Abbreviations spelled out (first use):

• GPU: Graphics Processing Unit (the AI accelerator chip; the key collateral here).
• P/E: Price-to-Earnings ratio; P/S: Price-to-Sales ratio; EPS: Earnings Per Share.
• GAAP: Generally Accepted Accounting Principles (US accounting standard).
• CDO: Collateralized Debt Obligation; SIV: Structured Investment Vehicle; ABS: Asset-Backed Security; CLO: Collateralized Loan Obligation. (2008-era structured-credit instruments the author maps onto today.)
• DSCR: Debt-Service Coverage Ratio (cash flow ÷ debt payments); NOI: Net Operating Income.
• DDTL: Delayed-Draw Term Loan; SOFR: Secured Overnight Financing Rate (the US benchmark lending rate); bps: basis points (1 bp = 0.01%).
• RPO: Remaining Performance Obligations (contracted future revenue / "backlog").
• ROI: Return on Investment; YoY: Year-over-Year; TPU: Tensor Processing Unit (Google's custom AI chip); ARM: Adjustable-Rate Mortgage; NINJA: "No Income, No Job or Assets" (2006 subprime loan).
• Neocloud: a new specialist cloud provider that buys GPUs and rents compute (e.g. CoreWeave); funds itself with GPU-backed debt.

---

The master framework: numerator vs. denominator bubbles#

The whole essay rests on distinguishing where the excess sits in a P/E multiple (Price ÷ Earnings).

[!tip] Intuition, which part of P/E is lying? A P/E has two parts. A numerator bubble = the Price (P) is too high on honest earnings → it deflates (price falls toward fundamentals; equity holders lose). A denominator bubble = the Earnings (E) are too high (flattered/temporary) so the stock looks cheap → when E reverses, debt taken against those earnings does not reverse with it → it detonates. "A numerator bubble deflates. It does not detonate… [a denominator bubble] does not deflate. It cascades."

| | 2000, Numerator bubble | 2008, Denominator bubble | 2026, author's claim | |, -|, -|, -|, -| | Where the excess lives | In stock prices | In inflated earnings (cheap-looking multiples) | In earnings (cheap-looking multiples) | | Example valuations | No-revenue firms at 20-30× P/S; Cisco ~150× P/E, Microsoft ~70× P/E | Banks at single-digit P/E on peak-cycle earnings | Nvidia low-30s forward P/E; hyperscalers low-to-high 20s | | Financed by | Equity (IPOs, venture capital) | Debt vs. deteriorating collateral | Debt vs. fast-depreciating GPUs | | Failure mode | Multiple compression | "When the E reversed, the debt did not reverse with it" → solvency cascade, forced selling, contagion | Solvency cascade (claim) | | Damage | Nasdaq −78%, but banking "barely scratched" | Systemic crisis | Systemic (claim) |

Why it's dangerous now: "We are trained by 2000 to look for a bubble in the price, and there isn't one in the price, so the screens stay calm and the value investors, reading low multiples, step in to buy what looks cheap." That marginal "value" bid is what lets insiders and early lenders exit.

---

The Overearning Engine, three stacked mechanisms inflating reported profit#

The author argues reported AI-complex earnings are overstated by three layered accruals (an accrual = profit booked now that isn't matched by cash now).

1. Under-depreciation (asset-side accrual)#

• The practice: 2020-2024, hyperscalers extended assumed server "useful life" from 3-4 years to 5-6 years. Longer life → smaller annual depreciation charge → higher reported profit. (Depreciation = spreading an asset's cost over its useful life; stretch the life and each year's expense shrinks.)
• Evidence: ~$18 billion/year of depreciation "saved" industry-wide; Amazon shortened GPU life on part of its fleet and took a ~$700 million charge; Meta pushed toward 5.5 years; Microsoft's Satya Nadella reportedly resisted being "stuck with four or five years of depreciation on one generation."
• The reality: H100 GPU rental fell from ~$8/hour (2023) to ~$2.35/hour (2026): a 70%+ decline. True value erosion ≈ 40%/year (stress) vs. ~17%/year (book). The gap is "pure overstatement of profit, deferred until reality forces the reckoning."
• Michael Burry's math: shortening assumed GPU lives toward the real 2-3-year replacement cycle would overstate profit by ~$176 billion across 2026-2028, leaving Oracle and Meta operating income "more than 20% above economic reality."
• 2008 parallel: banks holding too-thin loan-loss provisions in the boom, forcing huge reversals later.

