Follow the dollar. It never leaves the room.
NVIDIA invests in a neocloud. The neocloud signs a multi-year compute contract with a hyperscaler. The hyperscaler commits to buy NVIDIA accelerators. NVIDIA books the revenue, and the same capital re-enters the loop. One dollar, counted four times across three income statements. That is not a market clearing price. That is a closed circuit.
I have seen this exact shape before. In 2020 I deployed $50,000 across Uniswap V2 and Compound and wrote a Python bot to farm the spread. The yield looked real until a Sushiswap fork spiked gas and erased 40% of my gains in an hour. The lesson was not about gas. It was that recursive structures report profits that only exist as long as new capital keeps entering. Code doesn't lie. It just waits.
The structure, stripped of branding
Strip the logos and the AI buildout is a classic vendor-financing loop. A chip vendor takes equity in a compute provider. That provider pledges the contract as collateral. Private credit lenders underwrite the pledge against GPU collateral. A sovereign fund injects billions to secure compute sovereignty. The hyperscaler's balance sheet absorbs the demand signal. Each hop adds a layer of leverage and removes a layer of transparency.
This is the anatomy of the 2021 stETH loop, translated from Solidity into term sheets. On-chain, we called it recursive lending. Off-chain, bankers call it strategic investment. The mechanism is identical: the asset backing the loan is also the asset being purchased with the loan.
Concretely: NVIDIA has taken equity stakes in neoclouds and backed multi-gigawatt commitments; a model lab has signed a multi-year, multi-billion-dollar compute deal with a cloud provider that then buys the vendor's chips; a chip rival has locked a flagship customer through warrants, tying equity to purchase volume. Sovereign vehicles, MGX, PIF, Mubadala, and SoftBank's Vision Fund, sit on the other side, writing checks to secure compute and influence. None of this is illegal. All of it is circular. And circular is the one structure that makes every headline revenue number suspect.
The collateral is the problem
The mortgage in this trade is a GPU cluster. Unlike real estate, it does not hold value. A three-year-old accelerator is not a three-year-old building. It is a depreciating, power-hungry, generationally obsolete box. When SPV structures underwrite these assets on 20-year, low-rate logic, they import a real-estate mental model onto a semiconductor. That mismatch is where the losses hide.
The accounting is where it gets lethal. If a neocloud depreciates GPUs over five to six years while their economic life is two to three, book profit is fiction. The depreciation schedule is the single most important number in AI finance, and almost nobody reads it. I would treat any extension of GPU useful-life assumptions as a red flag equivalent to a bank marking down loan-loss reserves mid-cycle.
The debt, not the equity
In 2023 the loop ran on equity. That was survivable. Equity absorbs loss silently. In 2024 and 2025 the funding shifted to debt and quasi-debt: data-center ABS, private credit, SPVs, and mega bond issuance. Debt does not absorb loss. It transmits it. Equity forgives. Debt forecloses.
Now the tell. Much of the compute demand backing these loans is multi-year, non-binding, or phased. The headline contract value is used to secure financing, but the payment terms are softer than the press release. Ask the only question that matters: how much of it is take-or-pay? If a contract can be walked away from, it is not collateral. It is a press release with a payment schedule. Exit liquidity is a myth when the buyer is also the investor.
Overseas capital is not free capital
Middle Eastern sovereign funds and SoftBank are now anchor capital. They bring real money and a dual motive: returns and compute sovereignty. That dual motive means the pricing is not purely market-clearing. Non-market pricing masks the true risk premium. You cannot mark a political allocation to a market clearing level, and you cannot price the option the sovereign is really buying. Patient capital prices risk differently.
The on-chain mirror
Here is the part crypto natives should not ignore. Tokenized compute markets, decentralized GPU networks, render and inference protocols, are replicating the same loop on-chain, with the same circular logic, but with an oracle that reports utilization as if it were revenue. Measures what matters, not what feels good. On-chain utilization is not revenue. Token emissions subsidizing demand is not adoption. I have watched three cycles of decentralized compute tokenize a narrative and call it yield. Yield is just delayed volatility. The chain does not remove counterparty risk; it just relabels it as a smart contract. And smart contracts are brittle when the collateral behind them is a depreciating chip.
The analogy everyone gets wrong
The dominant warning is a 2008 comparison. It is wrong, and the error matters because it misleads you about contagion.
2008 was household leverage, securitization, rating failure, and a wholesale funding run. There was a bank-run mechanism, a fast, systemic, depositor-driven dash for the exits. What AI faces is vendor credit and capital misallocation. That is the 2000 telecom equipment bust, Lucent and Nortel financing their own customers, not Lehman. The contagion path is slower, more localized, and lands on the weakest link: the high-leverage neocloud, the single-customer compute lessor, the SPV with a re-pricing cliff.
The blind spot is the direction of pain. Everyone assumes a bubble pop destroys the giants. History says the opposite. Survival beats speculation. When the loop tightens, cash-rich incumbents buy distressed compute and absorb stranded engineers at a discount. The giants get stronger. The firms that took vendor money and pledged their future to it get consolidated. That is not a crash. That is a transfer.
What I am watching
The signals are concrete and readable if you know where to look. First, the depreciation disclosure: any lengthening of GPU useful life is the tell that earnings quality is deteriorating. Second, the neocloud tier: the first missed coupon, the first restructuring, the first fire sale of a cluster is the loop's earliest fracture point. Third, take-or-pay enforceability: watch whether the big compute contracts survive a demand slowdown in court, not in a deck. Fourth, private credit spreads on data-center paper: if they widen while equity stays calm, the smart money already left the room.
The dollar is still circling. It has not stopped. But loops do not unwind because someone decides to stop. They unwind when the next entrant does not arrive. Watch the spread, not the story. The only question left is who is holding the chip when the music stops, and whether they ever read the depreciation footnote.
