On 10 September 2025, Oracle experienced a stock-market moment almost unimaginable for a company approaching its fiftieth birthday. Its shares surged roughly 36% after investors learned that the value of its contracted future business had more than quadrupled over the preceding year. For a time, the rally propelled co-founder Larry Ellison past Elon Musk in the rankings of the world’s wealthiest people.
The excitement rested on an extraordinary promise. Oracle had committed to supplying immense amounts of computing power to artificial intelligence companies, including OpenAI, and expected its cloud infrastructure business to grow at a pace few established technology companies could match. Yet the buildings, chips and electricity needed to fulfil those contracts would cost tens of billions of dollars before much of the associated revenue arrived.
By 2026, that distinction had changed how investors viewed Oracle. Its revenue and order backlog were expanding, but capital expenditure, borrowing and future lease commitments were expanding too. Understanding this gamble requires going back to the business Oracle originally built—because almost everything that made its software empire financially attractive was different from what its AI infrastructure strategy would demand.
How Oracle Built a Software Empire That Printed Cash
Oracle began in California in 1977 as Software Development Laboratories, founded by Larry Ellison, Bob Miner and Ed Oates. Ellison and Miner had worked on a database-related project at Ampex, and the name Oracle was associated with their early work. Their new company set out to commercialise a technology that promised to make storing and retrieving business information far easier.
In 1970, IBM researcher Edgar F. Codd had described the relational model for databases. Rather than organising information primarily through rigid chains of records, a relational database represented information as related tables. A bank, for instance, could connect information about customers, accounts and transactions without manually building every possible relationship into a separate file. Structured Query Language, or SQL, eventually provided a standardised way to ask those databases questions.
IBM developed important experimental systems, but Ellison’s company moved quickly to market. In 1979, it released Oracle Version 2, an early commercial relational database built around SQL. The number was deliberately reassuring: there had been internal development before the release, but no commercially sold Version 1. Oracle’s account of its early database development describes the evolution of a product that became central to large organisations’ operations.
The database’s position inside a business made Oracle unusually powerful. Banks, retailers, airlines, hospitals and government agencies depended on their databases to process transactions and preserve records. Applications were written around particular systems; employees learned to administer them; outages could interrupt critical operations. Replacing a database therefore involved more than purchasing competing software. It meant migrating valuable information, adapting applications, retraining teams and managing the danger that something essential would stop working.
Oracle built attractive economics around that dependence. Customers could pay substantial upfront licence fees, followed by recurring support charges for updates, fixes and continued assistance. Annual support was often priced at around 22% of the licence value, although actual contracts and entitlements varied. Software could be reproduced for another customer at relatively little marginal cost, while support payments continued long after the original sale.
Licensing complexity added another dimension. The amount owed could depend on processors, deployments, user numbers and contract terms. Oracle retained contractual audit rights, and customers sometimes faced significant costs when their actual installations differed from licensed entitlements. The system did not make customers literally unable to leave, but it raised the practical and financial price of doing so. For Oracle, the combination of recurring revenue, switching costs and embedded customer relationships produced a formidable business.
Those cash flows helped finance expansion by acquisition. Oracle bought PeopleSoft in 2005 after a prolonged takeover struggle, then added businesses such as Siebel Systems and BEA Systems. Its 2010 acquisition of Sun Microsystems brought hardware, Java and MySQL into the portfolio; NetSuite extended its reach into cloud-based business applications. The $28.3 billion Cerner transaction, announced in 2021 and completed in 2022, took Oracle further into healthcare records. Oracle became a broad enterprise technology supplier, but its foundations remained deeply connected to the economics of software.
Why Oracle Resisted the Cloud Before Betting on It
When cloud computing became a fashionable industry term, Ellison was openly sceptical. At a 2008 event, he mocked the vague way technology companies used the phrase, arguing that established computing activities were simply being relabelled. His criticism had an element of truth: software delivered remotely was not a completely new invention. But it underestimated how profoundly the commercial and financial relationship between a customer and its computing supplier could change.
Amazon Web Services had already begun letting customers rent computing capacity rather than acquire and maintain all their own hardware. As cloud services matured, businesses could provision servers, storage and other resources on demand, paying for usage or subscriptions. Providers benefited from running immense shared infrastructure; customers gained flexibility and avoided some upfront costs. The model made the data-centre operator, not the customer, responsible for much of the initial investment.
