Oracle has reached a strange point in its AI business. The company is selling more cloud infrastructure than ever, signing enormous contracts and adding GPU capacity at a pace that would have been hard to imagine a few years ago. Yet the same business is forcing Oracle to spend cash faster than its revenue is coming in. And now some of the data-center projects tied to OpenAI, the customer at the centre of the story, are facing delays and changes.
The easiest headline would be that OpenAI is leaving Stargate and Oracle’s AI plan is falling apart. That would also be misleading. There has been no verified announcement that OpenAI has quit Oracle or cancelled the overall Stargate partnership. What has happened is serious enough without changing the facts: Oracle and OpenAI dropped plans for an additional expansion at a Texas site, a separate New Mexico project ran into a financing and construction dispute, and OpenAI has been lining up compute with other partners.
So here’s the question I’d ask if I were looking at Oracle as a business rather than a press release: what happens when a company makes a huge financial bet on AI infrastructure, but its biggest customer’s needs and timetable keep changing? That’s a better story than a dramatic declaration that Oracle is finished.
Access without medium partner: Oracle Bet Billions on Stargate
The Stargate Deal Was Too Big to Ignore
In January 2025, OpenAI, Oracle, SoftBank and other partners announced Stargate, with a proposed investment of up to $500 billion in US AI infrastructure over four years and a goal of 10 gigawatts of capacity. Those figures described an ambition and planned buildout, not money already spent or buildings already working. That distinction matters a lot in a project this big.
Then, on July 22, 2025, OpenAI said it had reached an agreement with Oracle for an additional 4.5 gigawatts of data-center capacity. Together with the first Stargate location at Abilene, Texas, OpenAI said the Oracle-linked work would take it beyond 5 gigawatts of capacity under development and involve more than two million chips. It was a huge statement about Oracle’s place in the AI market. A company that most people still associate with databases was suddenly being talked about as a major supplier of the computers behind frontier AI models.
Look at the size again. One gigawatt is 1,000 megawatts. A 4.5-gigawatt data-center commitment is an industrial construction and power problem, as much as a technology problem. It needs land, electricity, cooling equipment, networking, GPUs, construction crews and financing. The GPU purchase is only one part of the bill. Delays with any one of those can slow the entire site.
That is where the current trouble starts. A deal can be announced long before all the physical parts exist, and the timing of those parts matters to both sides. OpenAI wants computing capacity when its models need it. Oracle needs to build and finance that capacity without spending years on a site that isn’t earning enough revenue.
What OpenAI Actually Changed
On March 6, 2026, Bloomberg reported that Oracle and OpenAI had ended talks to expand their flagship Abilene data centre beyond the original plan. The report pointed to prolonged negotiations over financing and OpenAI’s changing computing requirements. Meta was reported to be considering the space that could have been used for the expansion, through developer Crusoe.
This was not a cancellation of Stargate I. The distinction sounds small until you think about how headlines spread. Losing a proposed extra phase of construction is different from losing an existing contract or watching a functioning data centre shut down. For investors, both developments can matter, but they have very different implications for revenue.
Then came the New Mexico problem. In September, Reuters reported that Oracle had issued a force majeure notice related to the Jupiter campus being developed by a Blue Owl-backed group through STACK Infrastructure. The facility was intended to support OpenAI. The notice raised questions about possible delays and sent concern through the AI-project financing market. Blue Owl said the notice did not change the parties’ commitment to the project.
A force majeure notice can be part of a contractual dispute over circumstances that interfere with delivery. It doesn’t automatically mean someone has walked away. Reuters reported this as a warning sign for project finance, not evidence that OpenAI had cancelled all its Oracle arrangements.
And OpenAI is not relying on Oracle alone. Its own Stargate announcements describe partnerships involving SoftBank, CoreWeave and Microsoft as well. More recent reporting has also described OpenAI’s efforts to secure large amounts of capacity through other infrastructure partners. From OpenAI’s point of view, spreading workloads across more suppliers can reduce the risk of waiting for any one project to finish. From Oracle’s side, it can mean more competition for each future expansion.
On October 15, The Wall Street Journal reported that OpenAI’s head of computing capacity, Sachin Katti, would leave the company the following month, with Uday Ruddarraju taking over his responsibilities. Management changes around an infrastructure programme this expensive deserve attention. But a personnel change is not proof that OpenAI is cancelling Oracle contracts, and I wouldn’t treat it as one.
