The AI-Economy
Fragility Scorecard
Not a price target. Not a date prediction. A bottom-up instrument that measures how far the AI sector’s narrative has drifted from its fundamentals — and flags the moment independent signals converge into the same warning.
Michael Burry didn’t predict the 2008 crash by reading sentiment. He read the loan tapes — the actual collateral beneath the AAA story — and found the underlying assets didn’t support the narrative being told about them. This is the same method, applied to the AI economy: go beneath the narrative, measure the fundamentals, and watch for divergence. When enough independent indicators converge into the red simultaneously, that convergence is the signal. Not any one metric. The convergence.
The donkey doesn’t time bubbles. The donkey reads the tape beneath the tape.
Nvidia (NVDA)
Data-center GPUs · Fabless semiconductor · Santa Clara, CA
AI infrastructure buildout narrative exceptionally strong · NVDA fwd P/E ~21x is NOT extreme vs. semiconductor sector (Damodaran Jan 2026: 37.3x fwd) · divergence rests on structural fundamentals, not a stretched multiple
Note on P/E: NVDA fwd P/E ~21x (trailing ~31x), as of Jun 2026 — NOT an extreme multiple. Semiconductor sector fwd P/E 37.3x / current 70.1x (Damodaran, Jan 2026). The divergence case here rests on structural fundamentals (Indicators 2, 4, 6), not a stretched multiple. ⚠ NOT SOURCED: 10–15yr SOX-index average P/E — labeled gap; do not cite until filled from primary data.
Composite of six indicators below · higher fragility = lower fundamentals score
Convergence Flag: Active
3–4 indicators elevated — Ind. 2, 4, 6 RED · Ind. 1 AMBER‑RED · Ind. 3 & 5 deliberately NOT elevated.
Three indicators are independently red and one is amber-red, derived from different filings, different economic variables, and different analytical methods. Each can be explained away in isolation. Their simultaneous convergence cannot. This is the Burry method. Note on Indicator 3: Jensen Huang’s $1B+ in share sales is less than 1% of his holdings on a pre-set 10b5-1 plan — we score it deliberately LOW (~30). Not cherry-picking every metric as red. That asymmetry is the credibility. Indicator 5 (Energy) is AMBER — genuinely contested among researchers; the hard figures are not yet sourced.
01 —
Depreciation Integrity
60–70 / 100
provisional
Depreciation Integrity
Nvidia is fabless — it doesn’t depreciate manufacturing equipment. The risk is the ecosystem effect: the earnings that underpin the AI-demand thesis are largely reported by four hyperscalers whose depreciation policy choices each add billions to reported net income. The critical finding from the filed 10-Ks is that three of the four extended useful lives (earnings-flattering), while one — Amazon — has already started to shorten. Amazon is the canary: it shortened a subset of servers/networking from 6 to 5 years effective January 1, 2025, citing “AI/ML obsolescence” verbatim — the first major hyperscaler to acknowledge the obsolescence problem in a filing. The other three ran the opposite direction: Meta extended to 5.5 years, Google from 4 to 6 years, Microsoft from 4 to 6 years. Michael Burry’s estimated ~$176B aggregate depreciation understatement (2026–2028) gives the order-of-magnitude — that figure is his attributed estimate, not an audited number, but the direction is confirmed by the filed disclosures.
Nvidia refreshes its GPU architecture roughly every 12–18 months. Meta is claiming a 5.5-year useful life on hardware in an industry where the CEO of the leading chip supplier announces the next architecture before the current one ships. The donkey just reads the footnotes and counts the years.
Amazon (canary — shortened): Shortened a subset of servers/networking 6→5 yrs, effective Jan 1 2025, citing AI/ML obsolescence verbatim; ~$920M accelerated depreciation Q4 2024 + ~$0.6B additional 2025 operating-income hit. (Amazon FY2024 10-K, filed 2025-02-07 — PRIMARY.)
Meta (extended): Stretched to 5.5 yrs (Jan 2025); favorable earnings impact: ~−$2.9B FY2025 depreciation reduction. (Meta FY2024 10-K.)
Google / Alphabet (extended): Servers 4→6 yrs (FY2023); FY2023 depreciation −$3.9B vs. shorter-life baseline; net income benefit +$3.0B. (Alphabet 10-K.)