For Oracle, this was an awkward trade. It had spent decades making money from organisations that maintained complex systems themselves. A move toward cloud rental threatened parts of that familiar licence-and-support arrangement, even if it also offered recurring subscription revenue. Cloud infrastructure additionally demanded continuous investment in processors, networking, buildings, cooling and electricity—expenses that distributing an extra copy of database software did not require.
Oracle nevertheless introduced its public cloud plans in 2011 and progressively built out services for applications and infrastructure. Early offerings struggled to match the reach, developer adoption and breadth of Amazon’s business. Microsoft Azure and Google Cloud also became major competitors. The difficult task was not merely creating cloud products; it was persuading customers that Oracle could deliver reliable capacity at global scale.
An important technical shift came with Oracle Cloud Infrastructure, particularly the Generation 2 architecture promoted from 2018. Oracle emphasised bare-metal computing, network isolation and infrastructure designed for demanding enterprise workloads. Those design choices later became useful for AI customers seeking large clusters of powerful processors connected by fast networks. The claim that OCI could handle such workloads was not sufficient by itself, however: Oracle still had to acquire chips, secure sites and demonstrate that its facilities could be delivered on schedule.
Leadership also changed as cloud grew more central. Safra Catz became co-CEO with Mark Hurd when Ellison stepped away from the chief executive role in 2014; Ellison remained chairman and chief technology officer. Oracle hired Clay Magouyrk, an engineer with experience at AWS, during that era of infrastructure development. In September 2025, Magouyrk and Mike Sicilia were named co-CEOs, while Catz moved to a board-level role. The succession underlined how important cloud infrastructure and industry-specific applications had become to Oracle’s identity.
The AI Boom and Oracle’s $455 Billion Turning Point
The public arrival of ChatGPT in late 2022 changed the economics of the computing market. Training advanced AI models required vast clusters of specialised chips, particularly Nvidia graphics processing units, or GPUs. Running the resulting models for millions of users also required continuous computation. Data-centre capacity was no longer simply an efficiency tool for conventional software; it was becoming a competitive bottleneck for companies racing to build more capable AI systems.
Oracle entered the spotlight in January 2025 when the Stargate infrastructure initiative was announced at the White House. OpenAI, SoftBank, Oracle and other partners outlined a plan targeting as much as $500 billion in AI infrastructure investment over four years, with an initial emphasis on facilities in the United States. The project was not a $500 billion bill that Oracle alone had agreed to pay. Different participants had different roles in developing, financing and operating the infrastructure.
The proposed Abilene, Texas, campus became a prominent example of the physical scale involved. A giant training site required more than installing servers in an existing warehouse. Facilities had to be built in stages, filled with expensive equipment and linked to power systems that could serve exceptionally heavy loads. Progress would be measured not just by signed contracts but by completed buildings and usable computing capacity.
Then came Oracle’s September 2025 earnings announcement. Its remaining performance obligations, or RPO—the value of contracted goods and services not yet recognised as revenue—reached approximately $455 billion, a 359% increase from a year earlier. Oracle’s September 2025 financial results also outlined expectations for extraordinary growth in cloud infrastructure revenue. The company reported four multibillion-dollar contracts with three customers, demonstrating how sharply a handful of large agreements could change its projected business.
One agreement attracted particular attention. According to reporting on the approximately $300 billion computing deal, Oracle had reached a major multiyear arrangement with OpenAI. The reported size suggested a relationship capable of reshaping Oracle’s future, but the full private terms were not disclosed. A headline contract value could not reveal precisely when services would be delivered, what minimum payments applied or how obligations might change under different circumstances.
Investors initially focused on the growth opportunity. Oracle’s shares surged around 36% on 10 September, its best one-day performance in decades, and the value of Ellison’s holdings soared. The market was no longer valuing Oracle only as a mature database company. It was considering whether Oracle could become one of the essential infrastructure suppliers to the AI industry.
The excitement contained a fundamental asymmetry. A database licence could begin producing attractive revenue without Oracle building a factory for each new buyer. An enormous AI cloud contract created demand for capacity that might not yet exist. Oracle could have the customer agreement in hand and still face years of construction, equipment purchases and financing before the economics of that agreement became clear.