Oracle’s Financial Numbers Are Both Excellent and Uncomfortable
Oracle’s September 10, 2026 earnings release gives us a more useful picture than rumours about Stargate. In the first quarter of fiscal 2027, which ended August 31, Oracle reported $19.3 billion in total revenue, up 30% from a year earlier. Cloud revenue reached $11.6 billion, up 62%. Within that, cloud infrastructure revenue grew 121% to $7.4 billion. Those are growth figures Oracle’s older cloud business would have loved to see.
Now look at spending. Capital expenditure in the same quarter was $28.5 billion. That is about $9.2 billion more than total quarterly revenue, or roughly 148% of it. Those two lines shouldn’t be confused with a profit-and-loss calculation: capex creates assets that may produce income for years, and revenue isn’t the same as cash flow. But putting them side by side shows the scale of the construction effort.
Oracle’s reported operating cash flow was $23.1 billion for the quarter, and its free cash flow was negative $5.4 billion after capex. The quarter’s operating cash flow was helped by sizeable customer prepayments. In its cash-outlay reconciliation, Oracle identified roughly $11.4 billion of customer prepayments with a financing component for capital expenditure. After accounting for those and other financing items, the company put its net cash outlay for capex at about $18 billion. That is still a lot of money, but it is less than the headline $28.5 billion.
There is an even larger number: $664 billion in remaining performance obligations, usually called RPO. Oracle said this balance had grown by $209 billion from a year earlier, and that it signed more than $30 billion of additional AI cloud contracts during the quarter. RPO is contracted work that has not yet been recognised as revenue. It isn’t $664 billion sitting in Oracle’s bank account. Revenue arrives over time, and some contracts could be revised under their terms or face delays in delivery.
The company also said it delivered more than 300,000 GPUs to AI cloud customers since the end of its previous fiscal quarter and added 850 megawatts of data-centre capacity. These numbers are part of the reason the bear case cannot just be “AI demand is imaginary.” Oracle is delivering capacity and growing its cloud infrastructure business, even while the financial cost of that growth becomes harder to ignore.
The $90 Billion Spending Plan Changes the Risk
Oracle’s guidance calls for around $90 billion to $95 billion in gross capital expenditure during fiscal 2027, according to its September earnings commentary. For comparison, it spent approximately $55.7 billion in the entire preceding fiscal year. Three years earlier, its annual capex was far smaller. The business has moved from running cloud services to financing an immense industrial buildout.
Oracle has said the net cash capex burden for fiscal 2027 should be no more than roughly $70 billion, after customer prepayments and related adjustments. That doesn’t make the physical spending disappear. It means some of the funding comes from customers and financing arrangements rather than being paid entirely from Oracle’s own funds.
And Oracle has been raising money. The company disclosed $43 billion in debt financing and $5 billion in equity financing in fiscal 2026. It planned to raise approximately $40 billion in fiscal 2027 through debt and equity, including a $20 billion at-the-market stock sale. In September, Oracle said it had completed that $20 billion common-stock sale during the first quarter. Existing shareholders should care about how much fresh equity is needed, even if they believe AI cloud demand will keep rising.
It is tempting to see the $664 billion backlog and decide all this spending is safe. I don’t think it works that way. A contract is valuable only if the company can deliver the capacity, collect the payments and earn enough after electricity, equipment, depreciation, financing and operating costs. GPUs also age quickly. A building can last decades; the accelerators inside it may need replacing much sooner. An AI data centre has a very different financial life from a traditional office building.
There are practical problems too. A delay in power delivery can leave costly equipment waiting. Construction can run over budget. A customer’s preferred chips or network layout can change before a planned site opens. Even a short postponement matters when large projects have financing arrangements that depend on reaching milestones.
This is the part of the Oracle story I find hardest to ignore. The company can be doing well operationally and still put pressure on its balance sheet because the investments need to happen before all the associated revenue is recognised. Strong demand and financial stress can exist at the same time.
If OpenAI Walked Away, What Would Actually Happen?
First, we need to separate a scenario from the present situation. There is no verified evidence that OpenAI has announced a complete exit from Oracle’s Stargate arrangements. A full departure is a hypothetical risk. A decision not to expand one location, on the other hand, has been reported. These are different events and should be analysed separately.
Suppose OpenAI cuts future capacity requests from Oracle but keeps existing commitments. The immediate problem would probably be slower growth relative to Oracle’s plans, rather than all current revenue disappearing. Oracle could seek alternative customers for uncommitted capacity, but replacing a customer at OpenAI’s scale would take time. It might also require different configurations or contract pricing.