[!tip] Intuition, the laptop financed like a building A GPU loses value like a laptop (fast), but its owner books depreciation like it's a building (slow). Every year that gap is booked as "profit" that isn't economically real.

2. Equity-method / fair-value mark-ups on AI-lab stakes (non-cash accrual)#

• The practice: hyperscalers own stakes in AI labs (Microsoft→OpenAI; Alphabet & Amazon→Anthropic). When a lab raises a new round at a higher valuation, the holder books a non-cash gain straight through GAAP earnings.
• Scale: "Tens of billions of dollars of reported 'profit'… is not cash from selling a product, it is the mark-to-market of a private holding that has only ever marked up."
• The problem (symmetry): marks that rise on the way up fall on the way down: exactly when every other indicator dims at the same time.
• 2008 parallel: "mark-to-model" accounting and booking a CDO's projected lifetime profit at origination.

3. Circular / round-trip revenue (demand-side accrual)#

• The mechanism: capital flows out of one node as investment and comes back as another node's revenue, and the market capitalises it as if it were external ("exogenous") demand.
• Examples:
• Nvidia sells chips to CoreWeave and invests $2 billion of equity into it.
• Microsoft invests in OpenAI; OpenAI commits hundreds of billions of spend back to Microsoft cloud. Microsoft's $627 billion of RPO backlog is 45% ($281 billion) a single counterparty, OpenAI, which Microsoft owns 27% of.
• Google funds Anthropic; Anthropic commits ~$50 billion to a neocloud; Google guarantees that neocloud's leases.
• Oracle: $300 billion OpenAI commitment; Amazon: $38 billion.
• The issue: "It is real, it is contracted, it is legal. It is not exogenous. It is the capital cycle feeding itself." The backlog the complex cites as proof of real demand is "to a remarkable degree, one bet, on AI's eventual return, wearing a great many costumes."
• 2008 parallel: originate-to-distribute fee income (origination, securitization, structuring fees) all booked on volume funded by the next node in the chain.

Stacked overearning, the illustrative model#

• Normalized durable earning power indexed at 100; the three accrual layers add +55. The stock looks like 28× reported earnings but is actually ~43× durable earning power (optimistic case: accruals merely stop). In a downturn the accruals reverse violently and the P/E "becomes undefined (like Citigroup)."

---

Originate-to-Distribute, mapping 2008 securitization onto 2026 GPU finance#

The 2008 machine passed each loan down a chain, every link booking a fee and handing the risk on; it held only while the end-borrower kept paying. The author maps it link-for-link:

| Stage | 2008 Mortgage | 2026 GPU | |, -|, -|, -| | Origination | Countrywide / New Century writes loan, takes fee, sells on | Nvidia sells GPU at ~75% gross margin, finances the buyer, takes equity in neoclouds | | Warehouse | Wall Street warehouse lines pool loans | Neoclouds raise GPU-backed debt against customer-lease cash flows | | Repackaging | CDO / CDO-squared, mark-to-model gains | Hyperscaler AI-lab markups; backlogs sold as "forward demand" | | Wrap | Monoline insurers stamp AAA on senior tranches | Google lends its investment-grade credit to structures that can't stand alone | | Rating | Agencies rate subprime AAA | Investment-grade ratings on "long-lived" assets that are demonstrably short-lived | | Distribution | Paper to pensions, SIVs, banks | GPU-backed bonds held by retiree annuities trusting AI revenue |

"Every link… books a fee and hands the risk on. And both chains hold only as long as the demand at the very end is real."

