Oracle's $900 Billion Collapse: From a 36% Surge to $153 Billion in Debt

The closing bell on a Tuesday in September 2025 wrote Oracle's history in a single candle: the share jumped 36%, adding $244 billion in a day to reach $922 billion in market value, and 81-year-old Larry Ellison briefly became the world's richest person. Per CNBC, it was Oracle's best day since 1992; the trigger was not the quarter's earnings but a promise — a roughly $300 billion, five-year compute contract with OpenAI and a 359% surge in remaining performance obligations (RPO — contracted future revenue) to $455 billion. Wall Street loved the story because for fifteen years Oracle had been a slow dividend stock; finally it had a growth narrative.

To understand the sparkle you must go back to 1977. Larry Ellison and two co-founders started with about $2,000 in capital while a 1970 paper by IBM researcher Edgar Codd on the relational database model — storing data as relations between tables — sat on the shelf. IBM owned the idea but moved slowly; Oracle shipped the first commercial relational database in 1979 and ran alone for years. Like a city's water grid, a database is invisible but indispensable: once bank balances, flight reservations and patient records sit on Oracle, moving them takes years, costs hundreds of millions and risks outages with every migration.

Oracle turned that stickiness into pricing power. On top of the license fee customers typically paid about 22% per year for support; in roughly five years you had paid for the software twice and the bills never stopped. Few companies in history sustained such margins, and Oracle became a cash machine. This is layer one — what the video says; layer two is why it was so profitable — the economics of fear. For the person choosing a database for a large bank, picking Oracle was the safest sentence in the room: if Oracle fails, the whole industry is blamed; if an unknown startup fails, you are blamed.

That safe choice fueled an appetite for buying rather than competing. After an 18-month fight, PeopleSoft was taken for $10.3 billion in 2005, Siebel followed the same year for $5.85 billion, then Hyperion, BEA and Sun Microsystems — those five deals alone cost more than $35 billion, with the 2022 purchase of health-software firm Cerner for $28.3 billion as the largest in company history. The formula never changed: buy the rival, migrate its customers onto Oracle support contracts, harvest cash. The machine ran inside an empire centrally controlled by Ellison — CEO for 37 years, chairman and chief technology officer after 2014, still owning about 40% — whose $300 million purchase of 98% of Hawaii's Lanai island, yacht racing and sports teams symbolized how big, personal bets succeed brilliantly when right and dangerously when wrong.

The Price of Dismissing the Cloud: A 3% Story

The muscle for buying made the company miss the wave it could not buy. In 2006 Amazon announced a modest service: AWS (Amazon Web Services — compute rented by the hour). The pitch was simple: rent power by the hour instead of buying expensive servers and licenses up front, a direct threat to Oracle's bulk-sale and forever-loyalty model. Ellison publicly dismissed cloud in 2008 as a marketing term for things already in the industry; Oracle did not ship a serious infrastructure cloud (IaaS — Infrastructure as a Service, infrastructure delivered as a service) until 2018. By then Amazon, Microsoft and Google controlled roughly two-thirds of the global infrastructure market. Synergy Research's Q2 2025 data clarifies the picture: global spend hit $99 billion, about 63% sat with the big three, Oracle stuck around 3% for quarters; its $6.7 billion in cloud revenue in June 2025, up 27% year-on-year, was a small slice of a huge pie.

Market share was not the only erosion; customer sentiment decayed. Many customers described Oracle's license audits as a sales tactic more than compliance, and in 2023 a shift in Java pricing from per-user to per-employee across the whole company led Gartner analysts to model two- to five-fold bill increases for some customers. Cerner was harsher: the U.S. Department of Veterans Affairs paused rollout in 2023 after outages and patient-safety concerns; the system meant to transform records for millions of veterans made headlines in Washington and subpoenas followed for Oracle Health executives. These stories remind us to translate jargon: not audit or license, but trust erosion. When alternatives finally existed, customers used them. By 2024 Oracle agreed to run its own database inside Microsoft, Google and Amazon data centers because customers refused to choose. The company that once demanded loyalty now begged to be a tenant inside rivals' buildings. The analogy helps: like an operator that collected tolls on the only bridge for years, then watches people build new roads and must carry its bridge onto others' land. This shows the collapse began in mindset before balance sheet; debt is just the invoice arriving.

