Home / How it works
How it works
A daily read on whether the AI debt boom is cracking, built from 17 signals in six layers.
How the score works
- Each signal is calm (0), watch (1) or flashing (2). Claude re-judges every one each morning from news, filings and the data feeds, and links the source of each reading.
- A signal flashes when its evidence crosses a set line, like CoreWeave’s CDS going above 1,000bp or several fund managers gating withdrawals at once. Watch means it’s heading that way. Each signal’s page spells out its lines.
- Each layer averages its signals: 1.5 or higher is flashing, 0.75 or higher is watch.
- The stress index averages the six layers on a 0 to 100 scale, split into five bands: Quiet (0–19), Stirring (20–39), Elevated (40–59), Warning (60–79), Panic (80–100).
- The chain reaction is the alarm that matters most: collateral, the weakest borrowers and the funds all flashing at once.
- Gauges, filings and book values aren’t scored into the index directly. They’re evidence Claude reads when judging the signals.
How we price GPU collateral
- The full-cost floor is what an hour of rent has to earn to pay off the chip (its price, life and cost of capital), its power and its data center space, spread over the hours it actually gets rented. The cash floor is power, space and running costs alone.
- We work it out three ways (cheap, mid, dear), moving life, cost of capital and occupancy together the way lenders do, and score on the mid case.
- The collateral cushion is the rent divided by the full-cost floor. It’s scored on fixed lines: the cash cost scores 100, 0.85× scores 75, 1× scores 50, 1.2× scores 25, 1.5× scores 0, with straight lines in between. It never switches to ranking against its own history.
- The floor isn’t a hard floor. Owners keep renting below full cost, down to the cash cost, so rents can fall a long way before they stop.
- The indexes disagree because they measure different things: Silicon Data Neo-Cloud rental index (SDH100RT, SDB200RT) uses standardised observations; Ornn Compute Price Index (OCPI) uses executed trades. We never average them.
- Rents are third-party index prices, taken from news reports for now. They’re shown with credit and never offered as data. Our inputs and floors are ours: collateral.json and collateral.csv list every input with its source and date.
What it isn’t
This isn’t a timing tool, and it isn’t financial advice. The readings are judgment calls based on public reporting, some of it secondhand or unconfirmed. Every one links to its source and date so you can check for yourself.
Where the data comes from
| Source | What we take | How often | Terms |
|---|---|---|---|
| SEC EDGAR | 8-K red flags, late-filing notices, BDC tender offers, and BDC book value and PIK income from quarterly reports | Daily | US government data; free to reuse. |
| OFR (US Treasury) | The OFR Financial Stress Index and its credit part | Daily, two days behind | US government work, free to reuse with credit. |
| Federal Reserve Board | The corporate bond (GZ) spread, the excess bond premium and its recession odds | Monthly | Public domain, free to reuse with credit. |
| FRED (St. Louis Fed) | The Fed’s loan officer survey, H.8 bank loans to nonbanks, the St. Louis Fed stress index, and the VIX | As published | Fed series are free to reuse with credit. The VIX is Cboe’s, so it isn’t in downloads. |
| Yahoo Finance | Daily closes for the 18 tape names and the 10 BDCs in the book check | Daily | Shown for reference; not licensed for redistribution. |
| ERCOT, PJM, Berkeley Lab | Data center hookup queues, capacity auction results, the national interconnection queue | Checked weekly | Each reading links its source. |
| News and filings | The 17 signal readings and the event log, written fresh each morning with a link to the source of every reading | Daily | Our own summaries; facts link to their sources. |
Using our data
The stress index, layer scores, signal readings and archive are ours, free to reuse under CC BY 4.0: credit “Glitch in the Credit” and link to the page you used. Third-party numbers (prices) keep their owners’ terms. The data page has everything as JSON and CSV.
Spotted a mistake? Email [email protected].