Larry Ellison Could Be Right About AI and Still Lose the Bet

Oracle Corp. reported $638 billion in remaining performance obligations for fiscal 2026, but S&P Global Ratings estimates OpenAI represents roughly half of that backlog, with a reported $300 billion commitment equaling about 47% of Oracle's total. Oracle is spending heavily and raising debt before much of the promised cloud revenue arrives, making its balance sheet a leveraged bet on OpenAI's continued growth and access to capital. Larry Ellison's vision of Oracle as a central AI infrastructure provider may be correct, but the financial structure may struggle to survive the journey.

TL;DR — Key Takeaways - Larry Ellison is positioning Oracle as a central infrastructure provider for the AI economy, but roughly half of its enormous backlog may depend on OpenAI. - Oracle is spending heavily, raising debt and reducing its workforce before much of the promised cloud revenue arrives. - The biggest risk is not that Ellison has misunderstood AI’s future, but that Oracle’s financial structure may struggle to survive the journey there. The New York Times Magazine profile of Larry Ellison https://www.nytimes.com/2026/07/31/magazine/larry-ellison-ai-oracle.html sees an empire being assembled. Oracle supplies cloud infrastructure and sits close to vast stores of enterprise data. Cerner puts it at the center of health care. Oracle’s position in the American TikTok entity provides proximity to one of the world’s most valuable collections of behavioral data. David Ellison’s Paramount Skydance controls an enormous library of entertainment and news content. Larry Ellison has influence in Washington and a growing role in the infrastructure supporting the AI economy. Put all of those pieces together and it is tempting to see a vertically integrated intelligence machine taking shape. There is just one problem. Much of the empire is still a construction site, and nearly half the projected rent may depend on a single tenant that is simultaneously cultivating several other landlords. Oracle reported an astonishing $638 billion in remaining performance obligations, or RPO, at the end of its 2026 fiscal year. RPO is revenue under contract that Oracle has not yet recognized because it has not yet delivered the corresponding services. It is not an account receivable. It is not money sitting in the bank. It is a promise of future business, contingent on Oracle performing and its customers remaining able to pay. S&P Global Ratings estimates that OpenAI represents roughly half of that $638 billion. The reported $300 billion OpenAI commitment alone would equal approximately 47% of Oracle’s entire backlog. S&P calls OpenAI a key credit risk https://www.spglobal.com/ratings/en/regulatory/article/-/view/sourceId/101695609 and says its ability to honor those commitments depends on continued growth in AI, continued access to outside capital and its models remaining market leaders. That is the number at the center of Ellison’s wager. Oracle’s $638 billion backlog is supposed to prove that the wager has already been won. Instead, roughly half of Oracle’s contracted future revenue may depend upon OpenAI. Oracle’s balance sheet is increasingly becoming a leveraged bet on Sam Altman’s ability to keep raising money. That is a very different story from the empire described by the Times. Ellison may understand perfectly well where the intelligence economy is heading. He may be right that AI will require an enormous new grid of data centers, chips, power and networking. He may even be right that Oracle can become one of the most important providers of that infrastructure. But recognizing the destination does not make Oracle’s financial route to it safe. Hosting the Data Is Not Owning the Intelligence The Times makes a critical category error by treating proximity to data as ownership of intelligence. Oracle hosts staggering quantities of enterprise information, including financial records, supply chain data, customer information, human resources systems and government workloads. Cerner places Oracle around some of the most sensitive health care information in the world. Its involvement with the American TikTok operation creates another potentially valuable connection to behavioral data. Paramount Skydance owns content that could be licensed for training or used in AI-driven products. These assets do not add up to a single Ellison-controlled training corpus. Most information residing in Oracle systems belongs to Oracle’s customers. Medical information is subject to the rights of patients, providers and health systems, along with extensive privacy and regulatory restrictions. Paramount Skydance is a separate public company controlled by David Ellison, not an Oracle subsidiary controlled by Larry. TikTok’s American operation is another entity with its own investors, contracts and continuing ties to ByteDance. A family relationship does not erase corporate boundaries, intellectual property rights, privacy protections or customer ownership of data. Even when data can legally be used for model training, supplying the data is not the same as owning the model trained upon it. Oracle could store the information, secure it, prepare it, broker access to it or provide a controlled environment in which a model maker uses it. Oracle could collect infrastructure fees, licensing revenue, access charges or a brokerage margin. Unless an agreement explicitly provides otherwise, however, the model maker retains the weights, the product, the customer relationship and the compounding value created by the resulting intelligence. Oracle may provide the land, the warehouse and access to some of the raw materials. It does not thereby own the factory or what the factory