The subprime layer
The AI bubble isn’t bursting. The capital structure around it might.
Four weeks ago, this newsletter argued that the AI cycle probably could not crash the way the dot-com bust did, because the world is too short of watts and wafers to build past demand the way fibre did in 2000. Call it a sorting rather than a crash: a slow repricing of which companies were ever capable of capturing the returns the market has priced into all of them.
Since then, one piece of analysis has arrived that does not change the sorting’s direction but does clarify its shape.
The structure underneath
Ed Zitron, who published OpenAI’s 2025 financials featured last time, sent a newsletter last week opening with Mark Baum in The Big Short. The scene: Baum realises, in conversation with a collateralised debt obligation (CDO) manager, that the market for insuring mortgage bonds is twenty times larger than the mortgage bond market itself.
At its peak, a single mortgage bond was routinely sliced into dozens of CDOs; tracking studies counted 5,500 bonds referenced across 36,000+ CDO positions. Housing demand was not independent of the structure financing it: low rates and loose underwriting inflated mortgage demand, and the CDO structure inflated it further, until the two were indistinguishable.
Zitron’s argument: the data centre build-out has developed a comparable property through the special purpose vehicle (SPV) – the CDO’s structural cousin. Each project gets its own legal entity; a holding company like CoreWeave spins one up per build, borrowing against projected customer cash flows and selling the debt in tranches to institutions, asset managers, or banks.
The AI company (say, OpenAI) contracts directly with the SPV, not the holding company. Revenue pays operating costs first, then works down the creditor stack by seniority; the remainder goes to the holding company. Most structures are non-recourse: if the customer defaults, investors claim the SPV’s assets but cannot immediately pursue the holding company’s broader balance sheet.
CoreWeave’s $8.5B DDTL 4.0 loan, for instance, is secured against a Meta contract and the underlying data centre assets, funded via CoreWeave Compute Acquisition Co VIII LLC. Its $2.6B DDTL 3.0, raised to “accelerate delivery of services from OpenAI”, uses two more SPVs. Each has individual debt service and counterparty dependencies, stacked on customers running significant losses.
Off the balance sheet
Meanwhile, hyperscalers run their own off-balance-sheet accounting. Meta has parked tens of billions in data centre obligations into named SPVs absent from its $58.7B balance sheet; actual obligations run higher.
More than $500B in data-centre securities are outstanding (Bloomberg Law). A Nikkei Asia investigation finds the five largest tech companies carry $1.35T reported debt plus $1.65T off-balance-sheet AI obligations – data centre leases and GPU contracts – with the hidden portion up eightfold in four years. Zitron’s verdict:
This is all legal, worrying, and yes, a little bit Enron.
What the contracts assume
SPVs are fragile only if underlying demand fails to arrive at contracted scale – the same problem CDO builders faced. AI services demand at the required scale is not today’s demand.
A Goldman Sachs analysis puts 2026 AI infrastructure spend at $765B. Anthropic and OpenAI together generate $70–95B annualised revenue (mid-2026). The gap between spend and revenue base is ~10x.
Contracts treat projected demand as being as established as the assets backing them. The CDO structure avoided asking whether borrowers could pay. This SPV structure has not yet been forced to answer the same question restated: Where does a tenfold revenue increase come from, and when?
What token pricing tells you
Benedict Evans this month analyses post-crunch token pricing. The crunch is real but driven by “really just one use case, software development” – “a pretty small field.” Next use cases, timing, and token needs remain unknown.
Across supply-side capex, inference efficiency, frontier-lab competition, every path to foundation models achieving market dominance, strategic leverage, or value capture “requires something to change” that we don’t yet see.
His parallel: mobile data. Mobile networks faced sudden marginal-cost capacity demand – as AI compute does now – but built out over decades, not years. As Evans writes:
Mobile data traffic has risen by several orders of magnitude, and this has become an enormous industry, with annual revenue of a trillion dollars and capex of $200 billion, but the stocks have gone nowhere, and all the value was captured by other people further up the stack.
Mobile networks had no subprime layer – high capital costs, low margins, but straightforward balance-sheet problems with understood collateral. AI infrastructure has Evans’s commodity trajectory plus Zitron’s synthetic complexity. That combination has no historical precedent.
Three concurrent sortings
The first piece in this series ended with a sorting. There are, more precisely, three concurrent sortings underway, running on different timelines and affecting different people.
The model layer
The model layer sorts through competitive dynamics pushing foundation models toward commodity infrastructure, with value captured above them. Apple owns the interface, NVIDIA owns compute without it, foundation model labs occupy a position Evans suggests won’t sustain a premium. Alphabet and Microsoft, straddling both, remain harder to call.
A serious counter-argument comes from Ben Thompson, in a recent Stratechery analysis: the agent paradigm will expand inference demand so dramatically that today’s 10x infrastructure-vs-revenue gap closes by demand-side explosion, not supply correction. This may prove right on the technology timeline.
