Last November I sat with a depreciation schedule pulled from a hyperscaler's filing and ran the arithmetic twice, because the first pass felt wrong. Five to six years, straight-line, for racks of H100 accelerators. Six years โ on silicon that had already been leapfrogged twice between the drafting of that table and the signing of the document. That single line, an accounting choice dressed up as a physical fact, is where this story begins. This month, reports have moved through crypto and private-credit desks that special purpose vehicles collateralized by AI chips now aggregate to roughly $300 billion. The figure is unaudited, relayed from a single industry brief, and I treat it the way I treat any number that leans on "reportedly." But the structure behind the number is not a rumor. It is the shape of how artificial intelligence is now financed โ and the shape is what should keep you awake.
Special purpose vehicles are not new. The structure is a legal shell that isolates a specific pool of assets, and the debt raised against them, from the parent balance sheet. What is new is what is being placed inside. Historically these shells held mortgages, receivables, aircraft leases, toll roads โ assets with long, legible lives. The AI variant holds graphics processors, networking fabric, and the memory wrapped around them. The supply chain feeding these shells runs from NVIDIA at the top, through server integrators like Dell, Supermicro, and Foxconn, into data centers that consume land and electricity, operated by hyperscalers such as Meta, Oracle, and Microsoft alongside a newer tier of "neoclouds" โ CoreWeave, Lambda, Crusoe. The capital stack beneath them draws on private credit houses โ Blue Owl, Blackstone, PIMCO โ then banks, then eventually the bond market. Meta's Hyperion data center, on the order of $27-30 billion, sits inside a Blue Owl-led SPV structure. Oracle's AI infrastructure debt runs into the tens of billions. CoreWeave has closed multiple GPU-backed loans. Stack them together and you approach the reported $300 billion.
The SPV here earns a securitization spread. It manufactures no chips and sells no compute. Its value is parasitic in the strict sense โ it lives entirely off the cash flow of the silicon underneath. The entire edifice rests on one assumption that nobody has stress-tested in public: that a GPU's accounting life matches its economic life. It does not. And that gap is the fault line.
The technology cadence is what drives the collapse. Trace the accelerator lineage: A100 arrived in 2020 on a 7nm node, H100 in 2022 on 4nm, H200 in 2024, Blackwell's B200 and GB200 in 2024-25, with Rubin on the 2026 roadmap. Each generation delivers roughly two to three times the performance and about twice the efficiency, on a refresh cycle compressed to eighteen months to two years. Traditional server CPUs refresh on three to five years. This is an accelerated Moore's Law, and its financial meaning is blunt: the physical useful life of the asset and its booked depreciation life are systematically mismatched.
Hyperscalers book AI servers over five to six years, some stretching to six. The skeptics โ Michael Burry went public with this in November 2025 โ argue the true economic life is closer to two to three, because each new generation crushes the rental price and utilization of the one before it. Burry's estimate puts under-depreciation across the leaders at roughly $176 billion cumulatively across 2026-2028. That figure is an external calculation, unaudited, and I flag it as such. But the mechanism does not require the number to be exact. Iteration accelerates, the resale and cloud-rental price of old cards falls, and any SPV collateralized by those cards watches its loan-to-value ratio drift upward on its own โ before a single payment is missed. The trigger for a margin call or a default is baked into the physics of the roadmap.
When I audited the early liquidity pools of Curve during the 2020 DeFi summer, I learned to read incentive structures the way you read a lease: not for what they promise, but for who absorbs the loss when the promise breaks. The chip-backed SPV is the same lesson in a different costume. The collateral is marked at book value, book value depends on the depreciation assumption, and an aggressive assumption inflates the collateral. The valuation chain is circular by construction.
There is a ghost here, and it has a name. In 2000, Lucent, Nortel, and Motorola lent money to customers so those customers could buy their own equipment. Vendor financing manufactured a boom that looked like demand and was in fact credit. When the demand was falsified, the equipment makers collapsed alongside their borrowers. The AI version of vendor financing is already visible: chip vendors and creditors co-funding compute purchases, a loop where NVIDIA invests in customers, customers buy chips, and revenue flows back to NVIDIA. The demand is real in accounting terms. In economic terms it may overstate how much of that demand is terminally paid for rather than credit-created. A dollar of revenue booked is not the same as a dollar of demand that survives a tightening cycle.
