The AI-CDO: Securitization Risk in AI-Infrastructure Credit
and the Revenue Variable That Governs Its Tail

In one line: I simulated AI debt. Losses hit the risky slices. The safe slice breaks only if revenue falls 40 percent short.

Abstract. The financing of artificial-intelligence infrastructure is reassembling the machinery of structured credit: data-center developers, “neocloud” compute providers, and chip-collateralized lenders linked by circular financing that the Bank for International Settlements flagged in its 2026 Annual Economic Report. I argue that the four defects behind the 2008 crisis (corrupt ratings, hidden correlation, inadequate tail capital, and misaligned incentives) are each observable in this complex today, with technology obsolescence adding a hazard mortgages never carried. A stylized Monte Carlo of a three-tranche pool (equity 15%, mezzanine 25%, senior 60%) over 200,000 draws shows the equity tranche losing 68% of its value on average, the mezzanine 28%, and the senior 1% while remaining intact in 93% of paths. The asymmetry makes the originator’s cash-out structural and isolates a single governing variable: realized revenue as a fraction of the promised AI economics. The senior claim bleeds only once revenue falls below roughly 60% of what was promised, the margin that separates this boom from the sub-prime one.

1. Introduction

Securitization did not cause the 2008 crisis. Pooling loans and selling tranched claims on the pool is a durable financing technology. What failed in 2008 was a set of four specific, nameable defects that traveled inside a securitization; once named, they are observable wherever the same machine reassembles. It is reassembling now around the financing of artificial intelligence. This note is an inventory of those four defects in the AI-credit complex and a stylized simulation of how losses move through it. It describes the present structure and does not forecast the timing of any outcome.

2. The AI-infrastructure financing complex

The AI buildout requires an extraordinary amount of capital, and the most profitable companies in the world have begun borrowing to fund it because internal cash flow no longer keeps pace. A financing complex has formed around that demand: data-center developers borrowing against future rents; “neocloud” providers borrowing against multi-year compute contracts; and lenders increasingly willing to take graphics chips themselves as collateral. The Bank for International Settlements, which warned about mortgage risk years before 2008, devoted part of its 2026 Annual Economic Report to the sector. Its language is deliberately careful: the sector shows “circular financing,” its “opacity” “compounds vulnerabilities,” collateral is in places “pledged multiple times,” and “signs of stress are already visible.” A chipmaker takes an equity stake in an AI lab; the lab commits to buy that maker’s chips; a cloud provider borrows to house them; an insurer is offered the senior claim on that debt as a safe asset. Money runs in a circle, and each turn is booked as growth.

Scale makes the loop dangerous rather than merely clever. Widely reported estimates put this year’s hyperscaler AI capital spending near three-quarters of a trillion dollars and the multi-year buildout in the trillions. The money entering from outside the circle, what end users and enterprises actually pay for the output, is smaller by roughly an order of magnitude. The structure therefore rests on a single bet: that the small outside number grows faster than the large recycled one. That bet is the whole risk.

3. The four failures of 2008

Securitization was the vehicle in 2008; the following four defects were its passengers.

1. The rating was corrupt. Pools of poor mortgages were stamped AAA. Institutions trusted the stamp rather than the mortgage, and the stamp was paid for by the parties being rated.
2. The correlation was hidden. The models assumed one borrower’s default said little about the next. The assumption held in calm and dissolved in stress. When house prices fell, they fell everywhere at once, and pools believed to be diversified proved to be a single bet.
3. The tail capital was inadequate. Buffers were sized for the normal distribution the models drew, not for the fat-tailed distribution the market realizes.
4. The incentives were misaligned. Under originate-to-distribute, whoever wrote a loan sold it onward and retained none of the risk, so the reward for volume swamped the reward for care.

4. The four failures in AI-credit form

4.1 Ratings

Debt backed by data-center rents and depreciating chips is new enough that its rating apparatus is young. The 2008 lesson is not that ratings are useless but that a rating is worth only the independence and the model behind it, and a young market is precisely where an under-modeled, conveniently generous grade does its damage.

4.2 Correlation and obsolescence

These loans are not independent. Nearly all of them lean on the continued spending of a handful of hyperscalers and on the fortunes of essentially one chip designer. A pool that looks diversified can be, underneath, a single concentrated bet on whether that spending continues. That is the 2008 correlation trap with fewer moving parts. AI credit also carries a hazard mortgages never had, technology obsolescence. A house securing a loan is worth roughly what it was worth last year; a graphics chip securing a loan sits on a steep replacement curve, and its collateral value can fall for reasons unrelated to the borrower’s health and entirely to do with a better chip shipping.

4.3 Tail placement

The tail increasingly lands on insurance balance sheets. Private credit, including the infrastructure paper discussed here, is sourced for the long-dated liabilities of annuity books precisely because it yields more than public bonds of the same stated grade. That excess yield is payment for illiquidity and tail risk. The question a risk seat must answer before the paper is bought is whether the buffer behind those liabilities is sized for the fat-tailed world or the modeled one.

4.4 Incentives

Originate-to-distribute is a structure, not an era; it returns wherever the writer of the risk can sell it onward. The honest tell is skin in the game: whether the party that structures the pool retains a real, loss-absorbing piece of it.

5. What the buyers say, on the record

The argument above is built from published aggregate data. In Spring 2026 Stanford ran a course on the economics of this build-out (MS&E 435), and the guests were the people doing the spending rather than the people underwriting it. Their numbers are worth recording precisely because none of them is a bear, and each has an interest in the scale sounding either large or manageable. I attribute rather than endorse.