Microsoft (extended): 4→6 yrs (FY2023); ~$3.7B favorable; still 6 yrs in FY2025 10-K. (Microsoft 10-K, med-high confidence.)
Burry (attributed estimate — not audited): ~$176B depreciation understatement 2026–2028; by 2028 Oracle earnings ~26.9% overstated, Meta ~20.8%. (Michael Burry, 2025-11-11; via Yahoo Finance / Fortune. Attributed, not independently audited.)
02 —
Capex vs. Demand Gap
75–85 / 100
provisional
Capex vs. Demand Gap
Nvidia’s FY2025 numbers are extraordinary on their face: total revenue $130.5B (+114% year-over-year), Data Center segment $115.2B (+142%, ~88% of total), GAAP net income $72.88B. This is real revenue from real customers writing real checks. The fragility question is not whether Nvidia is selling chips — it clearly is — but whether the hyperscalers buying those chips have end-user revenue that will justify the spend at this scale. The hyperscalers collectively guided toward combined AI capex approaching $300–400B+ for 2025–2026. That is not capex to revenue — that is capex to build capacity in anticipation of demand. If the demand materializes (see Indicator 6), the capex is rational. If the pilot-to-production gap remains wide, the capex cycle creates overcapacity that eventually slows Nvidia’s order book.
$115B in data-center chip revenue in one year. The donkey is genuinely impressed. The donkey is also reading Indicator 6, which shows that ~95% of enterprise GenAI pilots have not produced measurable P&L impact. Those two facts are both true, and their reconciliation is the scorecard’s question.
Nvidia FY2025 (ended 2025-01-26): Total revenue $130.5B (+114% YoY). Data Center segment $115.2B (+142%, ~88% of total revenue). GAAP net income $72.88B. (NVIDIA official earnings release 2025-02-26 + 10-K — PRIMARY.)
Demand-side context: See Indicator 6 (MIT NANDA Aug 2025: ~95% of enterprise GenAI pilots = no measurable P&L impact) for the end-user ROI data that tests the capex thesis.
NOT SOURCED — LABELED GAP: NVDA FY2026 quarterly revenue (8-Ks linked; pull from SEC EDGAR before next publish). FY2025 figures above are primary.
03 —
Insider‑Selling Intensity
25–35 / 100
confirmed
Insider‑Selling Intensity
Jensen Huang sold approximately $1B+ in Nvidia shares in the twelve months ending late October 2025 — all via a 10b5-1 plan adopted March 20, 2025 (6.0M shares authorized, completed late October 2025). The headline is large. The context is more important: this represents less than 1% of Huang’s ~859M-share stake (~4% of Nvidia’s total shares outstanding). The plan was pre-set with no evidence of abnormal acceleration triggered by current information. On a 10b5-1 at <1% of holdings, the correct score for this indicator is low. We score it deliberately green-amber (~25–35). Not every indicator should be red, and not every large dollar figure is a fragility signal. Retaining 99%+ of a position into an extraordinary price run is itself a form of insider confidence.
Huang sold more than a billion dollars of Nvidia stock and still owns roughly $100B+ worth. The donkey finds the math, not the headline, interesting — and the honest read is that this one doesn’t move the fragility needle. The credibility of the scorecard depends on not calling everything red.
Jensen Huang (CEO): 10b5-1 plan adopted 2025-03-20; up to 6.0M shares authorized; completed late October 2025; total proceeds ~$1B+. Huang’s stake ~859M shares (~4% of NVDA outstanding). Sale = <1% of holdings. All via pre-set plan. (SEC Form-4 filings; Bloomberg 2025-10-31; CNBC 2025-06-24.)
Assessment: WEAK signal. 10b5-1 plan removes information-asymmetry inference. Retained holdings overwhelmingly exceed shares sold. This is an honest low score — the scorecard’s credibility requires it.