produces. That puts Oracle in a fundamentally different position from OpenAI, Anthropic, Google or another model maker. A model company can use Oracle infrastructure and licensed data to create an intelligence asset that can be improved, replicated and deployed across thousands of applications. Oracle gets paid for supplying the inputs. The model maker retains the reusable asset produced from them. Even the long-term premium attached to proprietary training data is not assured. Model development is already moving beyond the proposition that accumulating more human-created data is the primary route to better intelligence. Synthetic data, reinforcement learning, specialized post-training, inference-time computation, retrieval and live access to external tools are becoming increasingly important. Oracle could become a broker of training data just as the economic value migrates from possessing static data to orchestrating intelligence against live data at the moment work is performed. Oracle is not a model maker climbing the stack. It is an infrastructure provider that may also become a data broker for the companies that do. The Liabilities Arrive Before the Rent Calling Oracle a landlord may itself be premature. A landlord owns completed property and collects rent. Oracle is still the developer. Much of the infrastructure is under construction. Capital leaves Oracle now. Debt is raised now. Interest accumulates now. The revenue arrives later, assuming Oracle completes the facilities on schedule, secures enough power, obtains the required chips and delivers the capacity its customers ordered. Oracle spent approximately $55.7 billion on capital expenditures during fiscal 2026, compared with $21.2 billion the previous year. That spending approached 83% of Oracle’s $67.4 billion in annual revenue. The company produced a record $32 billion in operating cash flow, yet finished the year with negative free cash flow of $23.7 billion. To sustain the expansion, Oracle raised $43 billion in debt and $5 billion in equity during fiscal 2026. It expects to raise another $40 billion through debt and equity in fiscal 2027. Interest expense in the fourth quarter was already 47% higher than a year earlier. Oracle’s fiscal 2026 results https://investor.oracle.com/investor-news/news-details/2026/Oracle-Announces-Record-Q4-and-FY-2026-Results-Driven-by-Cloud-Infrastructure--Cloud-Applications/default.aspx show a company growing rapidly while consuming extraordinary amounts of outside capital to do it. There is real demand behind this spending. Oracle Cloud Infrastructure revenue grew 77% during fiscal 2026. Oracle says $75 billion of the hardware associated with its large AI contracts has either been prepaid by customers or supplied directly by them. That reduces Oracle’s capital burden and is an important protection against customers ordering capacity without putting anything at risk. It does not eliminate the larger concentration problem. Prepaid GPUs are not payment for years of future cloud services. Nor do they ensure that Oracle will earn an adequate return after accounting for land, construction, power, networking, depreciation, leases, financing costs and the continuing replacement of equipment. RPO measures future revenue, not future profit. Oracle could eventually recognize hundreds of billions of dollars from these contracts and still discover that it financed the least attractive part of the AI economy. The OpenAI relationship makes the imbalance particularly dangerous. OpenAI has overlapping infrastructure, investment and financing relationships with Microsoft, Nvidia, Amazon, SoftBank, CoreWeave and others. It is deliberately distributing its computing requirements, financing needs and counterparty exposure across a growing network. From Oracle’s perspective, the OpenAI commitment looks like validation. From OpenAI’s perspective, Oracle is one part of a diversified system of capital and capacity providers. OpenAI does not need every announced infrastructure project to succeed. It needs enough of them to succeed. Some eggs can break while it makes the omelet because it has persuaded several other companies to supply the eggs. Oracle has far less optionality. Once Oracle raises the money, enters long-term leases and begins building specialized facilities, it cannot diversify away from those obligations. OpenAI can change model architectures, shift workloads, renegotiate capacity requirements or direct future demand toward another provider. Oracle remains responsible for the debt and construction commitments already made. A contract may be legally noncancelable without being economically guaranteed. Its full value still depends on Oracle performing, OpenAI paying and the relationship surviving long enough for hundreds of billions of dollars in future services to be delivered. This is not primarily an accounts receivable problem. Oracle is not waiting for OpenAI to pay a $300 billion invoice. The more dangerous mismatch is that Oracle’s financing and construction obligations are becoming real before most of the OpenAI backlog becomes recognized revenue. The liabilities arrive before the rent. Is Ellison Killing the Goose? Oracle is also financing this wager by taking resources out of the business that must support it. Its workforce fell from approximately 162,000 employees to 141,000 during fiscal 2026. That is a reduction of about 21,000 people, or 13% of the company. Oracle incurred $1.84 billion in severance and exit costs, compared with $374 million the previous year. The company has cited management and product changes, performance issues, acquisitions, strategic shifts and AI as factors in its restructuring. Reuters placed the workforce contraction