Case study, the Google guarantee (wrong-way risk)#

A concrete illustration of demand and credit originating at the same balance sheet:

• Google owns ~14% of Anthropic and supplies it TPUs. Anthropic announces a ~$50bn partnership with neocloud Fluidstack (Texas/New York data centres) → this anchors "demand" downstream.
• Fluidstack leases capacity from bitcoin miners repurposing sites: Hut 8 (245 MW, 15-yr River Bend LA lease, $7bn base→$17.7bn with extensions), TeraWulf (200+ MW, 10-yr Lake Mariner NY), Cipher Mining (300 MW Barber Lake). ~$19bn+ contracted.
• The wrap: Fluidstack and the miners lack investment-grade balance sheets, so Google guarantees the leases (~$1.8bn TeraWulf backstop; ~$1.3bn Abernathy JV; ~$1.4bn + $333m Cipher; full backstop on Hut 8's $7bn lease), and even owns equity warrants in the landlords (~8% TeraWulf, ~5.4% Cipher).
• Google sits at four nodes at once: (1) supplies the chips, (2) funds the tenant that creates demand, (3) guarantees the leases that make it financeable, (4) owns equity in the landlords. Banks (J.P. Morgan / Goldman Sachs) lend up to 85% of project cost "comfortable because Google made them so."

[!warning] Wrong-way risk (the key concept) Wrong-way risk = a guarantee that is most likely to be called exactly when the guarantor can least afford it. If Anthropic's economics disappoint enough to sink Fluidstack's lease payments, the same shock sinks Anthropic's compute spend and Google's AI returns simultaneously: "The guarantee comes due in the one world where Google least wants the bill. Correlation, when it matters, goes to one."

2008 parallel: SIVs were off-balance-sheet conduits the sponsor backed with liquidity puts, "the risk looked gone. It wasn't." Citigroup alone took $25bn+ of "sold" vehicles back onto its balance sheet.

---

GPU finance mechanics, a 3-year asset on a 6-year loan#

The single clearest mispricing, in the author's view.

• Loan structure: lender advances ~65% of GPU cost; borrower funds ~35% equity; the loan amortizes over ~5.5-6 years; secured by the chips + customer contract + data-centre lease; covenant = a minimum DSCR (debt-service coverage ratio).
• Base case (stable): GPU value erodes ~17%/yr → the asset value stays above the amortizing loan balance for the whole term → coverage holds.
• Stress case (glut): a new generation (Blackwell) makes old Hopper chips obsolete → value falls ~40%/yr → the asset value plunges below the loan balance by ~year 3 → the loan is underwater with years left to run.
• The error: "A GPU is, economically… closer to a laptop than to a building. It is being financed like a building." The interest-only GPU loan is "the interest-only mortgage of this cycle: it works only if the asset behaves like real estate, and the asset behaves like a laptop."
• 2006 parallel: 2/28 ARMs, option-ARMs, interest-only mortgages, all "sized to an asset assumed durable or appreciating"; when house prices fell 30% the collateral couldn't cover the loan.

CoreWeave, the public laboratory for GPU-backed credit#

The one listed pure-play, so its filings reveal the model. Q1 2026:

• Revenue $2.078bn (+112% YoY) but net loss $740m; interest expense $536m.
• Total debt $24.859bn; debt-to-equity 8.9×; current ratio 0.46 (near-term bills exceed liquid assets).
• Backlog $99.4bn but only 36% converts within 24 months (75% within 4 years).
• Nvidia bought $2bn of CoreWeave equity in the quarter, "the vendor financing the buyer's equity, the circular loop in miniature."
• Depreciation trap: much infrastructure is still construction-in-progress, "not yet being depreciated", the loss "is set to widen mechanically as the fleet comes online, before any demand shock at all."

The marginal loan tells the cycle stage (DDTL 4.0 → 5.0)#

The newest, "marginal" deal reveals where the cycle is. Two CoreWeave delayed-draw term loans seven weeks apart:

| | DDTL 4.0 (31 Mar 2026) | DDTL 5.0 (18 May 2026) | |, -|, -|, -| | Size | $8.5 bn | $3.1 bn | | Rating | A3 / A− (investment-grade) | Ba2 / BB+ (junk) | | Spread | SOFR +225 bps | SOFR +450 bps (double) | | Collateral | GPUs + ~$19bn Meta contract | "two large non-investment-grade customer contracts" | | Coverage covenant | 1.15× | tighter 1.35× |

"Issuance accelerating while the marginal deal gets worse: is the single most reliable signature of a late-stage credit cycle." 2006 parallel: prime → Alt-A → subprime → NINJA as good borrowers ran out. "The best collateral has been financed. What is left is the Alt-A."