04 —
Financing Opacity & Circular Leverage
78–88 / 100
provisional
Financing Opacity & Circular Leverage
CoreWeave’s S-1 (filed 2025-03-03) is the primary source for this indicator and makes the circular structure fully explicit. The chain: Nvidia is CoreWeave’s chip supplier; Nvidia held a >5% beneficial owner position at IPO per the S-1, later estimated at ~47.2M shares (~$3.66B, ~11%) per Q1-2026 13F data (third-party summary, medium confidence). CoreWeave financed its GPU buildout via a $7.6B GPU-collateralized debt facility (Blackstone/Magnetar; total debt $8.0B as of Dec 31 2024; >$14.5B raised across 12 financings). The GPU inventory securing that debt was purchased from Nvidia. Microsoft accounted for 62% of CoreWeave’s 2024 revenue; top two customers ~77%. Nvidia itself paid CoreWeave ~$320M through 2024 and is backstop-obligated to buy unsold capacity through 2032 (initial commitment ~$6.3B).
The full loop: Nvidia sells GPUs to CoreWeave → CoreWeave collateralizes those GPUs for debt → uses debt to buy more Nvidia GPUs → Nvidia books revenue → Nvidia holds equity in CoreWeave whose value depends on CoreWeave’s solvency → which depends on Microsoft contract renewal → which depends on enterprise AI demand. Every link is disclosed. The fragility is not the structure per se — it is the propagation speed if one link breaks. This is the 2008 CDO pattern applied precisely: disclosed, legal, and highly correlated when stressed.
Nvidia is the supplier, the investor, and the customer of the company that financed its purchases with debt secured by the inventory Nvidia sold it, with a government-style take-or-pay backstop running to 2032. The donkey just reads the S-1 and counts the links.
Nvidia stake: S-1 states Nvidia = “>5% beneficial owner” at IPO. Later ~47.2M shares (~$3.66B, ~11%) per Q1-2026 13F summary (third-party — med conf.; reject any “~1%” figure). (CoreWeave S-1 filed 2025-03-03 — PRIMARY.)
CoreWeave debt: $8.0B total debt (Dec 31 2024); $7.6B via GPU-collateralized term loan facility (Blackstone/Magnetar; DDTL 1.0 up to $2.3B + 2.0); >$14.5B raised across 12 financings total. (CoreWeave S-1.)
Revenue concentration: Microsoft = 62% of CoreWeave 2024 revenue; top two customers ~77%. (CoreWeave S-1.)
Circular chain: Nvidia paid CoreWeave ~$320M through 2024 (Nvidia = customer of its own investee). Nvidia backstop: obligated to purchase unsold CoreWeave capacity through 2032 (~$6.3B initial commitment). (CoreWeave S-1.)
05 —
Energy & Diminishing Returns
45–55 / 100
provisional
Energy & Diminishing Returns
The scaling-laws question is genuinely contested among serious researchers, which is why this indicator is amber rather than red. The empirical question is what the benchmark gain per dollar of training compute has been across successive model generations. The general finding from the published literature and independent tracking is that each major generation has required substantially more compute for progressively narrower benchmark improvements — more spend per capability point. This matters for Nvidia specifically because the entire AI infrastructure valuation rests on the assumption that more compute equals more capability equals more revenue. If the marginal return on compute is compressing, the capex arithmetic breaks. Energy is the physical constraint that makes this observable: training runs that cost 10× more to power than their predecessors for 1.2× the capability gain are the empirical signal.
The donkey is not calling the end of scaling laws. The donkey is noting that “we will spend more compute and get more capability” is a bet that has paid off historically and is now being made at a scale where the bet itself is the market.
This indicator is kept AMBER because the scaling-laws question is genuinely contested among researchers — the honest score is not red. The hard per-watt and cost-per-capability-point figures required to score it precisely are NOT SOURCED and are labeled as a tracked gap.
NOT SOURCED — LABELED GAP: MLPerf benchmark results across GPU generations (A100 / H100 / H200 / B200) — performance per watt and performance per dollar. Epoch AI training-compute tracking for GPT-3 / GPT-4 / GPT-5 vs. benchmark capability progress. GB200 NVL72 rack power consumption vs. H100 equivalent. Pull from MLPerf.org, Epoch AI published datasets, and Nvidia datacenter spec sheets before next publish.
