SpaceX does not build AI data centers.
That was the first thing I checked when the S&P Global note reached me ā relayed through two aggregators and one Web3 news feed, with no byline, no timestamp, and no link back to the original report. Six companies were named as carrying AI infrastructure debt that may exceed their disclosures. Amazon. Microsoft. Alphabet. Oracle. Meta. SpaceX.
Five of them file capital expenditure line items a forensic accountant can reconcile to the dollar. One of them launches rockets.
Glitch detected. Source traced.
The substance is real, which is what makes the packaging dangerous. S&P's actual claims ā that AI-related capex will run into the trillions through the end of the decade, that at least one of these firms will carry negative free cash flow into next year, that the largest risk is profit failing to materialize after the buildout finishes ā are all defensible. None of them are new. What is new is the funding structure underneath them, and that is where the note earns its attention and where the transcription chain loses it.
Start with the number everyone quoted: seven trillion dollars, 2025 to 2030, attributed to six companies. It does not belong to them. That figure tracks McKinsey's global data center capex estimate to 2030 ā roughly $6.7 trillion across every operator, every geography, sovereign builds and Chinese hyperscalers included. Six companies' AI capex over the same window is smaller by a factor of two or three. The scale is genuine. The attribution is fabricated, and the fabrication moves the risk from the entities that carry it to the entities that can absorb it.
What survives verification is narrower and more useful. In the second half of 2025, Meta issued roughly $30 billion in debt, Alphabet about $25 billion, Oracle about $18 billion, Amazon about $15 billion. More than $80 billion in a single half-year from four issuers who had been net cash generators for a decade. That is the regime change: AI capex has moved from self-funded to levered, and you can audit it in the exhibit pages rather than the press release.
The "negative free cash flow through next year" line deserves the same scrutiny the capex figure didn't get. Negative FCF from high-return investment and negative FCF from insufficient return look identical on a cash flow statement and are opposites inside a credit model. The note does not distinguish them. Neither did the transcription.
Then the accounting variable the coverage skipped entirely. Depreciation.
A GPU running continuous training load has an economic life closer to two or three years ā thermal cycling, HBM failure rates, interconnect obsolescence. The accounting life assigned to it is typically six years, sometimes five. Divide the capex by the schedule and you get the number nobody wants to print: seven trillion dollars at seven-year straight-line requires roughly a trillion dollars of annual gross profit just to cover the capital charge. Aggregate generative AI revenue in 2025, across every vendor on earth, is in the hundreds of billions. The gap is twenty to thirty times. It isn't a forecast. It's a division problem. And it stays invisible on an income statement right up until a CFO stretches the useful-life assumption ā which is precisely the lever pulled before a covenant breach, and precisely the footnote that never makes the aggregator feed.
The note's central allegation ā contingent liabilities above disclosed levels ā is correct and under-specified. The carriers are identifiable: the MetaāBlue Owl Hyperion joint venture at roughly $27 billion, Oracle's sale-leaseback arrangements with Vantage and similar developers, Microsoft's migration to operating leases and third-party capacity purchases from CoreWeave and Nebius. Each structure moves capex off the balance sheet and leaves a commitment behind. None of them appear as a single reconcilable line.
I have watched this latency before. In 2024 I built a Python model of IBIT institutional flows and found a nine-day gap between positioning data and the price at which the market actually repriced it. The AI credit complex runs the same structure with a longer fuse: sixty to ninety days between an SPV being signed and anything resembling disclosure reaching a lender's dashboard. Tokenized private credit funds already hold GPU-backed and SPV-backed paper, and DeFi lenders price that paper off headline yield, not covenant ladders. Oracle feeds update on a heartbeat. The underlying truth updates quarterly, if at all. Liquidity draining. Logic broken.
Now the sixth name. If the original list read xAI ā Colossus, SPV-heavy, vendor-backstopped ā then the entity with the least transparent debt on the planet was relabeled as a rocket company by three consecutive transcription hops, and nobody caught it. That is not an error about aerospace. It is the disclosure gap reproducing itself inside the reporting on the disclosure gap. Exchange volume anomaly flagged, and the anomaly was in the metadata.
The consensus reading says the six balance sheets are at risk. The arithmetic says otherwise. Four of these firms generate more than $500 billion in combined annual operating cash flow with net debt under half of EBITDA. S&P's refusal to cut ratings broadly isn't capture. It's math. The first-order exposure sits one hop downstream: roughly $1.4 trillion in multi-year compute commitments, neoclouds levered three to four times against GPU collateral, and private labs whose debt is invisible because nobody forced them to print it.
And then circular financing. NVIDIA committing up to $100 billion into a lab that simultaneously commits to ten gigawatts of NVIDIA compute. NVIDIA equity in CoreWeave, in Nebius, in xAI. The vendor funds the buyer, the buyer buys the vendor's product, and the vendor's stock becomes the collateral for the whole loop. Lucent ran this play. Nortel ran it. It works until the equity stops working, and then the collateral is the story.
So watch the useful-life footnote, not the headline. Watch Oracle's credit default swaps, which have already widened while the outlook sits at negative. Watch whether the trillion-dollar compute commitments ever receive a parent guarantee ā because a guarantee converts a rumor into a liability, and a liability into a covenant.
The debt is real. The disclosure is a rumor passed through three hands. Which one does the market price first?