The Stargate Casino: A $300 Billion Single-Customer Bet

That weakness met a seductive gamble in 2025. In January at the White House, Ellison stood beside OpenAI's Sam Altman and SoftBank's Masayoshi Son to announce Stargate, a plan to spend up to $500 billion on U.S. AI data centers, with Oracle building and operating most of them. September's RPO jump to $455 billion, driven mainly by the five-year $300 billion OpenAI deal starting in 2027, added about $100 billion to Ellison's net worth in a day. Yet RPO is not cash (Remaining Performance Obligations — contracted but not yet collected revenue); Oracle must build capacity first: land, power, buildings and especially Nvidia accelerators paid years before OpenAI revenues arrive.

Scale is frightening when quantified. The deal reportedly needs about 4.5 gigawatts of data-center power; one gigawatt can power roughly 750,000 homes, so Oracle committed to building several compute cities that consume the electricity of mid-sized metros — in a country where grid connections, transformers and cooling can take years. Then chip obsolescence: most data-center cost sits in Nvidia processors, a new generation appears almost yearly, and a campus full of prior-gen hardware loses value before its debt is repaid. The fix was borrowing: an $18 billion bond sale in September 2025, a $16 billion financing package for a single Michigan campus tied to OpenAI; total debt rose about 60% in a year to $153 billion, net debt about $88 billion after cash, and more than $55 billion in capital spending ate the free cash once used for dividends.

Why Did the Market Wake Up Late?

The worrying concentration hides inside the numbers. While backlog is quoted at $664 billion, $300 billion comes from one contract with one customer — nearly half. OpenAI, meanwhile, signed about $1.4 trillion in compute commitments across multiple providers over roughly eight years while annual revenue stayed in the tens of billions, a gap that openly raised sustainability questions in finance circles. In late 2025, with the share at its peak, credit markets quietly bid up the cost of insuring Oracle debt against default (CDS — Credit Default Swap, insurance against non-payment) and kept its rating a notch above junk, pattern-matching this single-customer debt-fueled expansion to prior telecom bubbles; the warnings were filed as bearish and ignored in the euphoria.

In April 2026 The Wall Street Journal reported that OpenAI missed its own user and revenue targets and that its finance chief privately warned that compute bills could be hard to meet if revenue did not accelerate; by July Oracle shares fell to a 14-month low, some traders openly calling the $300 billion deal a burden not a boon. Since then OpenAI pushed its IPO to 2027 at the earliest and, this month, news of talks to raise $1.2 trillion sent Oracle up 6% in a single day — a clear signal the stock no longer trades as a software company but as a bet on a startup's ability to keep raising money. The June 2026 SEC filing confirms the dual picture: Q4 revenue $19.2 billion, up 21%, cloud infrastructure $5.8 billion up 93%, RPO up $85 billion sequentially to $638 billion; demand is real, the question is whether the price is worth it.

Three Roads, One Truth: The Old Monopoly Is Over

Where does Oracle go from here? The three paths laid out must be read together with the numbers. In the bull case, OpenAI keeps raising, keeps paying, Oracle becomes one of the few physical backbones of the AI economy and today's expensive debt looks wise in hindsight — the selloff an overreaction. In the bear case, OpenAI slows, renegotiates or defers; data centers built for a customer that cannot fully use them sit half-empty while interest keeps ticking, forcing Oracle to sell new shares, dilute, sell assets or cut costs further — layoffs to protect cash have already been reported. The third, perhaps most insidious path is diversifying to more AI customers but at far lower margins than the legacy database business, a business eroding each quarter to cheaper alternatives, open-source Postgres and cloud-native offerings. For Ellison personally the trap is tight: owning about 40%, he cannot sell without signaling disbelief, and not selling leaves tens of billions exposed to swings. What is clear is that the Oracle that raised prices for 40 years because there was no alternative is gone; the new Oracle must compete on speed, delivery and cash it does not yet have.