Oracle turned itself into a leveraged bet on one customer

Oracle has reported remaining performance obligations of $638 billion, a figure reported as 363 percent higher than a year earlier and larger than the backlog of companies several times its size. Over half of it is reported to come from a single customer, OpenAI, under a contract reported at more than $300 billion over five years beginning in 2027. That is an unusual thing for an enterprise software company to be, and it explains a market behaviour that would otherwise make no sense: an AI model launch moving the share price of a database company. Start with what a backlog is and is not. Remaining performance obligations are contracted revenue not yet recognised. They are a real commitment and a genuinely useful disclosure, and they are not cash. Oracle said 12 percent of the backlog is expected to convert during the next twelve months, with a further 34 percent across the following two years. Roughly half, therefore, converts beyond three years, on infrastructure that in many cases does not exist yet. The building is the part that changes the risk profile. Fiscal 2026 revenue rose 17 percent to $67.4 billion and cloud infrastructure revenue grew 77 percent to $18.1 billion. Delivering against the backlog required $55.63 billion of capital expenditure in the same year, which pushed free cash flow to negative $23.7 billion. Set those numbers next to each other and the shape is clear. Oracle is spending close to its annual revenue on capacity, in advance, against contracts that mostly pay later, for a customer base concentrated in one name. Whatever that is, it is not a software business. It is closer to an infrastructure developer building to a long-term offtake agreement, and it should be assessed the way project finance is assessed: can the offtaker pay, over the full term, under conditions nobody can currently forecast. OpenAI's ability to pay more than $300 billion over five years from 2027 depends on OpenAI's revenue over that period, which depends on how the frontier model market develops. That is a question about a market that did not exist five years ago, resolved by companies whose economics are not public. Oracle's shareholders are, in effect, holding a position on it. Which is why the correlation exists. When OpenAI ships something significant, or when a competitor does, the market is repricing the probability that Oracle's largest contract performs. GPT-6 Astra began rolling out on 3 September, classified by OpenAI at the Critical level for cybersecurity capability under its own framework. NVIDIA confirmed a $12.93 billion acquisition of Hugging Face days earlier. Every such event is information about the same underlying question. The growth path Oracle has guided to compounds the sensitivity. Cloud infrastructure revenue is projected at $18 billion this fiscal year and then $32 billion, $73 billion, $114 billion and $144 billion over the subsequent four years. A curve of that shape has no room for a large customer renegotiating. There is a second-order effect worth noting for readers in this region. The capital expenditure has to go somewhere physical, and the supplier decisions that follow are large. Oracle agreed a multi-year global arrangement under which it will deploy HPE Juniper routing and switching across its AI data centres, an agreement in which HPE issued Oracle warrants over HPE shares. Warrants issued to a customer are a sign of how much bargaining power a buyer at this scale carries, and that bargaining power exists because of the backlog. The counterargument is straightforward and serious. Concentration risk is only risk if the counterparty fails, and OpenAI is among the best-capitalised private companies in the world with demand that visibly exceeds supply. Oracle spent a decade as the cloud provider nobody chose and has, in two years, acquired a position in AI infrastructure that its scale would not otherwise have bought. Taking a large concentrated bet was arguably the only way to get there, and the alternative was continued irrelevance in the fastest-growing part of enterprise technology. Backlog of this size also gives Oracle exactly the procurement leverage the HPE arrangement demonstrates. Both readings are consistent with the same facts, which is what makes this the most interesting position in enterprise technology right now. The disclosures that would settle it are specific: the proportion of RPO attributable to the largest customer, stated by Oracle rather than reported; the payment schedule and any minimum commitments; and whether free cash flow turns as capacity comes online or capital expenditure keeps pace with the guided revenue curve. The figures in this article are as reported and several are attributed to coverage rather than to Oracle's own statements. Anyone acting on them should read the company's filings, which is where the customer concentration disclosure, if it is made, will appear.