Abstract. A forecast, graded as such. I looked in my own work for a claim with the shape of the 1992 sterling episode, a public commitment that could be abandoned, and found one. Tested against that structure it fails today, so I register it and let the dates decide.
1. The commitment
On 1 February 2026 Oracle published a financing plan to keep an investment-grade balance sheet and said so again in March. My Backlog Ledger (No. 44) sets out the arithmetic behind that promise. Oracle spends $1.62 on capital spending and finance-lease repayments for every dollar of operating cash, against 87 cents across the seven builders together. It has $288bn of leases signed and not yet begun, trailing free cash flow of minus $28.7bn, and a backlog that needs $20.75bn of revenue a quarter against $19.3bn in its latest. Last quarter $11.4bn of its operating cash was customer prepayments, and its first $20bn equity programme is used up.
The registered claims test whether three things can hold together: the grade, the build the backlog promises, and funding that does not come from the buyers whose credit is the risk. In 1992 the parity, German rates and British mortgages could not.
2. Why it is not a 1992 bet today
| Element | 1992 | Oracle, October 2026 |
|---|---|---|
| Ammunition that cannot be printed | Reserves of about $40bn against the selling | Partial. Oracle cannot print cash, but it can dilute and draw prepayments |
| A clock the defender does not control | A referendum on a fixed date | None. S&P’s outlook is stable, and its own criteria call that a rating not likely to change |
| The ally refuses | The Bundesbank declined to help | Shareholders and customers supplied $31bn in one quarter; the labs still raise at higher marks |
| A bounded cost of being wrong | Carry under 20bp a year, and the Bank had to buy at the band | Protection costs about 249bp a year on the public tape, between the BB (191bp) and single-B (312bp) index averages, so a label change is largely in the price. No level pins the spread, and fallen-angel spreads have on average tightened after an index exit, a base rate from the literature, not a view on this issuer |
Three of the four fail and the first is partial. The closest analogue is the French franc of September 1992, which carried a similar reading and held because its ally stood behind it. The conditions under which the three failing elements would turn are stated here in advance: an agency has started a clock, protection trades at or inside the BB index average, and the ally has refused on a dated event. Until then the family stays graded FORECAST ONLY.
3. The claims
Nine claims, registered 5 October 2026, frozen by hash on the Forecast Wing and priced 7 October 2026, scored by Brier whichever way they fall. Six are on Oracle, three on PJM’s capacity price cap, carried because my forecast F7 already sits on it. Probabilities are mine; the engine supplied the base rates.
| Id | Claim, in short | Registered deadline | P |
|---|---|---|---|
| O1 | Oracle’s middle rating is BB+ or lower on a month-end between 30 Oct 2026 and 30 Sep 2027 | 30 Sep 2027 | 8% |
| O2 | An agency starts a clock (a watch or review by any of the three, or a cut by any of them) by 31 Mar 2027 | 31 Mar 2027 | 30% |
| O3 | Oracle’s 10-Q for the November quarter shows prepayments below $11,363m | 31 Jan 2027 | 75% |
| O4 | OpenAI closes at least $30bn at $852bn or better by 31 Mar 2027 | 31 Mar 2027 | 70% |
| O5 | O1, scored only if O4 fails | 30 Sep 2027 | 20% |
| O6 | Both: No. 44’s P3 is met and O1 resolves YES | 30 Sep 2027 | 5% |
| G1 | The 2029/30 auction clears below the $337.83 cap | 31 Dec 2027 | 35% |
| G2 | No cap is in force when the 2030/31 auction opens | 31 Dec 2027 | 30% |
| G3 | The 2030/31 price clears above the old cap, re-expressed | 31 Dec 2027 | 18% |
Full wording, sources and resolution rules are on the wing. The events come sooner than the deadlines: Oracle’s November-quarter 10-Q about December 2026, the 2029/30 auction results about 22 December 2026, the 2030/31 auction in May 2027. O6 is the cell that matters: the promises are kept and the grade is lost in the same year, the pattern of the fibre build of 2000.
4. What this does not claim
It is not advice and I hold no position; this site trades on paper only. The structure does not guarantee the outcome: the franc held. Nothing here says Oracle will lose its grade or that any obligation will go unpaid; my ledger shows 19 of 19 closed promises met. The ratings, the March reaffirmation and OpenAI’s valuation come from secondary sources. Nine claims that share one driver, the AI capital cycle, are not nine independent bets.
- Method
- The structure test is the 1992 trade dissected into twelve elements, four of them fatal if missing: unprintable ammunition, no deep-pocketed ally, a bounded cost of being wrong and a dated forcing event. Forty-eight looks from my research, overlaps not removed, were screened against it. Claim text, criteria and sources were frozen on 5 October 2026 (SHA-256 2472c493734caac294363bedb2e458afdccf52fb05b521f1b9c4265b3efe5211) before any probability was set.
- Sources
- Oracle 10-Q 0001193125-26-389274 and 10-K 0001193125-26-277521; Oracle’s financing plan of 1 February 2026 (SEC free-writing prospectus); the Bloomberg US Corporate Index factsheet; PJM’s 2029/30 planning parameters and FERC docket ER26-3515; Working Paper No. 44 and its spec. Protection levels are daily medians from the DTCC public price-dissemination tape for 1 and 2 October 2026, a partial tape of reportable trades; index spreads are from FRED.
- Honesty
- The Oracle claims are probabilities on a rating label, set with the current spread in view, though no probability was derived from it; they are graded FORECAST ONLY and that grade is the point. O2 sits above the engine’s band; the seven others priced against a band sit inside it, and O3 has none. G1 cuts against my own F7 (72%): the two are scored jointly so that two YES results count once. No claim here passed a gate; an event of one cannot. The cost of being wrong is a Brier score and a line in my own graveyard.
- Related
- No. 44, The Backlog Ledger. No. 40, The Gate Is the Product. F7 on the Forecast Wing.