$638B in Oracle Backlog, No Customer in the 10-K

Introduction

Oracle's FY2025 annual report described one of its customers as "an artificial intelligence (AI) product company that uses OCI to build and serve generative AI models." No name, but not exactly cryptic. In the FY2026 10-K, filed June 22, 2026, that sentence is gone, replaced by a link to oracle.com/customers. Oracle made the edit in the annual report for the year its contracted backlog grew from $138 billion to $638 billion, a haul S&P Global Ratings estimates is roughly half OpenAI's. If you hold an S&P 500 index fund in a 401k, you own a sliver of that bet, and the document you'd read to size it describes that customer only as a category.

$138 Billion to $638 Billion in Twelve Months

The Wall Street Journal's Berber Jin first reported the deal on September 10, 2025: roughly $300 billion of compute over about five years, 4.5 gigawatts of capacity. Neither company has ever confirmed that figure. What Oracle does confirm is the backlog. Remaining performance obligations, contracted work it hasn't billed yet, ran $97.3 billion in November 2024, $523.3 billion a year later, and $638 billion as of May 31, 2026.

Building that capacity costs real money. Capital expenditures went from $21.215 billion in FY2025 to $55.663 billion in FY2026, up 162%. Free cash flow ran negative $23.686 billion, against negative $394 million the year before. Oracle's own reconciliation puts free cash flow at -139% of net income, versus -3% in FY2025, on $17.087 billion of net income.

So Oracle went to the markets. Its February 1, 2026 financing plan called for $45 to $50 billion of gross proceeds: $25 billion of senior notes across eight tranches, $5 billion of mandatory convertible preferred, and a $20 billion at-the-market equity program. The 10-K also discloses $260 billion of off-balance-sheet lease commitments on terms of fifteen to nineteen years, including one where Oracle guarantees up to $3.3 billion of the lessor's borrowing. On July 9, 2026, S&P cut Oracle's long-term rating to BBB-, one notch above junk, naming OpenAI a key credit risk (MLQ News, heise).

The pattern running through all of it is a choice about where a name goes. Oracle uses OpenAI's in the filings that sell its story, never in the ones that carry legal liability for the story being wrong. No rule requires it to.

Where Oracle Does Name OpenAI

On February 2, 2026, the day before its $25 billion of notes priced, Oracle filed a free writing prospectus under the same registration statement as the offering. The text: "Oracle is raising money in order to build additional capacity to meet the contracted demand from our largest Oracle Cloud Infrastructure customers, including AMD, Meta, NVIDIA, OpenAI, TikTok, xAI and others." Oracle named OpenAI in its March and September 2025 earnings releases too. Those three are the entire set.

Search the FY2026 10-K, the FY2025 10-K, the November 2025 10-Q, or the final February 3 pricing supplement that actually sold the notes, and "OpenAI" returns nothing in any casing. The June 10, 2026 earnings release announcing that backlog skips it too. Section 11 liability attaches to a registration statement and the documents it incorporates by reference, not to a promotional free writing prospectus filed alongside it.

The deleted paragraph runs the same logic. FY2025's version gave four customer examples, the unnamed AI product company among them. FY2026 swaps all four for one sentence and a marketing link. What Oracle added instead is two lines of new risk language: revenue is now "more concentrated among a number of large customers," and a new Concentrations of Risk note says returns on its infrastructure investments depend on "the ability of our key customers to meet their contractual obligations." Two new sentences, for a year in which backlog grew 4.6x and capex grew 162%.

The 10% Rule Oracle Didn't Break

Oracle has a real answer, printed in the same filing: "No single customer accounted for 10% or more of our total revenues in fiscal 2026, 2025 or 2024." That's the standard that forces a company to name a major customer, and it keys on recognized revenue. OpenAI's exposure sits almost entirely in backlog, and nothing in GAAP requires a company to break a backlog down by counterparty. Oracle wasn't required to name anyone, and it didn't.

Two things cut Oracle's way. The FY2026 cash-flow statement shows $4.592 billion of deferred revenue from customer prepayments, against zero in each of the prior two years, and rent on leased capacity doesn't start until the facilities are running, which keeps that $260 billion off the current cash clock.

The number moving the other way is the recognition schedule. In November 2024, Oracle expected 39% of its backlog to convert to revenue within twelve months. By May 31, 2026 that share was 12%, while the slice landing 37 to 60 months out climbed from 19% to 34%. The backlog quadrupled and got slower to collect at the same time, which is not the direction you'd want both numbers moving.

