Three hundred billion dollars. That is the reported size of the chip-backed special purpose vehicle market now underwriting Big Tech's artificial intelligence buildout. I did not find this figure in an audited filing. I found it in a single crypto media dispatch, hedged by the word 'reportedly.' That hedge is the first red flag. The second is the collateral itself. These vehicles pledge AI accelerators โ H100s, H200s, Blackwell racks โ against debt, and they value that collateral on book. The book assumes a five-to-six-year useful life. The physical secondary market clears the same silicon in two to three. A single line of logic can unravel a thousand lies: if the pledged asset is worth less than the loan secured against it, the loan is already impaired โ it simply has not been marked yet.
This is not a story about chips. It is a story about where AI risk has been relocated.
Context: how compute became collateral
For three years the AI trade ran on equity and retained earnings. That channel is closing. Hyperscaler capital expenditure crossed roughly $200 billion in 2024 and is tracking above $300 billion for 2025, a spend intensity that now rivals the telecom buildout of the late 1990s. When internal cash flow stops covering the bill, the balance sheet reaches its leverage ceiling, and the financing migrates off-book. That migration has a name: the special purpose vehicle.
The architecture is consistent. A vehicle is stood up around a specific asset โ a Meta Hyperion data center, a CoreWeave GPU cluster, an Oracle AI campus โ and funded through private credit. Blue Owl, Blackstone, PIMCO supply the junior capital; JPMorgan, MUFG and others layer in bank debt; the whole stack is securitized and sold into the bond market. The SPV sits in the middle as a pure financial intermediary. It manufactures no compute. It earns a spread on asset securitization, and its entire value is parasitic on one thing: the rental cash flow the underlying chips can generate before they depreciate into obsolescence.
That is the load-bearing assumption. Everything else is packaging.
Core: the depreciation gap is the whole trade
Here is the mechanical problem, and it is not subtle. NVIDIA ships a new accelerator generation roughly every eighteen to twenty-four months. A100 gave way to H100 in 2022, H200 in 2024, Blackwell across 2024 and 2025, with Rubin on the 2026 roadmap. Each generation delivers two to three times the throughput and roughly double the efficiency. The direct consequence is that older silicon does not hold value โ it loses it on a schedule set by the vendor's own release cadence.
Hyperscalers depreciate these assets over five to six years. The economic reality, argued publicly by skeptics including Michael Burry in late 2025, is closer to two to three. The gap is not academic. One external estimate puts the cumulative under-depreciation at roughly $176 billion across 2026 to 2028. That figure is unaudited and I treat it as directional, not precise. But the direction is unambiguous.
Cold eyes see what warm hearts ignore: depreciation is an accounting choice, not a physical fact. The SPV values its collateral at book. Book value is a function of the depreciation schedule. Stretch the schedule and the collateral looks healthier than the market would ever pay for it. The net asset value of the vehicle and the liquidation value of its chips are two different numbers wearing one label.
Now follow the chain. Accelerated iteration pushes used-card prices and cloud rental rates down. A100 rental economics have already deteriorated sharply. When the collateral is old silicon, its loan-to-value ratio drifts upward without anyone touching the loan. Push LTV past a covenant threshold and the borrower faces a margin call. Fail the call and the collateral is sold into a market that is already falling. That is a negative feedback loop, and it is structurally identical to the 2008 mortgage machine โ mark-to-model collateral, off-balance-sheet leverage, and a trigger that only fires when prices move.
There is a second layer the bull case ignores. These vehicles are non-recourse or limited-recourse to the sponsor. The depreciation risk does not land on Meta's income statement. It lands on the private credit funds and, ultimately, on the bondholders. That is the exact point where financial opacity stops being a talking point and becomes a transmission channel.
And there is a crypto-native edge to this. The private credit funds and GPU lenders here overlap heavily with the same capital that financed the 2022 crypto lending complex. Blue Owl sits on both a Meta Hyperion vehicle and CoreWeave's balance sheet. When one borrower's collateral reprices, the same fund's other positions take the mark. Concentration like that does not diversify risk. It clusters it.
The historical rhyme is exact enough to be uncomfortable. In 2000, telecom equipment vendors lent their customers the money to buy the vendors' own gear, booking revenue against demand they had financed into existence. Lucent and Nortel collapsed when the loans went bad and the demand evaporated in the same quarter. The AI version substitutes GPUs for fiber and private credit funds for vendor balance sheets, but the circularity is intact: capital creates the demand that justifies the capital.
I have audited this shape before. In 2022 I traced the UST de-peg block by block and watched a $40 billion liquidity drain execute in real time while the market narrated it as betrayal. It was never betrayal. It was a broken incentive mechanism with a predictable terminal state. The chip-backed SPV is the same species of error: a structure whose solvency depends on an assumption the underlying technology is actively destroying.
Contrarian: the bulls are not wrong about the demand
I will not pretend the bear case is airtight, because it is not. The cash flows here are real in a way that most crypto credit was not. CoreWeave rents genuine GPUs to genuine customers at genuine margins. Hyperscaler training demand is concentrated but it is paid for. NVIDIA's gross margins above 70% reflect a real, defensible monopoly, and no ASIC program has yet cracked it. Unlike the 2022 crypto lending complex โ Celsius, Three Arrows, BlockFi โ which pledged illiquid tokens against illiquid tokens in a closed loop, these vehicles pledge a productive asset with an observable rental yield. That distinction matters, and dismissing it is lazy.
The bulls are also right that the demand curve, not the supply curve, is the variable that decides this. If AI inference monetizes at scale, rental income can outrun depreciation and the LTV problem never triggers. The bet is not fraudulent. It is levered. And there is a real difference between a levered bet and a lie.
Takeaway
Watch the depreciation schedule, not the headline number. The day a major hyperscaler or its auditor is forced to shorten useful life from six years to three, the collateral behind hundreds of billions in private credit reprices overnight. The vendor-financing parallel to the 2000 telecom bust is not decorative โ it is the blueprint. The question is not whether AI compute is real. It is who holds the bag when the book value and the clearing price are finally forced to agree. The structure is not illegal. It is merely fragile, and fragility has a habit of arriving all at once rather than on schedule.