How a $664 Billion Backlog Becomes a Financing Problem
RPO is easy to mistake for a pile of cash waiting to arrive. It is not. The figure represents contracted future performance: services and products Oracle expects to provide and recognise as revenue over time. Some commitments may stretch across many years, and their conversion into reported sales depends on delivery schedules and contractual conditions. Even when an agreement is legally binding, the headline amount says little about the profit Oracle will ultimately earn from fulfilling it.
That distinction became increasingly important as the backlog grew to roughly $664 billion by 31 August 2026. The number indicated substantial demand for Oracle’s future offerings, but every additional infrastructure commitment brought questions about the equipment and facilities needed to honour it. If a customer contracts for enormous computing capacity, Oracle must make sure servers, networks and buildings are available when the customer needs them.
The financing can involve several parties. A developer might acquire a site, construct a data centre and borrow against expected rental income. Oracle might sign a long lease for that facility, install or operate computing equipment, and sell the resulting services to OpenAI or another customer. A chip supplier might provide financing for hardware, while some customers might prepay for capacity or supply equipment themselves. Each arrangement shifts costs, timing and risks among participants rather than making them disappear.
This creates different kinds of obligations, which should not be added together as though they were equivalent. Corporate bonds are debt that typically carries interest and repayment terms. Recognised lease liabilities reflect existing lease arrangements under accounting rules. Additional leases that have been signed but have not commenced can create substantial future commitments without yet appearing as recognised lease liabilities. Customer contracts, meanwhile, represent expected future receipts for services Oracle still needs to deliver.
Oracle’s filings make the scale unusually visible. At the end of November 2025, it disclosed around $248 billion in additional lease commitments not yet commenced, associated with long-term data-centre and cloud arrangements. By 31 August 2026, that amount had reached approximately $288 billion. These future commitments were distinct from the leases already reflected on its balance sheet. The figures should not be interpreted as a single immediately payable debt balance, but neither can they be ignored simply because many payments begin later.
The company’s August 2026 quarterly filing also provides the context for understanding its much larger RPO and associated financial arrangements. The critical question is whether contracted income arrives early enough, and with sufficient margin, to support the costs of creating the capacity. Revenue recognised over a decade does not automatically pay for equipment bought today.
There is also a mismatch between how long different assets remain useful. A data-centre building may serve customers for many years, while the processors inside it can become less competitive as newer chips offer more performance or greater energy efficiency. Oracle must recover not just construction and financing expenses but also the economic cost of replacing or upgrading hardware. Higher utilisation can help spread those costs over more revenue; underused capacity can do the opposite.
Electricity introduces another constraint. An AI campus can require immense power, and its profitability depends partly on the cost and availability of that power. Delays in grid connections, cooling systems or chip deliveries can postpone the moment Oracle begins earning revenue from a promised deployment, even while development and financing commitments continue. Large-scale infrastructure is therefore partly a scheduling challenge, not simply a demand forecast.
Oracle has argued that its model reduces the amount of external capital it must raise. Certain customers have agreed to prepayments or to provide their own hardware, while vendor financing and other structures can reduce immediate cash requirements. Management cited approximately $75 billion of customer prepayments and customer-supplied hardware commitments in 2026. Those arrangements can improve the funding equation, but their precise financial effects depend on when cash or equipment is actually delivered and what obligations accompany it.
The result is a business that looks simultaneously secure and exposed. The substantial backlog provides unusually strong visibility into potential future demand. Yet Oracle must orchestrate landlords, contractors, chipmakers, utilities, lenders and customers to turn those contracts into working capacity and profitable cash flow. Its financial health depends as much on the pace and cost of execution as on the number printed beside RPO.
The Cost of Growth: Debt, Cash Burn and Investor Anxiety
Oracle’s fiscal 2026 results captured the tension. Revenue increased to approximately $67.4 billion, up 17%, while operating cash flow reached around $32 billion. For many companies, those would be impressive figures indicating a healthy core business. But Oracle’s capital expenditure climbed to roughly $55.7 billion, up 162%, leaving free cash flow at approximately negative $23.7 billion.
Free cash flow measures the cash remaining after operating cash generation is reduced by capital spending. A negative figure does not necessarily mean the company is unprofitable under accounting rules, and it certainly does not mean it is bankrupt. It does indicate that internal cash generation was insufficient to fund the scale of investment being made during that period. Oracle needed financing beyond the cash its operations produced.
