Securitization did not cause the 2008 crisis. Four defects riding inside it did, and they are reassembling around artificial intelligence: developers borrowing against future data-centre rents, neoclouds against multi-year compute contracts, lenders taking graphics chips as collateral. The Bank for International Settlements gave the sector part of its 2026 Annual Economic Report, naming “circular financing,” an “opacity” that “compounds vulnerabilities,” and collateral “pledged multiple times.” What enterprises pay for the output is small beside what circulates inside the loop. That gap is the bet.
| Defect of 2008 | Its form in AI credit today |
|---|---|
| The rating was bought | Debt on data-centre rents and depreciating chips is new enough that its rating apparatus is young, and a young market is where an under-modelled, conveniently generous grade does its damage. |
| The correlation was hidden | Nearly every borrower leans on a handful of hyperscalers continuing to spend and on essentially one chip designer, with a single upstream node behind them all. A pool that looks diversified is one concentrated bet. |
| The tail capital was inadequate | The tail is sourced for insurers’ long-dated annuity books because it yields more than public bonds of the same stated grade. That excess is payment for illiquidity and tail risk. |
| The incentives were misaligned | Originate-to-distribute is a structure, not an era. The tell is whether the party that builds the pool retains a piece of it that absorbs loss. |
| New hazard, absent from mortgages | Technology obsolescence. A building holds roughly last year’s price; a chip loses collateral value the day a better one ships, for reasons unrelated to the borrower. |
On the record
Stanford’s MS&E 435 put the buyers of compute in front of a class in Spring 2026. Table 2 records what they said, attributed rather than endorsed; none of them is a bear. Depreciating silicon over a schedule longer than its design cycle is not a wear estimate but an unhedged bet on an unsettled scientific question, taken by the party whose earnings improve if it wins. Debt itself never comes up, which is how such a complex assembles with nobody minding the liability side.
| Measure | Value | Source |
|---|---|---|
| Hyperscaler AI capital spending this year | Near three-quarters of a trillion dollars | Widely reported estimates |
| The buildout over multiple years | Trillions of dollars | Widely reported estimates |
| Money entering from outside the circle | Smaller by roughly an order of magnitude | Widely reported estimates |
| A further round of announced data-centre spending | Another hundred billion dollars, which measures the promise and not the payment | Widely reported estimates |
| All-in cost of a gigawatt of AI capacity | $70 billion | OpenAI compute lead, MS&E 435 |
| GPUs in a gigawatt | Roughly half a million | OpenAI compute lead |
| OpenAI aspiration by the end of the decade | 30 gigawatts | OpenAI compute lead |
| US hyperscaler plans, collectively | About 100 gigawatts | OpenAI compute lead |
| Share of US grid capacity that implies | A percentage in the double digits | OpenAI compute lead |
| Compute and revenue growth | Each tripled year over year for three years, unaudited | OpenAI compute lead |
| Single chokepoint of the supply chain | ASML | OpenAI compute lead |
| Chip design cycle | Three years | OpenAI compute lead |
| Operator depreciation of that hardware | Four to six years | Standard practice |
| Nvidia margin | Roughly 75% | Founder, same course |
| Insourcing indifference point | Custom silicon 80% as effective | Founder, same course |
| Debt, leases, vehicles or covenants discussed in roughly three hours | None | Author’s count |
Where the losses land
The debt is modelled as one pool sliced into three claims that absorb loss from the bottom up, tranched as in Table 3. Everything turns on R, the fraction of promised AI economics arriving as revenue.