06 —
Organic End‑User Demand
60–72 / 100
sourced
Organic End‑User Demand
This indicator moved from amber to red-amber on the strength of a single, large-sample finding: MIT Project NANDA’s August 2025 “GenAI Divide” report found that approximately 95% of enterprise GenAI pilots showed zero measurable P&L impact; only ~5% produced any measurable ROI. This is not a technical failure rate — it is a financial productivity finding. Gartner corroborates the direction: at least 30% of enterprise GenAI projects were abandoned post proof-of-concept by end-2025; Gartner further projects that over 40% of agentic-AI initiatives will be canceled by 2027. Nvidia’s FY2025 Data Center revenue of $115.2B represents real chips sold to real customers. The structural question this indicator poses is whether the enterprise demand that ultimately justifies hyperscaler capacity purchases — and therefore Nvidia’s continued order book at this scale — can clear the pilot-to-production gap before the next hardware cycle requires the capex to begin again.
“Enterprises are buying AI” is true. “95% of enterprise GenAI pilots have shown zero measurable P&L impact” is also true. Both sentences are in the public record. The donkey just puts them in the same paragraph.
MIT Project NANDA (PRIMARY): “GenAI Divide” report (Aug 2025): ~95% of enterprise GenAI pilots showed zero measurable P&L impact; only ~5% measurable ROI. Precisely: “zero measurable P&L,” not “technically failed.” (MIT NANDA via Fortune, 2025-08-18.)
Gartner: ≥30% of enterprise GenAI projects abandoned post-PoC by end-2025. Forecast: >40% of agentic-AI initiatives canceled by 2027. (Gartner, published 2025.)
NOT SOURCED — LABELED GAP: NVDA FY2026 quarterly AI-segment revenue detail; hyperscaler AI contract renewal rates; enterprise AI ARPU / churn data. (Note: the previously cited “Gartner Jun 2026” figure was not verified and has been removed; replaced by the confirmed sources above.)
Six independent probes, one convergence verdict
Each indicator is scored 0–100 fragility per company, sourced to a specific filing, date, and figure. No single indicator is the tell — they’re designed to be independent so that when three or more turn red simultaneously, the simultaneous convergence cannot be explained away as noise.
Useful-life stretch
Are AI chips depreciated over 5–6 years when realistic obsolescence runs 2–3?
The gap between claimed and realistic useful life is a direct earnings-inflation estimate. Disclosed in PP&E footnotes almost nobody reads.
Build vs. buy
How far does AI infrastructure capex outrun the revenue actually attributable to AI end-user demand?
Capex can be rational before revenue appears — if the demand curve is credible. The fragility score rises when the gap is wide and demand proxies are thin.
Insiders vote with shares
Are executives net-selling at elevated price levels, sustained and at scale?
Insiders have information asymmetry. Sustained selling into a multi-year run by senior leadership is the market’s best read on what the people closest to the business think the asset is worth.
Circular leverage
Is demand being financed in ways that make it self-referential? Vendor financing, GPU-collateralized debt, equity in your own customers?
Circular structures inflate apparent demand without creating end-user revenue. When the chain breaks, the demand signal reverses sharply. The 2008 CDO analog is exact.
Returns per compute dollar
Is the cost per marginal unit of AI capability gain increasing? Is the scaling-law curve flattening at current spend levels?
The entire infrastructure valuation rests on more compute equaling more capability equaling more revenue. If the marginal return is compressing, the capex arithmetic breaks.
Real willingness‑to‑pay
Paid vs. free usage, enterprise ROI, retention, churn. Is there durable demand, or is it FOMO buying that churns when budget scrutiny arrives?
Free-tier MAUs are an option on demand, not demand. Enterprise pilots that never reach production are not revenue. The structural demand question is the last and hardest to answer.
Convergence, not coincidence
Any single indicator can be explained away. The convergence rule is the instrument’s core: when three or more independent indicators turn red simultaneously, the probability that each is independently explained by benign factors drops dramatically. Three independent data sources — different filings, different methods, different economic variables — converging on the same warning is the regime-change signal. This is the Burry method. He didn’t find one problem with the mortgage market. He found the same structural problem reflected simultaneously in the loan-level data, the CDO structure, the rating-agency model inputs, and the banks’ own hedging behavior.
| Indicators Red | Interpretation |
|---|---|
| 0–1 |
Baseline Elevated valuations in a growth sector. Not unusual. Watch and update. |
| 2 |
Watch Independent warnings emerging. Review each source chain. Determine if they share an underlying cause or are truly independent. |
| 3 |
Convergence Flag Independent methods agree. Regime fragile. This is the primary signal. Not a trade trigger — a structural reading. |
| 4+ |
High‑Alert Structural fragility, not noise. Divergence from fundamentals is now the dominant story. The question is not whether but when and what catalyzes the correction. |
The donkey doesn’t call the top. The donkey counts the red flags until the pattern isn’t deniable anymore.
