Last quarter, a mid-tier neocloud operator refinanced a GPU-backed facility at an implied collateral life of six years. The secondary market for the same silicon — H100s, mostly — was clearing at prices consistent with two to three. That thirty-six-month gap is not an accounting footnote. It is a solvency event waiting for a trigger. Jefferies' Chris Wood gave it a headline last week: "massive capital destruction." The crypto commentariat responded the only way it knows how — repricing every AI-adjacent token by reflex. The reflex is wrong. The structure is what matters. I have audited this structure before. Not in data centers. In crypto lending.
In 2022, I modeled Compound Finance's oracle exposure through the Terra collapse and calculated that a 15% deviation in price feeds could liquidate roughly $2 billion in positions, because the lighthouse nodes updated slower than the market moved. The lesson was never that oracles lie. It was that failure propagates through the financing layer, not the application layer. Applications survive. Leverage does not.
In 2022 the crypto lending complex — Celsius, BlockFi, Three Arrows — failed for a reason that had nothing to do with crypto being fraudulent and everything to do with duration. They borrowed short against assets they assumed were long, priced those assets off feeds that lagged, and discovered in a single week that the assumption was the whole business. The AI infrastructure stack is running the same play with better branding.
The AI buildout has reproduced that architecture one asset class over. Through 2023, hyperscaler capex came from operating cash flow — Microsoft, Google, Meta, Amazon funding data centers out of earnings. That is equity-financed risk, and equity absorbs disappointment. Through 2024 and 2025, the marginal dollar migrated. Neocloud operators levered up. Data centers moved into special-purpose vehicles. Private credit funds took the senior tranches. GPU-backed lending desks appeared, treating silicon the way 2021 treated Bitcoin.
Read that list again, because it is a protocol diagram. The GPU is collateral. The residual-value assumption is the price feed. The debt maturity is the liquidation threshold. And the depreciation schedule — six years, five years, occasionally four — is the oracle.
The depreciation schedule is the oracle, and nobody is auditing it. This is where the crypto analogy stops being rhetorical. In DeFi lending, a collateral asset is only as safe as its price feed's latency and manipulation resistance. In AI infrastructure finance, the collateral asset is a GPU whose true economic life nobody can measure, and whose price feed is a vendor's own residual-value assumption. There is no Chainlink for silicon. There is a spreadsheet, a leasing executive, and an incentive to stretch.
Here is the mechanical problem. If H100-class silicon has an economic life of six years, a GPU-backed loan amortizes comfortably. If the true figure is three — and the arrival of cheaper, more efficient inference silicon keeps pushing the estimate down — then every such facility is under-amortized by half. The borrower's equity is gone before the loan matures. That is not a valuation question. It is a duration mismatch, the same one that killed crypto lenders in 2022: long-duration assets, short-duration liabilities, and a price assumption nobody stress-tested.
The efficiency shock compounds it. DeepSeek's January release was, in structure, an oracle deviation. It told the market that a given unit of capability no longer required a given unit of compute. When the cost of producing intelligence falls, the value of the capital stock built to produce it at the old cost falls with it. On the application layer, that is a gift. On the infrastructure layer, it is a markdown.
The revenue side makes the arithmetic harsher. AI application revenue is growing fast — nobody disputes that. But it is growing from a small base against a capital stock measured in hundreds of billions. The gap between annual AI revenue and annual AI capex is the only number that settles this argument, and it is a number neither Wood nor his critics published. If the ratio is converging, the debt is serviceable. If it is diverging, the depreciation schedule is the only thing standing between the buildout and a write-down cycle.
Note what Wood's warning does not contain. No debt figure. No maturity wall. No trigger condition. "Massive capital destruction" is an unfalsifiable phrase, which is exactly what makes it durable as a strategy note. Code does not lie, but it often omits the truth — and so do strategists. The real quantity that matters is the debt share of AI capex, its weighted-average maturity, and the residual-value assumptions embedded in it. None of those were in the article. All of them are the actual risk.
Now the counter-intuitive part, and it is where crypto stops being a metaphor. The chain is only as strong as its weakest node, and the weakest node in the AI capital stack is not Microsoft. It is the leveraged middle: neoclouds, data-center SPVs, private-credit AI exposure, and the power projects built to feed them. Hyperscalers absorb disappointment with cash flow. The middle tier absorbs it with equity, and the middle tier has very little. So the failure, when it comes, will be a credit event, not an earnings event.
This is also why the reflexive crypto repricing is wrong. Decentralized compute networks — the DePIN layer, Render, Akash, and their peers — do not carry residual-value risk, because they do not own the asset. They price spot, they monetize the depreciating tail of someone else's capex, and they are short-duration by construction. A GPU write-down at a neocloud is, structurally, a supply increase to the spot compute market. The bear case for AI infrastructure is the bull case for decentralized compute. Almost nobody is trading it that way.
I ran a version of this benchmark myself. In 2025 I designed a protocol to verify AI inference with zero-knowledge proofs, cutting verification overhead about 30% against existing methods. The point was never throughput. It was auditability. If inference can be proven, then utilization can be proven, and if utilization can be proven, residual value stops being a spreadsheet and becomes a verifiable metric. That is the actual fix for the depreciation problem — not better forecasting, but cryptographic attestation of the asset's real workload. AI compute financing has an oracle problem. Cryptography is the oracle.
Scalability is a trilemma, not a promise — and so is collateral. Watch three signals over the next two quarters: whether any hyperscaler shortens its GPU depreciation schedule, whether secondary H100 prices keep sliding, and whether neocloud credit spreads widen ahead of earnings. Those three numbers, not the headline, will tell you when the markdown starts. The capital will not be destroyed. It will be reallocated — to whoever can prove what the hardware actually did. When strategists reach for words like destruction, the market has stopped asking whether the buildout is real and started asking whether it pays. That shift is the event.