The shift becomes clearer in Oracle’s fiscal 2026 annual filing. Over that fiscal year, Oracle raised substantial funds through debt and equity, including approximately $43 billion of debt and $5 billion of stock financing as described in its funding programme. Borrowing can be rational when it supports projects with dependable future returns, but it also creates repayment and interest obligations that do not vanish when enthusiasm for a technology changes.
In September 2025, Oracle completed a major bond offering. As its data-centre ambitions expanded, attention turned to how much more debt and lease exposure the company might accumulate. Bondholders have a different priority from shareholders: they generally want to know whether a borrower can service its obligations, not whether its stock might double if everything goes right. A spectacular backlog can excite equity investors while leaving creditors concerned about the cash needed to fulfil it.
That concern intensified in 2026. Credit default swaps, which investors can use to insure against a borrower’s default, became more expensive, signalling greater perceived credit risk. S&P subsequently lowered Oracle’s credit rating to one notch above speculative-grade territory. That remained an investment-grade rating, not a declaration that the company was about to fail, but it illustrated how the financing burden had changed investors’ assessment of risk.
The bond market also became the setting for a legal dispute. In January 2026, investors brought a lawsuit alleging that documents accompanying Oracle’s September 2025 debt offering inadequately described the company’s existing plans for additional borrowing. Those were claims in litigation, not judicially established findings of wrongdoing. The importance of the dispute lies in the wider concern it expressed: investors wanted greater clarity about how much capital Oracle’s expansion would ultimately require.
The costs of transformation extended to the workforce. Oracle’s employee count declined by approximately 21,000 in the fiscal year ending May 2026, to roughly 141,000. The reduction included restructuring and ordinary attrition, so it would be misleading to describe every departure as a confirmed layoff. Oracle also recorded about $1.8 billion in restructuring charges and linked part of its operational changes to AI adoption and changing organisational needs.
Market reactions reflected both possibilities. Investors could recognise rapid revenue growth and still sell the shares when spending exceeded expectations or financing requirements appeared larger than anticipated. The stock suffered a severe decline in June 2026 as scrutiny of free cash flow and credit exposure intensified. In the space of months, the dominant question shifted from how large Oracle might become to how the company would finance the route there.
Two Concentrated Risks: OpenAI and Larry Ellison
Among the largest uncertainties is Oracle’s dependence on a relatively small number of extraordinary customers. A diversified database business serves thousands of organisations across industries, with no single new deployment necessarily determining its future. Large AI infrastructure agreements can concentrate years of projected revenue in just a handful of counterparties. Losing, delaying or renegotiating one relationship can therefore alter the financial outlook substantially.
OpenAI is the clearest example. The reported agreement valued at around $300 billion over roughly five years is exceptionally large relative to Oracle’s historical cloud business. Analysts have estimated that OpenAI represents a substantial share of Oracle’s total backlog, although exact customer-by-customer contractual terms are not publicly disclosed in full. Estimating concentration by dividing a reported contract value by overall RPO is illustrative, not a substitute for a disclosed accounting breakdown.
The apparent mismatch between the two companies’ scale is difficult to ignore. Spread evenly across five years, a $300 billion contract would average $60 billion annually. OpenAI’s reported annualised revenue run rate approached $50 billion by October 2026, but a run rate is not the same as completed-year revenue, profit or cash available for infrastructure purchases. Nor does the illustrative $60 billion average establish the deal’s actual payment schedule. It simply shows why Oracle’s assumptions about its customer’s future growth matter so much.
OpenAI has raised substantial financing and developed relationships with multiple technology suppliers. Its rising commercial revenue provides a credible reason to expect demand for computing capacity, but its future ability to honour the full range of its agreements depends on sustained customer adoption and access to funding. The broader dynamics of OpenAI’s wider infrastructure commitments illustrate how suppliers, investors and customers can become financially interconnected without every announced dollar representing an immediate cash payment.
Oracle has tried to reduce the concentration risk. Management said in September 2026 that it had signed more than $30 billion in new AI-related agreements during the quarter and that backlog from customers other than OpenAI had more than doubled over the preceding year. This provides evidence that demand is broadening. It does not eliminate the exposure created by the largest existing commitments, especially while their associated facilities are still being built.
The second concentration concerns Ellison’s personal wealth. His fortune is closely tied to his Oracle holdings, giving him strong financial incentives in the company’s long-term success. Oracle’s September 2026 proxy disclosures reported that he beneficially owned approximately 1.16 billion shares and had pledged about 413 million shares as collateral for personal obligations. The pledged figure was larger than an earlier disclosed total of 346 million and represented a significant portion of his stake.