Narrative score minus fundamentals score
Narrative strength captures how bullish the market story is: current P/E vs. sector historical P/E, analyst consensus sentiment, forward-guidance language from management calls. A high narrative score means the story being told about the company is very strong.
Fundamentals score is the inverse of the composite fragility score. Companies with low fragility (strong fundamentals) score high; high fragility pulls the fundamentals number down. It represents how well the underlying economic substance supports the narrative.
The divergence is the gap between the two. A large positive divergence — strong narrative, weak fundamentals — is the fragility condition. The regime-change alert fires when: (1) the divergence crosses a critical threshold, historically when sector P/E exceeds 3× the 10-year average without proportional earnings growth; AND (2) three or more fragility indicators are simultaneously red (the convergence flag is active). Both conditions must be met.
A large divergence is not a shorting signal. It is a fragility signal. The timing of mean reversion is unknowable; the direction and eventual magnitude of correction, in a high-divergence regime, is what history suggests is predictable. This is the instrument’s purpose: not to call the date, but to measure the gap while it is open.
Every number traces to a filing
The scorecard is built on provenance discipline. The same standard that makes a legal evidence ledger credible — verbatim source, document reference, date, page or section — applies here. Every fragility score is backed by a chain: filing → date → figure → calculation → score. Estimates are labeled as estimates. Provisional scores are labeled provisional until the underlying filing is confirmed.
| Indicator | Primary Source |
|---|---|
| 01 — Depreciation Integrity | SEC EDGAR: 10-K PP&E footnotes, useful-life disclosures by asset class |
| 02 — Capex vs. Demand Gap | SEC EDGAR: cash-flow statements; earnings-call AI-segment revenue disclosures |
| 03 — Insider‑Selling Intensity | SEC EDGAR: Form-4 filings (ADC Insider tab already ingests this) |
| 04 — Financing Opacity | SEC EDGAR: 10-K related-party notes, debt footnotes; S-1 / prospectus filings |
| 05 — Energy & Diminishing Returns | Published training-compute estimates; MLPerf benchmarks; hardware spec sheets |
| 06 — Organic End‑User Demand | Earnings-call disclosures; Gartner / Forrester surveys; analyst channel checks |
Scoring is 0–100 per indicator. Weighting is equal (1/6 each) in v0.1; the architect may adjust in v0.2. The convergence flag operates independently of the composite score — it fires on the count of red/amber-red indicators, not the average score, because convergence across independent methods is more informative than an average that can be dragged by one extreme value. Scores update when new filings publish; a “last updated” timestamp appears per indicator.
This is an instrument. Not an oracle.
Four things are true about this scorecard, and we say them plainly because anyone who uses it deserves the honest version of what it can and cannot do.
It does not predict when.
The scorecard detects fragility and divergence. It does not know whether the correction is six weeks or two years away. Timing is the one thing the instrument explicitly disclaims. Burry was early by eighteen months and nearly got liquidated by his own investors before the thesis resolved.
Being early is the known failure mode.
High fragility scores in a market with strong momentum can persist for a long time. The instrument is not a trigger. It is a reading of how much structural risk has accumulated. What the catalyst will be, and when it arrives, is outside the model.
Quality is bounded by data quality.
Where data is thin — depreciation policies, off-balance-sheet structures, enterprise demand economics — scores are provisional. A scorecard built on incomplete data is better than none. It is worse than one built on complete data. We are transparent about the difference.
Analysis, not a recommendation.
Nothing on this page is investment advice. Fragility scores do not tell you whether to buy, sell, or hold anything. They show where the structural risks in the AI economy are concentrated, sourced to public filings, in the same way an epidemiologist tracks disease vectors: not to make you afraid, but to show you where the exposures are.
The donkey’s job is to point at the gap between the story and the tape.
What you do with that information is yours.