating agencies, looking at the hyperscaler's pristine balance sheet, see a fortress. They are not wrong about the fortress. They are wrong about where the moat ends and the swamp begins.

The Monetization Math

Strip away the rhetoric and the "profits won't materialize" warning becomes a subtraction problem. If $7 trillion of capex is depreciated over six to eight years, the annual gross-profit contribution required just to cover the capital charge is on the order of $0.9 trillion to $1.2 trillion per year. Against that, 2025's global generative-AI revenue — OpenAI at roughly $13 billion annualized, Anthropic at roughly $7 billion, plus the AI-attributable slice of cloud revenue — totals in the hundreds of billions at most.

The gap is 20 to 30 times. That is the arithmetic behind every credit warning you will read this cycle, stated plainly. The capex is not being financed by the revenue it generates. It is being financed by the expectation of revenue that does not yet exist, intermediated through debt and leases. This is not a scandal. It is how every infrastructure buildout in history has been financed — railroads, fiber, telecom. The question is never whether the gap exists. It is who is holding the paper when the gap fails to close on schedule.

The Circular Financing Nobody Names

The note is silent on the mechanism that should worry every forensic reader most: vendor financing. NVIDIA has committed to invest up to $100 billion in OpenAI while simultaneously backstopping OpenAI's 10-gigawatt compute procurement — compute that is, overwhelmingly, NVIDIA silicon. NVIDIA has also taken positions in CoreWeave, Nebius, and xAI. Trace the flow and you find capital leaving the chip vendor, entering the customer, and returning to the chip vendor as revenue. The supplier is financing its own demand.

This is not new. Lucent and Nortel ran the identical playbook in the late 1990s, lending to carriers so the carriers could buy their equipment. It worked until it didn't, and when it stopped, the write-downs were catastrophic precisely because the revenue and the loan were the same dollar. Cold eyes see what warm hearts ignore: a dollar that appears twice in a revenue line is not growth. It is a loop. And loops are invisible on any single company's income statement.

Wallet Anatomy: Mapping the Real Exposure

When I traced fund flows during the Terra collapse, the lesson was that the entity on the press release is rarely the entity holding the risk. The same topology applies here. Rank the six names by first-order credit exposure and the ranking inverts the headlines.

The big four — Amazon, Microsoft, Alphabet, Meta — carry combined annual operating cash flow north of $500 billion and net debt-to-EBITDA generally below 0.5x. Their negative free cash flow is the benign kind: heavy investment against a business that prints cash. S&P's decision not to downgrade them broadly is not negligence. It is correct. For them, AI debt is a second-order problem.

The first-order problems sit elsewhere, and they are barely mentioned:

Oracle is the single most levered major, with a negative outlook and widening CDS, funding a backlog it must build before it can bill. The neoclouds — CoreWeave and its peers — are GPU-collateralized lenders whose credit quality is a derivative of NVIDIA's roadmap and rental pricing. Private AI labs, xAI chief among them, raise through SPVs and vendor debt with the thinnest disclosure of all. The single largest exposure in the entire system, in fact, is not on any of these balance sheets: it is OpenAI's reported ~$1.4 trillion in multi-year compute commitments — a figure larger than the GDP of most nations, held by an entity whose own revenue is a rounding error against it.

That is the true concentration. Not six public balance sheets. One private counterparty's contractual promises, distributed across a dozen vendors and financiers who each believe someone else holds the senior claim.

The Depreciation Variable

The variable every media account omits is the one that governs everything: assumed useful life. A GPU cluster depreciated over six years versus three years is a difference of billions in annual earnings for a single hyperscaler. Extend the assumed life and you flatter near-term profit while pushing the obsolescence cliff into a future quarter. This is the quiet accounting lever of the entire cycle, and it is almost never audited in the coverage. When you see margins holding despite a capex surge, ask what depreciation schedule is holding them up.

Contrarian: The Bulls Are Not Wrong About the Fortress

Here is where I part company with the doom chorus, because the panic is mislocated and a mislocated panic is its own kind of negligence.

The bulls are correct on the central point: the big four can self-fund. Their cash generation dwarfs their debt service. They are not Enron. They are not even close. The AI buildout, for them, is a strategic bet funded from strength, and if returns disappoint, they will slow the spend and absorb the hit — a hit measured in valuation multiples, not solvency. Anyone predicting the collapse of Microsoft or Alphabet from AI capex is reading the wrong ledger.

The genuine contrarian insight is narrower and colder: the systemic risk is not in the companies being warned about — it is in the counterparties they are quietly backstopping. The hyperscalers' strength is precisely what allows them to push capital intensity onto neoclouds and SPVs, and it is those thinner nodes that will crack first if AI revenue growth decelerates from triple digits to fifty percent. The rating agencies are watching the fortress. The swamp is where the leverage migrated, and it did so with the fortress's blessing.

There is a second bull point the bears ignore: infrastructure cycles that look catastrophic at the trough often build the substrate for the next decade. The fiber glut of 2001 bankrupted dozens of carriers and left the world with cheap bandwidth that powered Web 2.0. The $7 trillion may well be overbuilt. It will also leave behind compute that is cheaper and more abundant than any model today assumes. The question is not whether the infrastructure is wasted. It is who is holding the debt when the repricing happens, and whether that holder can survive the mark-to-market.

Takeaway: The Disclosure Gap Is the Signal

The lesson of the SpaceX error is not that the note was wrong. It is that the note was unreadable by the people who propagated it — and an unreadable warning about opaque debt is worse than no warning, because it manufactures false confidence that someone, somewhere, has done the forensic work. No one has.

Watch the wrong things and you will be surprised by nothing and prepared for everything. Watch the right things: the depreciation schedules, the SPV filings, the vendor-financing loops, the neocloud credit spreads, and the counterparty that is too private to rate but too large to fail quietly. The next time a credit agency says disclosure may be insufficient, do not count the companies. Read the ledgers. A single line of logic can unravel a thousand lies — but only if someone bothers to read the line. Who is auditing the auditors of the buildout? As of this writing, the honest answer is: the market, badly, and late.