OpenAI’s compute lead put the all-in cost of a gigawatt of AI capacity at $70 billion, and a gigawatt at roughly half a million GPUs. Against his own stated targets, that arithmetic is the story: OpenAI describes an aspirational 30 gigawatts by the end of the decade, and he estimates US hyperscalers collectively plan about 100 gigawatts, which he says would consume a double-digit percentage of US grid capacity. He also states that compute and revenue have each tripled year over year for three years, a claim that is unaudited and made by the person paid to make it true.

Three details matter more to a credit analyst than the headline totals. First, he names ASML as the single chokepoint of the entire supply chain. That is the hidden-correlation problem stated by the buyer: nominally independent AI-infrastructure credits share one upstream node that no tranche structure diversifies across. Second, he describes a three-year chip design cycle and hardware that throttles delivered performance, alongside an admission that general-purpose GPU compute is already inadequate for the target workload. Laid against operators depreciating that hardware over four to six years, that is an impairment case assembled from friendly testimony. Third, a founder on the same course priced Nvidia’s margin at roughly 75% and put the insourcing indifference point at custom silicon 80% as effective. If that option is ever exercised, the supplier’s revenue and the borrowers’ credit are revealed as one risk written twice, because the supplier’s customers and the credit’s borrowers are the same handful of firms.

The depreciation point deserves its own sentence. A four-to-six-year write-off on silicon co-designed for one model architecture is not a wear estimate. It is an unhedged position on one side of an unsettled scientific question about whether that architecture persists, taken by the party whose reported earnings improve if the bet wins. That is mark-to-model risk in the 2008 sense, relocated from a ratings model to a fixed-asset schedule.

What is absent is as informative as what is present. Across roughly three hours with senior operators, debt is never discussed. There is no financing structure, no lease, no special-purpose vehicle, no covenant. The people building the demand side are not thinking about the liability side at all, which is precisely the condition under which a financing complex assembles itself without anyone treating it as their problem.

6. A stylized loss-waterfall simulation

5.1 Design

The useful question is not whether the complex is fragile but where losses land when they arrive. I model the AI-infrastructure debt as a single securitized pool sliced into three claims that absorb losses from the bottom up. The equity tranche (15% of the pool) is the neocloud, lab, and circular-financing paper and takes the first loss. The mezzanine (25%) is the GPU- and equipment-backed debt. The senior (60%) is secured by the assets that survive a technology cycle: land, buildings, and power contracts. Every tranche depends on one variable, R, the fraction of the promised AI economics that arrives as revenue. I let R range from a bear case (half of what was promised) to a bull case (more than promised), draw 200,000 paths, and let losses waterfall up through the stack. The design is deliberately simple, its inputs are declared in the code, and it is not a forecast.

5.2 Results

The columns of Figure 1 read left to right as revenue deteriorates; red marks loss filling each stack from the bottom. In the good state (revenue near-promised, the 25th-percentile draw) equity holders take a haircut and every senior claim is untouched. In the bad state (the median draw, revenue merely disappoints) the equity is wiped out entirely, and its holder matters. Increasingly it is yield-seeking buyers sold the tranche as “diversified AI exposure.” In the ugly state (the 95th-percentile draw, revenue collapsing toward half) equity and the GPU-backed mezzanine are both destroyed and the senior claim finally bleeds, because obsolescence leaves chip collateral little recovery and circular financing unwinds faster than a single-name default.

Across all 200,000 paths the asymmetry is stark: the equity tranche loses 68% of its value on average and is all but wiped in half of them, the mezzanine loses 28%, and the senior loses 1% and stays fully intact 93% of the time. That gap makes the originator’s cash-out structural. Origination fees are earned whatever happens; the senior, real-asset-backed claim that survives almost everything is retained; and the equity and mezzanine, the tranches that carry the technology risk, are sold to insurers and pensions reaching for yield. The 2008 question is whether the sellers kept a real slice of the loss for themselves.

7. Scenarios the model rejects

The simulation is honest only about the myths it refuses. Three that it rejects:

• “GPUs hold value like real estate.” A building is worth roughly last year’s price; a chip sits on a replacement curve and can lose collateral value the day a better one ships, for reasons unrelated to the borrower. That is why the mezzanine, not the equity, is the tranche that surprises.
• “It is diversified across hundreds of borrowers.” Diversification across borrowers who all depend on the same handful of hyperscalers continuing to spend is not diversification; it is the mechanism that turned “uncorrelated” mortgage pools into a single wager in 2008.
• “The senior tranche is safe.” Only if it is secured by real assets rather than claims on other claims. A senior slice of circular financing is senior in name and equity in substance. Safety lives in the collateral, not the label.

8. The revenue variable

The variable that separates this boom from 2008 is the one the simulation turns on. A house produces no cash flow and only ever costs; compute can produce cash flow if the applications built on it are paid for. The senior claim survives in 93% of the draws for a single reason: revenue must collapse below roughly 60% of what was promised before it bleeds, and sustained revenue at that level is a materially different world from the sub-prime one, where the underlying asset could never service anything. The saving grace of the AI boom is revenue: enterprise adoption, inference demand actually invoiced, applications that are renewed. That is the flow that fills the waterfall from the top and services the stack on the way down. It is the number to watch rather than the capital-expenditure headlines. Another hundred billion dollars of announced data-center spending measures the size of the promise; the payment has not yet arrived.

9. Scope and discipline

This note does not call a top, name a date, or claim the AI buildout will fail; on a decade’s horizon it may pay off, as the internet did. Timing is a forecast, and forecasts belong in the Forecast Wing under resolution criteria. The exercise here is narrower: to show that each of the four defects that broke in 2008 is observable in the AI-credit complex today, to run the stack forward for the shape of the outcomes, and to name the single variable that decides between them. 2008 was not a failure to predict the future but a failure to describe the present, to state, while the music played, that the diversification was fake, the rating was bought, and the safe tranche was only ever as safe as the revenue beneath it.