Abstract. An index of the backlog six sellers of cloud and AI computing capacity file with the SEC, built on rules frozen in advance, with four registered tests. Backlog is four and a half times its level on file at March 2024.
1. The question
Steve Eisman argued in early October 2026 that the AI build-out is a chain resting on two customers, OpenAI and Anthropic. Companies filing 10-Qs must disclose each quarter their remaining performance obligations, or RPO, the revenue under contract but not yet earned, and when they expect to earn it. From it I can measure what was promised, what falls due soon and whether revenue later fell short.
2. The ledger
| Line | What it measures | Filed by 30 Sep 2026 | Base, filed by 31 Mar 2024 |
|---|---|---|---|
| Promised | RPO summed across the sellers that pass the rules, four at the base and six now | $2,472.3bn, index 445.9 | $539.0bn, index 100 |
| Due | Share expected within twelve months, at the sellers that file one (Microsoft, Oracle and, since 2026 Q3, IREN), which hold 54.7% of Promised | 21.6% | 44.5% |
| Delivered | Twelve-month promises whose year has closed | 19 closed, none short | |
| Spent | Capital spending plus finance-lease repayments per dollar of operating cash flow, at seven builders (the six sellers and Meta); 1.0 means no free cash flow | 0.87 | 0.42 (five builders) |
| Named | RPO a filing’s data tags to a named customer | 0.0%; peak 1.7% in 2026 Q2, when Amazon tagged an OpenAI figure | 0.0% |
| Filed by end of | Promised, $bn | Promised index | Due share, filers only | Spent |
|---|---|---|---|---|
| 2024 Q1 | 539.0 | 100.0 | 44.5% | 0.42 |
| 2024 Q2 | 570.1 | 105.8 | 43.3% | 0.41 |
| 2024 Q3 | 609.5 | 113.1 | 43.1% | 0.44 |
| 2024 Q4 | 614.1 | 113.9 | 43.4% | 0.49 |
| 2025 Q1 | 704.4 | 130.7 | 37.3% | 0.52 |
| 2025 Q2 | 740.2 | 137.3 | 37.9% | 0.57 |
| 2025 Q3 | 1,163.6 | 210.3 | 23.5% | 0.63 |
| 2025 Q4 | 1,329.0 | 240.2 | 23.0% | 0.66 |
| 2026 Q1 | 1,731.1 | 312.9 | 18.9% | 0.73 |
| 2026 Q2 | 2,201.4 | 397.9 | 21.0% | 0.79 |
| 2026 Q3 | 2,472.3 | 445.9 | 21.6% | 0.87 |
3. What it reads
Like for like, backlog is 4.46 times its base. Where a twelve-month share is filed, that share halved, though the dollars due doubled. Oracle’s fell from 43% at the base to 13%, most of it in the one quarter its backlog went from $137.8bn to $455.3bn.
Builders now put 87 cents of each dollar of operating cash into capital spending and finance-lease repayments, against 42 cents at the base, and over a dollar at Amazon, Oracle, CoreWeave and IREN.
The delivery record is clean and nearly uninformative. Nineteen twelve-month promises, all Microsoft’s or Oracle’s, have closed without a shortfall, and the most demanding needed only 78% of the prior year’s revenue.
Who owes it is filed only in part. Amazon’s filings tie a $38.0bn OpenAI commitment to its backlog and record announced expansions of $100bn with OpenAI and more than $100bn with Anthropic. CoreWeave’s name OpenAI, Meta and Microsoft, with contract ceilings but no backlog share. Microsoft’s 10-K reports $24.1bn of fiscal 2026 revenue from OpenAI. Its finance chief put OpenAI at about 45% of commercial RPO on a January 2026 call, a share I did not find in its filings. One Oracle earnings release names OpenAI as a customer, with no amount. Any other lab share is press.
4. The tests
| Registered 3 October 2026. Each asks whether a filed near-term share is a firm schedule | Resolves | |
|---|---|---|
| P1 | No twelve-month promise whose test closes between 4 Oct 2026 and 31 Oct 2027 falls short. Nine are on file today and three are already met | Oct 2027 |
| P2 | The mid-2026 promises of Microsoft (30% of $684bn), Oracle (12% of $638bn) and IREN ($0.9bn of $5.1bn) are met at the low edge of each rounded share, or at IREN’s filed amount: $188.1bn, $73.4bn and $0.9bn | Aug 2027 |
| P3 | Oracle’s revenue in the four quarters to 31 Aug 2027 reaches $83.0bn, the low edge of the 13% of $664bn it expects within a year. That is $20.75bn a quarter against $19.3bn in its latest | Sep 2027 |
| P4 | CoreWeave’s revenue in the eight quarters to 30 Jun 2027 reaches $14.9bn, the low edge of its filed 50% of $30.1bn. $7.6bn is filed, so the rest needs $1.8bn a quarter against $2.6bn in its latest | Aug 2027 |
Each bar is the company’s own filed figure, not my forecast. Of the registered tests, only Oracle’s latest and IREN’s need revenue above the latest quarter. I will publish every result.
5. What this does not claim
RPO is whole-company, not AI alone, carries no adjustment for the buyer’s credit, and stops where a buyer may cancel without substantive penalty. I chose the fourteen-registrant panel by hand; rules decide membership only within it. A met test does not prove conversion, since a year’s revenue includes contracts signed later. Only a miss proves anything. Nothing here says who will pay, or whether promises dated beyond a year convert. CCIR, Calcbench and Cloud Wars already publish cloud backlog totals near this level. What I add is the frozen rules and the delivery test. This is a research measurement, not a benchmark, not investment advice and not a view on any security.
