What an edition contains

Each quarterly edition covers calendar-quarter filings for nine entities — Microsoft, Alphabet, Amazon, Meta, Oracle, CoreWeave, Nvidia, Meta’s Hyperion financing vehicle and, once its filings are public, Anthropic — built on one method and delivered as documents a subscriber can put into their own systems:

• The registers. Obligations on and off balance sheet (recorded debt, leases commenced and not yet commenced, purchase obligations classed by kind, guarantees and residual-value support at their maximum, exposure to unconsolidated vehicles, vendor financing, customer prepayments, positions in other members of the stack); the facilities that finance compute, with their collateral, covenants, amortisation and reporting provisions as filed; the flows between members of the stack, gross and netted; the holders of each instrument, with the share that cannot be identified published as such; and a register of every item sought in the filings and not found, with the reason it is not public.
• The exposure sheets. One page per entity in an identical layout — the headline triple of recorded, committed and contingent obligations; coverage of fixed charges and the forward funding gap by quarter; loan-to-value on two published residual curves for secured compute debt; holder shares; dependence on the two largest model builders; four standard sensitivities; and, in the corner of every sheet, the share of its inputs that rests on a filed document — and a stack view.
• The note. Ten to twelve pages: what is owed, what carries it, what secures it, who holds it, what is circular, what is not disclosed, and what changed since the last edition. Every number in the text traces to a register row, and every register row to a page of a filing.
• The dataset. The registers as a workbook and as CSV with a data package, the assumptions file with its version, the methodology note and a change log.

Edition 1, on calendar Q3-2026 filings, is in preparation.

How it is built

The sources are the filings themselves — annual and quarterly reports, the credit agreements and guarantees filed as exhibits, fund and business-development-company holdings, registries, rating-agency disclosures — read in full, quoted and page-stamped before anything becomes a row. Nothing is typed into a computed table: a change in an assumption restates every entity and every back-quarter together, and the assumptions are published. Licensed data may calibrate a curve; it never appears as data, so every published number traces to a document the subscriber could obtain. The method page sets out the model.

What stays free

The Return on Thinking publishes the free layer — the Fragility dial, the deal ledger and the collectors behind them — complete at its own level of aggregation, under an open licence, and nothing there is withdrawn or thinned to sell this tier. Thinking on Credit adds what takes judgement and maintenance rather than collection, cites the free rows, and uses the same conventions: a source, a retrieval timestamp, a tier and a basis on every row.

What it is not

Thinking on Credit assigns no rating, grade or probability of default to any issuer, and makes no recommendation to buy, sell or hold any security. It is accounting and cash-flow arithmetic on disclosed commitments; where judgement enters, it is an assumption in a versioned file applied to everyone alike.

Who it is for

Credit funds and insurers lending into the stack or holding its paper; the bank desks that arrange it; asset managers and family offices with exposure to it; and the supervisors, researchers and journalists who need the filed facts in one place. Founding subscriptions are open for the first editions.