Pledged stock introduces a distinct risk. If share prices fall sharply and collateral arrangements require additional security or repayment, a shareholder may face pressure unrelated to the operating performance of the company in a particular quarter. That does not mean a forced sale is inevitable or that the company itself owes the shareholder’s personal debt. Oracle’s board expressed confidence in Ellison’s ability to meet the relevant obligations without selling pledged shares.
Ellison was also involved in financing connected to his son David Ellison’s entertainment-sector acquisition activity, including the Paramount–Warner Bros. Discovery transaction completed in October 2026. The scale of these arrangements drew attention because Oracle shares underpin so much of Ellison’s wealth. Yet personal guarantees, share pledges and entertainment-company debt must remain distinct from Oracle Corporation’s borrowing and lease commitments. Treating them as one corporate debt pile would badly distort the financial picture.
Can Oracle’s AI Gamble Pay for Itself?
By the autumn of 2026, Oracle was attempting to demonstrate that the acceleration in spending had a credible endpoint. Its quarter ended 31 August showed revenue of approximately $19.3 billion, nearly 30% higher than a year earlier, and cloud infrastructure revenue growth of 121%. RPO reached $664 billion. These numbers suggested that the demand side of the strategy was developing rapidly, even as its financing remained under scrutiny.
On 28 October, management presented an unusually ambitious medium-term plan. At Oracle’s October 2026 Investor Day, the company reaffirmed a target of at least $225 billion in annual revenue by fiscal 2030. It projected that free cash flow would turn positive in the second half of fiscal 2028 and reach at least $30 billion in fiscal 2029. Those are forward-looking targets, not financial results already achieved.
Oracle also said that additional capital-markets fundraising through fiscal 2030 would be limited to approximately $20 billion, split between debt and equity, excluding refinancing. That limit followed an already completed $20 billion equity raise in fiscal 2027 and therefore should not be interpreted as a cap on everything Oracle had ever borrowed or raised. Management’s case depended in part on customer prepayments, customer-provided equipment, vendor financing and eventually lower capital requirements as major infrastructure deployments were completed.
The optimistic case has several credible elements. AI usage could continue to expand across consumer services, business software, coding, scientific research and automated workflows. Oracle’s large contracts could provide a reliable stream of revenue once customers begin consuming the promised capacity. Higher utilisation, efficient hardware deployment and prepayments could allow cash receipts to catch up with investment. If Oracle executes successfully, it could emerge with a substantially larger recurring infrastructure business while continuing to benefit from its established database and applications franchises.
The sceptical case rests on equally concrete considerations. Oracle expected around $90–95 billion of capital expenditure in fiscal 2027, a figure that underscores the extraordinary scale of the buildout. Construction and electrical delays could postpone revenue; rising financing costs could reduce returns; cheaper and more efficient chips could accelerate hardware obsolescence. A customer’s strong growth today does not guarantee that every planned deployment will remain economically attractive years from now.
The most useful indicators are therefore more specific than the stock price or the headline backlog. Investors need to watch how much RPO converts into recognised revenue and operating cash, whether new facilities become operational on time, and whether cloud margins improve as deployments mature. They should also examine capital expenditure relative to management’s guidance, the direction of lease commitments, the cost of new debt, and whether Oracle can reduce its reliance on a small set of customers. Positive free cash flow would be particularly meaningful because it would indicate that the business is beginning to fund its investments internally rather than depending so heavily on new capital.
There is a final distinction between a successful AI industry and a successful investment in AI infrastructure. Artificial intelligence could transform the economy while some individual projects earn disappointing returns. An infrastructure supplier can be right about future demand but wrong about the price it pays to secure that demand, or the timing of the revenue that follows. Oracle has made a wager on both the importance of AI and its own ability to build capacity profitably at unprecedented scale.
Oracle did not reach this point because its database empire had collapsed. It made the leap from a position of substantial strength, using its established business and customer relationships to pursue a far larger opportunity. The decisive question is whether its new cloud contracts can eventually deliver the kind of durable, self-financing cash flow that made the old Oracle so formidable. Until facilities are operating, customers are paying and free cash flow consistently improves, the answer remains open.
Last Updated on November 2, 2026 by Aseem Gupta