https://www.reuters.com/business/world-at-work/oracle-workforce-shrinks-by-about-13-2026-06-22/ against the cash demands of Oracle’s infrastructure expansion. We cannot prove that each eliminated job directly financed a data center, and Oracle says AI-assisted development allows smaller teams to produce more software. Some cuts may eliminate duplication, improve efficiency and make the company more competitive. But it would be equally misleading to present the layoffs simply as AI replacing workers. Oracle is moving capital from human and operating capacity into physical AI infrastructure. Those 21,000 people included some combination of engineers, salespeople, consultants, health care specialists, product managers and support personnel. They helped develop products, implement them for customers, integrate Cerner, sell additional services and maintain the enterprise relationships that generate Oracle’s recurring revenue. Ellison is assuming Oracle can remove 13% of its workforce without materially damaging product velocity, implementations, sales execution, customer support or retention. That assumption is now part of the AI wager. Oracle’s established businesses must continue producing the cash, growth and customer relationships that make the infrastructure expansion financeable. If the workforce reductions weaken execution, Oracle could create a negative feedback loop. Cuts intended to release cash for construction could slow growth in the existing business. Slower growth could increase Oracle’s reliance on borrowing. Higher interest costs could create pressure for further reductions. Oracle would then become increasingly dependent on infrastructure revenue that has not yet arrived. Ellison may be weakening the goose that lays the golden eggs to finance a much larger goose that has not begun laying. Microsoft, Amazon and Alphabet also spend extraordinary sums on AI infrastructure, but they finance those investments from larger, more diversified cash-producing businesses. Nvidia sells the scarce equipment to almost everyone and captures some of the richest margins in the market. Oracle is attempting its transformation with less room for delay and much greater exposure to the financing needs of one customer. S&P has already lowered Oracle’s credit rating to BBB-minus, the lowest investment-grade level, citing higher leverage, negative free cash flow and OpenAI counterparty risk. The Samuel Insull Problem This is where Samuel Insull enters the story https://www.repository.law.indiana.edu/facpub/308/ . Insull was not a fool who misunderstood electricity. He was one of the great visionaries of the electrical age. He understood that electricity would become an essential utility, that scale would reduce its cost and that controlling the infrastructure delivering it would create enormous power. He was right about the technology, the demand and the future. What failed was the financial architecture he constructed to reach it. Insull expanded through layers of holding companies supported by leverage and continued investor confidence. His system worked as long as demand grew, cash flowed and refinancing remained available. When the Depression weakened revenue and closed access to capital, the structure collapsed. Electricity did not fail. Insull’s ability to finance his position in the electrical economy failed. Oracle is not Insull’s pyramidal holding-company empire. It has substantial recurring revenue, a huge installed base, genuine cloud growth and real contracted demand. Predicting Oracle’s collapse would be premature and irresponsible. The analogy concerns the difference between understanding the future and surviving the journey to it. Ellison could be right that AI will become an indispensable utility. He could be right that inference demand will grow for decades. He could be right that enterprises and governments will require a new grid of data centers, chips, power and networking. None of those conclusions proves Oracle can safely finance its portion of that grid. Construction can be delayed. Power projects can slip. Chip architectures can change. Training can become more efficient. Infrastructure margins can fall. OpenAI can redirect demand. Credit conditions can tighten. Oracle’s existing business can slow after losing tens of thousands of employees. Any one of these risks might be manageable. Several arriving together could test a company that is already consuming extraordinary amounts of capital. Oracle is committing that capital to the ground floor of the intelligence economy while much of the richest value continues moving upward. Model makers own the weights. Agent platforms control orchestration. Applications control workflows. Distribution platforms control access to customers. Oracle may host the systems, supply the compute and broker access to data without owning the most valuable intelligence produced inside its facilities. That is the indispensability trap. Oracle could become essential to operating the intelligence economy without capturing the best economics the intelligence creates. The Times sees Ellison assembling an AI empire. I see a successful software company leveraging its balance sheet, reducing its workforce and consuming its cash to become a prospective landlord for model makers that continue shopping among several landlords. The wager deserves respect. Ellison may once again have recognized an industry transition before most of the market understood its scale. The danger is not simply that he could be wrong. Samuel Insull did not fail because America stopped needing electricity. His empire failed because its financial structure could not survive long enough to collect the returns from a future he had correctly predicted. Larry Ellison is now walking that same fine line.