---

The ROI Wall, why the math must eventually break#

Return on AI build-out = (end-customer profitable revenue it can support) ÷ (cost to build it).

• Numerator is capped by economics, what buyers can actually earn from AI and afford to pay.
• Denominator is rising fast due to physical bottlenecks: power, transformers, grid, skilled labour, land. Build cost went from ~$42bn per gigawatt (2023-24) to ~$62bn per gigawatt (2026) and "the bottlenecks only push it higher."
• Hold the numerator flat while the denominator rises and the return falls through the ~10% cost-of-capital line. Past that, "the build does not earn a low return; it destroys value outright." (Context: Microsoft alone plans $190bn of capex this year.)
• "Each incremental gigawatt costs more than the last and earns no more, because the end-demand that pays for it is bounded."
• 2007 parallel: the housing machine ran out of people who could afford a house at an economic rate, so it lent to people who could only afford the teaser rate. "The marginal AI dollar is in the same position… real until the capex normalizes, which it must, because spending that cannot earn its cost of capital is, eventually, spending that gets cut."

---

The unwind sequence (the cascade)#

Trigger: end-AI returns disappoint vs. the capex assumption; the marginal circular dollar stops circulating.

1. Deceleration & mark reversals: hyperscaler growth slows; Alphabet/Amazon mark down AI-lab stakes; tens of billions of non-cash "earnings" reverse; the capital-budget justification is undercut.
2. Capex rationalization: hyperscalers trim capex guidance; chip order book softens; Nvidia growth decelerates; days-sales-outstanding lengthen; supply commitments become inventory write-down risk.
3. Collateral impairment: GPU rental rates fall (the "load-bearing variable" for every neocloud loan).
4. Covenant cascade (the heart of it): falling rents push neocloud cash flow below debt service → coverage covenant breaks → forced prepayment/default → GPUs dumped into a thin secondary market → each forced sale marks down resale values and rents for everyone → cuts collateral/coverage on every other loan → trips the next covenant → forces the next sale. "It is 2008 exactly: the margin call that forces the sale that lowers the price that triggers the next margin call."
5. Backlog evaporation: labs can't raise the next round or honour cloud commitments; Microsoft's $281bn and Oracle's $300bn backlogs "go to fiction." But "the debt stays real."

[!tip] The asymmetry that defines the whole thesis "Equity and backlog evaporate, debt does not: is the whole difference between a deflation and a cascade."

Recognition timing: "The bell… does not ring at the top. It rings on the way down, in the form of a covenant breach," because the financing was committed years earlier against collateral that was always going to depreciate.

---

Where the systemic risk sits, the new "AIG seat"#

• The danger is not the brokers writing the loans but the regulated, levered "safe" holders of the senior, investment-grade paper: the life-insurance & annuity complex funding it with sticky retirement money.
• Example: Athene anchored Meta's $29bn data-centre financing; Apollo syndicated the rest to insurers/pensions. The data-centre asset-backed-bond market is up 8× in five years.
• "That is the AIG seat of this cycle: the regulated, levered, 'safe' holder of senior exposure to collateral that depreciates faster than the loan against it."
• The tell: Apollo, which pioneered the model, is running off its CLO book and sitting on a $40bn cash pile, its CEO Mark Rowan saying it is "preparing its balance sheet for a period of systemic instability." "When the originator is hoarding dry powder against the assets it just sold you, that is the tell."

---

The bull case (what would make this thesis WRONG)#

The author fairly lists four invalidating scenarios, important for balance:

1. End-AI revenue inflection: enterprises/consumers actually pay for AI output at scale and margin that clears the ROI wall; the circular dollar becomes genuinely external; backlogs convert to cash. (The author says this is the most important one: "if that proposition is wrong, the thesis is wrong.")
2. GPU value retention: inference demand grows fast enough to keep rental rates firm; the ~6-year life assumption is roughly right; collateral holds.
3. Better debt structure: more termed-out, better-covenanted, more equity underneath, less forced-seller reflexivity than 2008 → orderly losses, not a cascade.
4. Survivable guarantees: hyperscaler balance sheets are so large they absorb the lease guarantees and write-downs without contagion.

