Last quarter, a single dollar crossed four balance sheets and was booked as revenue three times. It left a chipmaker as a strategic investment, landed in an AI lab as equity, came back as a multi-year compute commitment to a cloud provider, and returned home as a hardware order. The headline that followed — "AI infrastructure demand up double digits again" — was technically accurate and entirely hollow.
That is the anomaly. Not fraud. Not even unusual. It is vendor financing, repackaged in GPU silicon and sold as a growth narrative. The street calls it strategic investment. I call it a closed loop with a marketing department.
I hunt for the story the data refuses to tell. And this one is hiding in plain sight, in the gap between what AI companies report as revenue and what their customers actually pay for with outside money.
Context
Vendor financing is not new. In the late 1990s, Lucent and Nortel extended credit to telecom carriers so those carriers could buy Lucent and Nortel gear. The revenue was real. The cash was circular. When the carriers defaulted, the equipment makers discovered they had financed their own collapse. Nortel's vendor loan book — billions — became a write-down. Lucent followed. The lesson, written in red ink, was simple: a supplier that funds its customer's purchase has not sold a product. It has made a bet, and booked the proceeds as if the bet had already won.
Now re-read the last two years of AI headlines with that sentence in mind.
In 2023, the AI buildout ran on equity. Dilutive, transparent, priced honestly. By 2025, the funding stack had mutated. Data center asset-backed securities. Private credit. Special purpose vehicles that park debt off the sponsor's balance sheet. Sovereign wealth from the Gulf and Tokyo, chasing compute sovereignty as much as financial return. The instrument changed. The risk did not.
I spent six weeks in 2017 reverse-engineering token vesting schedules, and one insight has never stopped paying rent: when a structure is designed so that the same capital can be counted as value by two parties at the same moment, someone downstream is holding a bag they cannot see. Crypto learned this at scale. Terra, yield farming, the entire reflexive token economy — all of it was vendor financing in a different costume. The AI capital stack is now relearning the same lesson, and the crypto natives, who should recognize the shape instantly, are mostly watching from the cheap seats, amused.
The mechanics of the loop
Here is the closed loop, stripped of branding. A chipmaker invests in a model company. The model company commits to buying compute. The compute provider buys chips from the chipmaker. The chipmaker's investment returns, partly, as orders. Round and round.
Name the nodes and the shape is undeniable. Chipmaker equity flows into neoclouds and model labs. Those labs sign multi-year, multi-billion compute commitments with hyperscalers. The hyperscalers order GPUs. The GPU vendor books the revenue and, sometimes, reinvests it as equity in the next lab. AMD has structured deals with warrant-based equity to bind customers. NVIDIA has seeded compute providers that then buy NVIDIA hardware. Oracle has signed compute commitments so large they dwarf its historical revenue base.
Nothing in that chain is illegal. Everything in it is fragile, because the same dollar is being counted as demand at every node. The industry is measuring its own capital expenditure and calling it market demand.
This is where the accounting gets interesting, and where most coverage stops too early. The question is not whether the revenue is real. It is whether the revenue would survive the removal of related-party money. If a meaningful share of AI "revenue" is the industry paying itself, then the growth rate is an optical illusion — a number that describes the velocity of capital inside a closed system, not the arrival of external customers.
I built a version of this test in 2020, dissecting DeFi yields. The projected APYs on Compound and Uniswap looked spectacular until you separated emissions from fees. Strip the governance-token subsidy and the "yield" was mostly dilution wearing a percentage sign. AI's headline growth is running the same play: strip the related-party flows and ask what a stranger paid for.
The depreciation knife
There is a second mechanism, quieter and more lethal. GPU depreciation. If a cloud provider assumes its GPUs have a five-to-six-year economic life, it spreads the cost of those assets across more quarters and reports higher profit. If the real economic life is two to three years — and the pace of silicon iteration suggests it might be — then the reported profit is a policy choice, not a fact.