Now imagine something worse: large contractual commitments are reduced or delayed after Oracle has already ordered equipment or backed construction. The financial exposure would depend on the exact agreements, penalties, customer deposits, ownership of hardware and timing of payments. Those terms aren’t all public. Anyone producing a precise dollar estimate of Oracle’s losses without access to those contracts is guessing.
A simple illustration helps. If a hypothetical $10 billion of future work were delayed by one year, Oracle would not automatically lose $10 billion of cash that day. The damage might show up as delayed recognition of revenue, extra financing costs, underused assets or lower margins. If the customer ultimately pays, the hit could be mainly about timing. If the project is cancelled and the costs cannot be recovered, the effect could be much worse. Neither example is a claim about an actual $10 billion Stargate cancellation.
There is a possible upside too. Demand for AI computing comes from more than OpenAI. Other model developers, enterprises and cloud resellers need GPUs. If Oracle can fill available space with customers that pay good prices and have reliable contracts, losing a proposed expansion with one customer might be manageable. But a data-centre tenant isn’t as easy to swap as a guest at a hotel. Power contracts, network design, hardware ownership and security requirements can complicate the change.
OpenAI Has Its Own Problem With the Math
OpenAI wants more compute, but it also needs a business that can support those bills. In October 2026, news coverage reported that OpenAI’s projected annual revenue was roughly $50 billion rather than the previously discussed $70 billion, a $20 billion gap in expectations. The reporting also described disagreements about how revenue should be counted relative to competitors. These are reported projections, not an audited statement of actual 2026 sales.
This matters for Oracle because the customer making the biggest infrastructure promises still has to convert AI usage into money. Training larger models, running inference, supporting enterprise customers and serving everyday chatbot users all cost money. Nobody knows with certainty how quickly revenue, model efficiency and compute prices will change over the next few years.
OpenAI also has a reason to avoid a single-provider dependency. If Oracle’s site is late, OpenAI needs alternatives. If another supplier offers a better deal, it has reason to negotiate. If newer accelerator designs are better for future models, it may want to redirect capacity plans. That is normal behaviour for a fast-growing customer, but difficult for an infrastructure provider putting billions into long-lived commitments.
It would be a mistake to assume that OpenAI’s diversification means Oracle loses by default. The demand for compute can be so large that several providers grow at once. The question is how much profitable business Oracle captures, and how much capital it must commit to capture it.
Is This the End of Oracle’s AI Plan?
Not yet. And a collapse isn’t what Oracle’s latest reported results show. Revenue is growing fast, cloud infrastructure is expanding at triple-digit rates, and the company has disclosed an enormous contracted backlog. There is no confirmed blanket withdrawal by OpenAI from the Stargate partnership.
But there is a real risk hiding under the good numbers: Oracle is increasingly behaving like an AI infrastructure builder whose success depends on completing giant construction projects on time, securing capital, and collecting payments over years. The old Oracle could sell databases and software subscriptions without needing to add hundreds of megawatts of electricity capacity in a quarter. The new Oracle must do that and keep the rest of its business working.
I would watch four things in upcoming reports. First, whether the New Mexico campus gets back onto a clear construction and financing schedule. Second, whether the Abilene project continues to deliver its originally planned capacity even without the extra expansion. Third, whether RPO begins converting to recognised cloud revenue at the promised pace. And fourth, whether free cash flow improves as more sites start earning money, or stays negative because the spending keeps moving faster than collections.
The more worrying scenario is not that OpenAI sends a single message saying it is done with Oracle. It is that several sites slip by six or twelve months, customers renegotiate later phases, equipment costs stay high and financing becomes more expensive. Each individual delay could sound manageable. Taken together, they could change the economics of a project.
The positive scenario is equally concrete. Oracle builds on schedule, keeps customers locked into long-term contracts, uses prepayments to reduce its own cash burden and spreads infrastructure demand across multiple AI companies. If it can turn that $664 billion backlog into profitable revenue over time, its AI investment could make it a much larger cloud provider than it was before Stargate.
For now, the evidence supports neither an obituary nor a victory speech. Oracle’s AI business is growing, but its financial exposure has grown with it. OpenAI has not publicly quit Stargate. What it has done is remind investors that even deals measured in hundreds of billions of dollars remain subject to changing plans, construction deadlines and the uncomfortable question of who pays when a project runs late.
That is why the next Oracle earnings report matters more than another giant Stargate announcement. I want to see how much capacity actually came online, how much of the backlog became revenue, and how much money was left after building it.