Oracle Piled Up Hundreds of Billions in AI Debt, and Investors Started Asking About the Money

Oracle has taken on hundreds of billions of dollars in AI infrastructure commitments, and its free cash flow has turned negative. At the end of the first quarter of fiscal 2026, its remaining performance obligations (RPO) rose to $455 billion, and by December 2025 they exceeded $520 billion. By then the company's total debt had crossed $100 billion. Habr described September 2026 as the moment when the question "who will pay" stopped being theoretical. The article's announcement gives no facts about that event, so what follows covers the published data and the logic that led to it.

Orders without the money to fill them

Oracle's model in the AI cycle works like this: it signs multi-year contracts for compute capacity, then builds or leases data centers, buys GPUs, and only afterward receives revenue. Years pass between spending the money and getting paid. Throughout that time, construction has to be financed somehow.

In September 2025, Oracle issued $18 billion in bonds. It also raised project financing for individual campuses. Its capital expenditure forecast for the fiscal year rose to roughly $50 billion. Operating cash flow does not cover spending on that scale. In the quarter reported at the end of 2025, free cash flow came in at roughly negative $10 billion.

One customer accounts for tens of percent of the backlog

Most of the RPO growth comes from a compute contract with OpenAI. According to press reports, it is worth about $300 billion over five years, with main deliveries scheduled to begin in 2027. For Oracle, it is both the main growth driver and the main concentration risk. OpenAI spends noticeably more than it earns, and it depends on further funding rounds, partners, and loans.

The chain is circular. The cloud provider needs market money to build capacity. The customer needs market money to pay for it. If the cost of capital rises for either link, both suffer.

Key metrics investors are watching:

• RPO: $455 billion as of September 2025, over $520 billion at year-end;
• total debt: more than $100 billion;
• fiscal-year capital expenditure: around $50 billion;
• quarterly free cash flow: about −$10 billion;
• key contract: about $300 billion with OpenAI, deliveries from 2027.

The market's reaction was mixed. In September 2025, the stock jumped more than a third in a single day, then fell noticeably, and Oracle's credit default swap spreads widened. Investors moved from judging the size of the orders to judging what money will be used to fulfil them.

What to do if you build on this market

For companies buying AI capacity, the takeaway is down to earth. Don't put your entire stack on a single cloud provider if its balance sheet rests on borrowed money. Contracts should spell out capacity delivery dates and penalties for missing them, because a construction delay costs the customer more than a discount saves.

For investors and startups, the main indicator is now the cost of capital, not the number of announced gigawatts. If a project depends on cheap debt, any rise in borrowing costs pushes its payback back by years. So in the reporting of Oracle and other large data center operators, watch not RPO but the ratio of capital expenditure to operating cash flow and the share of revenue from a single customer.

A limitation of this piece: the details of September 2026 are not disclosed in the original announcement, and assessing the event itself will require the full text of the analysis and Oracle's latest financial reports.