A Michigan Pension Fund Wants Its Money Back

On July 1, 2026, the City of Sterling Heights Police & Fire Retirement System sued Oracle in Davidson County Chancery Court, Tennessee, naming executives including Larry Ellison and Safra Catz alongside four banks that underwrote the February notes: Deutsche Bank, BNP Paribas, HSBC, and Santander. The 20-page complaint alleges Oracle "was required to disclose at the time of the February 2026 Note Offering" that OpenAI, the customer it calls "primarily responsible for the staggering [remaining performance obligations] growth claimed in the registration statement," had "missed its internal revenue and new user targets" and that "OpenAI's own CFO had doubts raised about OpenAI's ability to pay for cloud computing power." The fund alleges a violation of Item 105 of Regulation S-K and seeks damages and rescission.

A separate federal case, Barrows v. Oracle, was filed in Delaware on February 3, 2026, covering buyers from June 12 to December 16, 2025 and naming Ellison, Catz, Clay Magouyrk, Mike Sicilia, Douglas Kehring, and Maria Smith. No court has ruled on the merits of either complaint.

The Sterling Heights theory carries a timing problem: the CFO-doubts allegation traces to a Wall Street Journal report published April 27, 2026, nearly three months after the notes priced. Barrows alleges something different, that Oracle misled investors about AI-driven growth while insiders sold, and its class period ends more than four months before that report ran. OpenAI denies the characterization on the record: Sam Altman and CFO Sarah Friar told Reuters, "This is ridiculous. We are totally aligned on buying as much compute as we can and working hard on it together every day." Spokesperson Steve Sharpe called the Journal's report "clickbait" and told Forbes the business is "firing on all cylinders."

Who Benefits

Oracle, first and most directly. Remaining performance obligations is the metric the market grades AI-era cloud companies on, and a backlog that size is why Oracle could sell $25 billion of bonds in a single week at coupons from 4.55% to 6.85% and authorize $20 billion of new equity behind it. Goldman Sachs led the bonds, and the underwriters collected roughly $75 million in fees.

Then the executives. Four Oracle insiders sold $1,871,638,462 of stock during the Barrows class period, per their Form 4 filings. Most of it is Catz: 8,694,918 shares for $1.83 billion, executed under a Rule 10b5-1 plan adopted September 25, 2024, nine to fifteen months before any event either complaint calls a corrective disclosure, a timeline that makes the plan a legitimate defense. The sales with no plan footnote are smaller, and carry no such defense on the face of the filings: Magouyrk's 40,000 shares on October 21, 2025 for about $11.07 million, Sicilia's 16,323 shares on September 16 that year for $5.12 million, and both of Smith's.

OpenAI's benefit is structural: it gets pre-committed capacity while Oracle carries the construction risk, the debt, the credit-rating exposure, and the fifteen-to-nineteen-year lease obligations. OpenAI points at the scale of its supply contracts while raising money; in March 2026 it closed $122 billion at an $852 billion valuation.

Your retirement account sits at the other end: real exposure, but several contingent steps away. Oracle's contracts would have to sour and the market would have to reprice the AI mega-caps your index fund is heavy in before any of this reaches your balance.

What OpenAI's Leaked 2025 Books Show

OpenAI's leaked 2025 financials, reported by Ed Zitron and independently verified by the Financial Times (via Fortune), show $13.07 billion of revenue against $34 billion of costs and a $20.92 billion operating loss. Net loss was $38.53 billion, including a one-time $41.55 billion non-cash conversion charge.

Oracle is profitable, with $67.4 billion of FY2026 revenue, up 17%, and Q4 cloud infrastructure revenue up 93%. OpenAI beat its own internal 2025 revenue goal, and Anthropic is running near a $47 billion annualized rate with an operating profit projected for Q2 2026, so "AI companies can't pay their compute bills" isn't the story either. Nvidia's Jensen Huang called the framing "ridiculous" back in January, and Bloomberg reported on July 27 that Google has effectively backstopped roughly $35 billion of Anthropic's leases.

What's specific to Oracle is that it borrowed against the number, and S&P now projects its free cash flow deficit widening to negative $42 billion in FY2027 on capex of $90 to $95 billion.

The Bottom Line

Oracle closed at $129.87 on July 31, up about 13% from its July 24 low after announcing a partnership putting Google's Gemini models on OCI alongside Meta's, OpenAI's, and xAI's, but the stock isn't the story: the downgrade, the leases, the negative free cash flow, and both lawsuits survived that rebound unchanged. The $3.3 billion lessor guarantee matures in September, and Q1 FY2027 earnings land September 8.

The 10% rule cuts both ways. The threshold keeping OpenAI out of Oracle's annual report is the one that will eventually put it in: backlog converts to billed revenue, and once a customer crosses 10% of total revenues, GAAP requires the name. S&P has cloud infrastructure heading from 27% of Oracle's revenue toward 60% by fiscal 2028. The disclosure investors would want today arrives on its own, at the point where the exposure has already been poured into concrete and fifteen-year leases. Which leaves a question neither lawsuit will settle: whether a rule keyed to billed revenue tells you anything useful about a company whose valuation rests almost entirely on revenue it hasn't billed.