- Method
- Inputs are XBRL facts from 10-Q and 10-K filings, dated by filing day, so each quarter holds the latest balances filed by its end, mostly one quarter older. 2026 Q3 holds June balances and Oracle’s of August. The index is chain-linked to 100 at 31 March 2024, measuring each quarter’s change only on sellers on file in that quarter and the one before. That is what like for like means here, and why an entry does not move it. The level, which includes CoreWeave and IREN, is 4.59 times the base and the index 4.46. Dollars due were $137.5bn at the base and $292.4bn now. A test fails only if revenue falls below the low edge of the filed share’s rounding, 27.5% for a share filed as 30%. IREN files an amount, so its bar is that amount. Table values are rounded from unrounded sums, so a few differ by one in the last place from the four-decimal series file. Spent is called FUNDED in the specification. The specification, which fixes the four tests, was frozen before the series was computed, though after I had summed the latest quarter’s backlog for the panel. Its SHA-256 is 9ce9f1796f3fa6be60661f2e2501f21b8bb96105f31f2238366a654bafc9aec9.
- Sources
- SEC EDGAR, a public record, read under the SEC’s fair-access policy. Oracle 10-Q 0000950170-24-029904 for the 43% of $80.2bn, 10-K 0000950170-25-087926 and 10-Q 0001193125-25-200095 for the $137.8bn and $455.3bn, 10-K 0001193125-26-277521 and 10-Q 0001193125-26-389274. Microsoft 10-K 0001193125-26-323660, including the $24.1bn of revenue from OpenAI, revenue sharing included, the $6.0bn OpenAI owed it at 30 June 2026 and its interest of about 25% in OpenAI. Microsoft 10-Q 0001193125-26-027207, the notes of 10-Q 0001193125-26-191507 and 8-K 0001193125-26-027198 were read for an OpenAI share of RPO and give none. Amazon 10-Qs 0001018724-26-000014 and 0001018724-26-000026. CoreWeave 10-Q 0001769628-25-000041, 10-K 0001769628-26-000104 and 10-Q 0001769628-26-000366. IREN 10-K 0001878848-26-000052. Alphabet 10-Q 0001652044-26-000071 for its two-year share. Oracle earnings release, 8-K exhibit 0000950170-25-036295, for the customer list. Microsoft FY26 Q2 earnings call, 28 January 2026, on its investor relations site. The Eisman interview is New Money on YouTube, video Zw0I7kb6zg4, posted about 2 October 2026. Summed cloud backlog: CCIR (ccir.io), Signed AI Commitments monitor, read 3 October 2026; Calcbench, blog post of 26 September 2026; and Cloud Wars, press, of 11 August 2026.
- Honesty
- This is one prototype vintage. Due covers only the three members that file a twelve-month share. Amazon files no split, and Alphabet and CoreWeave file a 24-month one; Alphabet’s is still half. Promised counts some capacity twice, since CoreWeave and IREN sell to Microsoft. Without both it is $2,363.5bn. Two sellers also fund buyers they name. Amazon’s 10-Q for the June quarter records $28.7bn invested in OpenAI preferred stock by 30 June 2026 and $21.3bn more afterwards, $8.0bn of Anthropic convertible notes bought through 2025 and $10.0bn of Anthropic preferred stock bought in the June quarter. Microsoft’s 10-K records $11.9bn of a $13.0bn commitment to its OpenAI investment funded by 30 June 2026. Sellers may leave contracts of a year or less and some usage fees out of RPO. Spent leaves out leases not yet begun, $288bn at Oracle alone, and counts missing finance-lease principal as zero. Named counts only RPO that a filing’s data tags to a named customer, so Amazon’s announced expansions, which are not balances, enter no line. It reads zero now because Amazon’s 10-Q for the June quarter tagged no OpenAI balance at 30 June, only the March figure again. P2 and P3 sit inside P1. Of the registered tests not yet met, Microsoft’s two need 31% and 52% of its latest quarterly revenue each quarter, Oracle’s three 65%, 93% and 107%, CoreWeave’s 71% and IREN’s 164%. CoreWeave’s four later 24-month promises, which no registered test covers, need 1.1 to 2.0 times its latest quarterly revenue. Every closed twelve-month test with a prior year on file, 18 of 19, closed on revenue above that year, so none was tried in a year of falling revenue. Dropping any one member leaves the index between 376 and 553. Of the rise in Spent, from 0.42 at the base and 0.63 a year ago, about 0.02 is new members. Oracle’s 10-Ks for fiscal 2021 and 2022 tag a full year’s revenue on a three-month period, so I derive those two quarters from year-to-date figures. The specification is version 1.1. Version 1.0 was replaced before any value was computed. The build code was edited twice after its first run. The first edit’s prior output was not kept, so that change cannot be audited. The second is the Oracle fix above, made after the series was computed. Its prior output is kept, the series did not change, and one closed Oracle test’s bar, unchanged at $26.3bn, went from 36% to 62% of the corrected prior year’s revenue. Use of this index as a reference in any financial contract, instrument or fund requires my written permission.
- Related
- No. 40, The Gate Is the Product, for the testing rules every note here follows. No. 36, Permissioned Abundance II, for the macro argument.