The demand structure sharpens the risk rather than softening it. Model training still consumes an estimated forty to fifty percent of capacity, concentrated among a handful of frontier labs โ a client base narrow enough that one slowdown cuts deep. Inference is rising, perhaps thirty to forty percent, but it carries lower margin and is more easily served by cheaper silicon: in-house ASICs, older cards, algorithmic efficiency gains that quietly retire demand for the newest hardware. So the exposure is not that demand disappears. The exposure is that demand degrades โ shifting from high-margin training to low-margin inference while the collateral underneath it depreciates. That combination compresses the SPV's revenue side and its collateral side at the same moment. A double squeeze. Revenue falls, collateral shrinks, LTV worsens, default risk climbs. None of this needs a villain. It needs only time.
Which raises the question I keep returning to: why does this financing exist off-balance-sheet at all? The answer is itself the warning. If a company could fund AI capex cheaply on its own balance sheet, it would. The migration to SPVs, private credit, sovereign funds, and off-book structures tells you that on-book leverage is already near its ceiling. The shadow-ization of AI capex is not a financing innovation; it is a symptom that the transparent channel is full. And it carries a second hazard, one the coverage rarely names: participant overlap. Blue Owl sits on both the Meta Hyperion structure and CoreWeave's funding. When the same capital appears on both sides of a trade, you are not looking at diversification. You are looking at concentration wearing a diversified mask. The "financial opacity" these deals are politely described as having is, in plain terms, related-party risk that no outside investor can price.
Now add the competitive geometry. NVIDIA holds an estimated eighty to ninety percent of the accelerator market โ it sets the pricing and the cadence that define every collateral value in the chain. The SPV has no pricing power upstream. Downstream, its tenants are a handful of hyperscalers and labs with enormous negotiating leverage; a single tenant default can pierce an entire SPV. The intermediary is squeezed at both ends, earning a thin spread while carrying the depreciation risk of an asset it cannot control. And the competition among these intermediaries is, at bottom, a leverage contest: whoever dares to use more off-balance-sheet debt expands faster in the short run. That is a race to the bottom, and it incentivizes the whole industry toward over-indebtedness.
Here is where the conventional read gets it backwards, and where I part with the crowd. Everyone watching this space is braced for an AI bubble popping in equities โ a violent, visible, ticker-tape event. The more probable path is quieter and therefore more dangerous: an AI credit event. The risk has been transferred from the equity market, which can absorb volatility, into the credit market, which is unforgiving of default. Liquidity flows, but trust evaporates โ and in private credit, trust evaporates without a headline, because there is no exchange to ring a bell. Repricing happens in rooms, not on screens, until a margin call cascades and the refinancing window slams shut.
There is a contrarian wrinkle in the geopolitics that almost nobody prices. Export controls on advanced AI chips are typically read as a headwind for the hardware trade. For the SPV holder, they may be a perverse tailwind. Stricter controls slow the cadence of new generations reaching restricted markets, which slows the depreciation of the installed base. Loosen the controls, fragment the market, and you accelerate the decay of the collateral instead. The same logic runs through the localization push โ CHIPS Act subsidies and national fab programs raise long-run supply, which feeds price competition and speeds the obsolescence of every card currently pledged as security. The tail risk embedded in these structures is policy-driven depreciation: a rule change that, overnight, alters which markets a batch of chips can be sold into and what they are worth. No depreciation schedule on earth captures that.
I spent part of 2025 inside a Frankfurt bank, drafting narrative strategy for its first crypto allocation and translating blockchain mechanics into the language of intergenerational wealth. The lesson I carried out of those closed-door workshops is the one I keep applying here: institutions do not buy assets, they buy stories that let them sleep. The chip-backed SPV is a very good story โ AI, growth, collateral, yield. What it is not is a story with a stress test. And this is a bear market, which means the question is no longer how much you can gain. It is whether the thing holding your collateral can survive a repricing it never modeled.
The tell to watch is not the price of NVIDIA, nor the next model release. It is the depreciation schedule. If a major hyperscaler or its auditor is forced to shorten the assumed life of AI servers from six years toward three, the collateral revalues downward, margin calls follow, forced selling follows, and prices fall further โ the same reflexive loop that ran through 2008, wearing different clothes. Code is law, but narrative is truth, and the narrative currently holding $300 billion of collateral in place is a single accounting line. Don't trade the chart; trade the story โ and right now the story has a footnote that nobody wants to read aloud. The next credit event will not announce itself. It will be filed, quietly, as a change in estimate.