This is not a rounding error. On a fleet of hundreds of thousands of accelerators, moving the depreciation schedule by two years swings billions in reported earnings. It is the single most consequential assumption in the entire AI infrastructure complex, and it is buried in footnotes that almost nobody reads.
Watch for the tell. If a major AI cloud provider extends its depreciation window — stretching the tail of an asset to flatter the present — that is not a sign of confidence. It is a sign that the loop needs more time to close.
The loop also has a physical ceiling, and it is not silicon. It is electricity. Compute expansion has shifted from a chip constraint to a power, interconnection, and land constraint. You cannot finance your way past a substation queue. That physical limit matters for the narrative because it caps how fast the "infinite demand" assumption can be validated — and if the capex lands before the demand arrives, the oversupply is baked in before anyone admits it.
Where the debt hides
The equity phase was honest about risk. The debt phase is not, because debt is where the loop goes to hide.
Special purpose vehicles let sponsors keep leverage off the parent balance sheet. Data center ABS packages future rental income into tradable securities, and private credit funds have rushed to hold them, drawn by yield and insulated from mark-to-market discipline. The valuations are opaque by design. There is no public tape for a private credit position in a data center SPV, so the loss — when it comes — arrives late, recognized not by the market but by a committee.
That delay is the whole point. It manufactures what I would call paper prosperity: a stretch of time during which the assets look sound, the income looks durable, and the mismatch between short-term funding and long-term assets sits quietly in a spreadsheet. Until it does not.
The overseas capital adds another layer. Sovereign funds from the Gulf and Japan are not pure financial investors. They are buying compute sovereignty, energy leverage, and geopolitical positioning. That means their pricing may not reflect the same risk premium a market-clearing buyer would demand. When the marginal buyer is not optimizing for return, the clearing price stops telling you the truth about risk. Add the export-control and foreign-investment review overlay, and the whole structure acquires a political tail that no spreadsheet prices.
The contrarian angle
Here is where I part company with the loudest voices. The dominant narrative has already settled: this is 2008 again. It is not. It is 2000.
The distinction matters, and getting it wrong will misprice the whole event. 2008 was household leverage, securitized mortgages, ratings failure, and a wholesale funding run. Its signature was a systemic stampede — overnight funding vanished and the contagion was instant and everywhere. What we have here is different in kind. It is corporate capex, supplier credit, and private-market opacity. The collateral is not a house that keeps its value; it is a GPU that loses it every eighteen months. The leverage sits with firms, not families. The contagion is slower, more localized, and more likely to be recognized by accountants than triggered by a run.
That difference flips the intuition about who gets hurt. In a 2008-style panic, the giants bleed. In a 2000-style telecom unwind, the giants survive and consolidate. The carriers that overextended died; the equipment makers that financed them took the write-downs and then bought the wreckage cheap. The parallel here is uncomfortable for everyone cheering for a crash. The most likely outcome is not a collapse. It is a clearing — and the cash-rich incumbents are the ones holding the broom.
And the crypto crowd, gleefully forecasting AI's fall, is missing the punchline. Crypto did not escape this structure. It wrote the playbook. Reflexive incentives, emissions dressed as yield, tokens whose value depended on new capital entering to pay the old — that is the same loop, just with faster feedback. The AI industry simply added lawyers, auditors, and a depreciation schedule. Crypto is not the victim of the AI bubble. It is the ancestor. Chaos is just a pattern you have not decoded yet, and this pattern has been running since the first whitepaper promised more than it could pay.
Takeaway
So here is what I am watching, and it is not the crash headline. The related-party share of AI revenue, disclosed or inferred. The depreciation policy, because a lengthened schedule is a confession. The first credit event in a neocloud or a compute-leasing SPV, which will be the earliest signal that the loop has stopped closing. And the regulators circling private credit exposure to data centers, because their attention will raise the cost of the next turn of the wheel.
Decode the script before you bet on the actor. The AI boom is real. The demand it reports is partly a mirror. The question that decides who survives is not whether the bubble bursts — it is who holds the assets when the music stops, and whether they bought them with someone else's money.