[!warning] But even the bull case runs into atoms "Every one of those bullish outs runs through physical reality: power, land, transformers, the cost to build a gigawatt. The ROI wall is a denominator problem, and the denominator is made of atoms whose costs are rising faster than the returns meant to justify them."

---

Historical anchor, Citigroup 2007 ("the most expensive cheap stock")#

Citigroup traded at 9× earnings on 30% EPS growth weeks before those very earnings proved to be peak, then fell >75%. "The multiple was never the margin of safety… the most expensive cheap stock in the market." (Other tells the essay cites: homebuilders ~6× P/E in 2007, then −80%; banks levered ~30:1 assets-to-equity.)

---

Coined terms / frameworks to remember#

• Peak Cheap: a low multiple on artificially inflated peak earnings; looks cheap, is the highest-risk position.
• Numerator vs. Denominator bubble: price bubble (deflates) vs. earnings bubble (cascades).
• The Overearning Engine: under-depreciation + non-cash markups + circular revenue.
• Originate-to-Distribute (GPU edition): the 2008 securitization chain re-mapped.
• Wrong-way risk: the guarantee called exactly when the guarantor is weakest.
• Forced-seller cascade: reflexive covenant-break → fire-sale → lower marks → next break.
• The ROI Wall: bounded AI demand meeting ever-rising per-gigawatt build cost.

---

Common pitfalls / critical reading#

• It's a bear thesis, and partly anonymous ("Groundbreaker"). Treat it as a risk argument to test, not a forecast. The author concedes the bull case (esp. #1) could invalidate everything.
• Timing is unknowable. "The bell rings on the way down", even if right, the essay offers no date; bubbles can run far longer than the leverage math suggests.
• Not all names are equally exposed. Cash-rich hyperscalers with real cash earnings differ hugely from levered neoclouds (CoreWeave) and circular-demand counterparties.
• Figures are the author's: many are estimates/illustrative (the indexed 100+55 model, the $176bn Burry figure). Directional, re-verify before use.

---

Exam / desk focus, why this matters for PCA SOF#

[!warning] This is a portfolio-level risk note, not a single stock The Strategic Opportunities Fund is long the entire AI build-out (The AI Value Chain), the precise structure this essay says is a 2008-style credit bubble. Read it against the book:

• Compute / collateral at the centre of the thesis: NVIDIA (vendor-financing CoreWeave; circular revenue), Micron Technology (memory tied to the same capex), TSMC, Marvell Technology, most exposed to a capex normalization + GPU-glut.
• Hyperscalers = the balance sheets the whole structure leans on: Microsoft ($627bn RPO, 45% OpenAI), Alphabet (Anthropic stake + lease guarantees = wrong-way risk), Amazon, Meta Platforms ($29bn Athene-anchored financing). Watch for AI-lab stake mark-downs and depreciation-life disclosures.
• Power names Constellation Energy / Vistra, demand is downstream of the same data-centre capex that the ROI wall threatens.
• The screening trap is the key takeaway: these names may look reasonably valued on P/E, the essay's whole point is that a cheap multiple on inflated earnings is the danger, not the safety. Pair with the moat/valuation work in Risk Master Note and PCA SOF Investment Thesis.

The signpost to monitor (per the essay): GPU rental rates, the quality of the marginal GPU-backed loan (rating/spread/covenant drift, à la DDTL 4.0→5.0), hyperscaler capex guidance, and AI-lab funding rounds. The cascade "rings on the way down."

Linked Notes#

• Related Readings: Return on Tokens (ROT) (the value/efficiency counterpart, ROT asks if AI spend earns; Peak Cheap asks if the financing of that spend is solvent).
• Related Themes: The AI Value Chain · AI Power Demand.
• Related Risk / Thesis: Risk Master Note · PCA SOF Investment Thesis.
• Related Holdings: NVIDIA · Microsoft · Alphabet · Meta Platforms · Micron Technology · Constellation Energy · Vistra.