The subprime layer, however, operates on a financial timeline that doesn’t wait for the agent era. SPVs have quarterly debt service. Thompson’s demand is real; the question is whether it arrives before the capital structure it supports buckles.
The capital layer
The capital layer sorts through the exposure the infrastructure build-out created. Oracle is patient zero. Larry Ellison has pledged 40 per cent of his shares – $60 billion – as collateral for personal loans. The company signed a $300B, 7.1 GW contract with OpenAI it cannot afford; free cash flow is negative $24B. When the SPV structure meets a balance sheet this leveraged, the non-recourse fiction collapses. The margin call on Ellison’s shares becomes the margin call on the edifice.
Exposure isn’t unique to Oracle: it runs through any company taking 15–19-year positions to serve five-year promises, collateralised lending against hardware with uncertain depreciation, and circular equity inflating hyperscaler earnings. This sorting could arrive quietly or suddenly. It doesn’t require technology failure.
A concrete signal arrived this week. Chinese memory-chip maker CXMT listed in Shanghai and briefly became the country’s most valuable company at $400B, price multiplying sixfold day one. Same day: sell-off wiping 10% from Samsung/SK Hynix, 9% from ASML, 5% from NVIDIA – Korean exchange paused trading.
The trigger was Chinese DUV lithography production, but underneath, two capital-structure stories broke:
One: NVIDIA reportedly guaranteeing $250B in debt for a 10 GW SB Energy data centre in Ohio (SoftBank affiliate) that buys NVIDIA chips to rent to OpenAI – all three mutually invested and cross-guaranteed. Zitron’s follow-up “The More You Buy, The More You Lose” called this the “final boss of circular financing”: chipmaker, cloud builder, and AI lab mutually insure each other’s debt.
Two: market pricing chip corrections in hours, not quarters. The slow repricing is accelerating. It’s not theoretical.
The deployment layer
The deployment layer sorts through whether AI investment in large organisations lands where it generates returns. An NBER study tracking 100,000 GitHub developers found 80% more app releases against flat end-user adoption. Uber burned its annual AI coding budget in four months; the COO noted no direct line from spend to user-facing features.
The pattern: deployment at scale requires integration depth a standalone model doesn’t provide; most enterprise investment buys the model, not the integration. Companies with genuine AI leverage – the top 1% in Ramp’s AI Index, spending $7,449/employee/month vs. $11.38 median – built that integration. The median hasn’t.
These three sortings are structurally independent. Model-layer sorting could complete while deployment-layer sorting barely starts. Capital-layer sorting could accelerate without touching deployment.
What the subprime layer is
None of this argues against the technology. In a LinkedIn essay, Raoul Pal describes an economic singularity: AI, robotics, energy, crypto converging at inflection points, each amplifying the others. Demis Hassabis argues AGI is years, not decades, away; the comparison is electricity, not the internet.
The mortgage market was also real. In 2007, Americans needed housing; demographics supported building. The problem wasn’t the asset class but the synthetic leverage layer: CDOs turning individual mortgage risk into opaque, system-wide exposure nobody could price because it never traded with price discovery.
The AI infrastructure buildout has produced the same structure, same opacity, same reason. The GPU is the mortgage – a physical asset valued against projected cash flows, financed through structures treating projections as certainty. The SPV is the CDO – a legal vehicle isolating risk on paper but distributing it across investors, asset managers, and bank balance sheets in ways revealed only during repricing.
The subprime layer of the AI economy is the part that never trades: junior tranches, circular equity, off-balance-sheet lease obligations living in SEC footnotes. Because each SPV is a bilateral contract between one AI customer and one capital pool, there is no secondary market, no price discovery, no public signal – the leverage stays invisible until an SPV fails and the chain reaction starts.
Surviving the AI sorting
Amazon survived the dot-com crash; Google was built on its rubble. NASDAQ lost 80% over two years; the internet happened anyway. Transformation and financial crises ran in parallel, each with its own logic and timeline. The crash didn’t refute the internet’s worth; it settled the capital structure question while the technology question was decided simultaneously.
That parallel is the likeliest template. Pal’s S-curves are real. Evans’s commodity trajectory is real. Zitron’s structural exposure is real. None cancels the others; each operates on a different stack level. The subprime layer is what you see when you stop arguing about whether the technology is real and ask who pays for the structure financing it.
The first piece in this series described a sorting underway while headlines waited for Seoul. The sorting has gained detail: three of them, not one, not sequential. What’s clearer is which layer matters most to understand not where AI is going, but what breaks first on the way there.
The subprime layer is what you see when you stop arguing about whether the technology is real and ask who pays for the structure built to finance it.
Photo by Mesut Yalçın on Unsplash