Equity absorbs the shock and the senior claim usually holds, and that asymmetry makes the originator’s cash-out structural: fees are earned regardless, the senior slice is retained, and the tranches carrying the technology risk are sold to insurers and pensions reaching for yield. When the senior does bleed, the cause is the circular-financing amplifier lifting pool loss through the attachment point, not obsolescence striking the senior.
| Measure | Value |
|---|---|
| Draws | 200,000, seed fixed |
| Equity tranche, first loss | 15% of the pool: neoclouds, labs, circular financing |
| Mezzanine tranche | 25%: GPUs and equipment |
| Senior tranche | 60%: land, buildings, power contracts, attaching at 40% of pool loss |
| R, promised AI economics realised | Bear at half of promise to bull above promise |
| Equity mean loss | 78% |
| Equity all but wiped | 61% of paths |
| Mezzanine mean loss | 29% |
| Senior mean loss | 1.2% |
| Senior fully intact | 90% of paths |
| Senior bleeds | 9.9% of draws |
| Odds of a senior loss cross one half | R near 65% of promise |
| Bleed events at R of 60% or above | 59% |
| Richest path on which bleeding is seen | R of 75% |
| Figure 1, good | The draw at the 25th percentile |
| Figure 1, bad | The median draw |
| Figure 1, ugly | The draw at the 95th percentile |
| Claim | What the waterfall says |
|---|---|
| “GPUs hold value like real estate” | Rejected. Chips sit on a replacement curve, which is why the mezzanine surprises before the equity does. |
| “It is diversified across hundreds of borrowers” | Rejected. Borrowers who all depend on the same hyperscalers continuing to spend are one wager, as the mortgage pools of 2008 were. |
| “The senior tranche is safe” | Rejected. Only where real assets secure it rather than claims on other claims. Safety lives in the collateral, not the label. |
What this does not claim
No top, no date, no claim that the buildout fails; over a decade it may pay off. Timing belongs in the Forecast Wing under resolution criteria. The simulation is stylized, its inputs declared in the code. Every probability follows from an uncalibrated prior on R; only the attachment map is structural, so a different prior moves the probabilities but not the ordering of the tranches. Unlike a house, compute can service debt if what is built on it is paid for, so the line between this boom and the sub-prime one is a band of revenue rather than a floor. Invoiced revenue is the number to watch, not announced capital expenditure.
10. Correction, 19 August 2026
The probabilities on this page are not calibrated, and that belonged here from the start. Every number is a direct function of an uncalibrated triangular(0.50, 0.85, 1.20) prior on R, the fraction of promised AI economics actually realized. Only the attachment map, equity 0 to 15, mezzanine 15 to 40, senior 40 to 100, is structural. Change the prior and every probability moves; the ordering of the tranches does not. That caveat sat in a private working file at publication rather than on the page.
A defect in the waterfall was found and fixed, and two published headline numbers move. The obsolescence haircut was being added to mezzanine loss after the waterfall rather than impairing the collateral before it. Two consequences, both wrong. Tranche losses no longer reconstructed to pool loss: 17.74% of loss allocated against a 16.50% pool, so 124 basis points were manufactured at the mezzanine level, with a maximum single-path gap of 7.38%. And the senior tranche was left structurally blind to obsolescence, the correlation of senior loss with the haircut being −0.002, while section 6 told the reader that obsolescence was what made the senior bleed. The fix impairs the 25% GPU slice of the pool before the waterfall and carries the original 0.5 scaling unchanged, so it corrects the structure without also re-parameterizing the haircut. The waterfall now closes exactly.
One quotation is also withdrawn from section 2. The phrase “signs of stress are already visible” was published inside a list of the BIS report’s findings on the AI complex. In the report it introduces redemption pressure at retail-facing direct lending funds, which is private credit broadly rather than AI credit specifically. Quoting it in an AI list borrowed the regulator’s alarm for a subject it was not describing, so it is removed rather than relocated.
What moved: equity mean loss 68% to 78%, equity wiped in 52% of paths to 61%, mezzanine 28% to 29%, senior mean loss 0.8% to 1.2%, and P(senior intact) 92.9% to 90.1%. Every conclusion holds in direction and the senior tranche is modestly worse than published. The causal sentence in section 6 now names the actual mechanism, which is the circular-financing amplifier lifting pool loss through the 40% attachment point.
References
- Bank for International Settlements (2026). “I. Progress and peril,” Annual Economic Report 2026, 28 June 2026. Bank for International Settlements, Basel.