In Austin this summer, the accountants who decide what Oracle pays to borrow finished their work before the customers who buy what Oracle sells did theirs. On July 13, S&P Global Ratings convened a webinar to explain its downgrade of Oracle Corporation to BBB-minus, the lowest rung of investment grade, citing elevated debt levels from the artificial-intelligence buildout (S&P Global Ratings event listing, Jul 13; Cryptonomist, Aug 17). One month earlier, the same company had reported the strongest orderbook in its history: remaining performance obligations grew by $85 billion in a single quarter, from $553 billion to $638 billion (Oracle Q4 FY2026 results announcement, Jun 10). A business cannot be both the safest kind of borrower and the shakiest kind at once. The bond market has decided which one it believes.
Start with what each actor wants, because the wants no longer fit inside one balance sheet. Larry Ellison's Oracle signed OpenAI to a roughly $300 billion, five-year computing contract beginning in 2027, the largest deal in the company's life, and to deliver it, Oracle must build power plants' worth of machines before collecting most of the revenue (Wall Street Journal reporting via TechRepublic, Sep 2025). To fund fiscal 2027, management guided to about $70 billion in net capital spending, $90 to $95 billion gross counting customer prepayments, after spending $55.7 billion in fiscal 2026 (analysis.org, Jun 11). That is more than the company earned in revenue all of last year, when total revenues reached $67.4 billion (Oracle earnings statement, Jun 10). OpenAI gets compute without putting it on its own books; Oracle gets revenue it must first pay for; the lenders get caught in between.
The trigger this season was the ratings decision itself, but the pressure underneath has been building since September 2025, when that first giant contract turned a software company with fat margins into a construction firm with a financing plan. In February, Oracle told investors it would raise half its year's funding through equity and complete a single one-time investment-grade bond issue, promising no further bonds during calendar 2026 (Oracle investor update, Feb 1). Promises made at BBB-plus do not always survive contact with a $95 billion build program. By early August, the cost of insuring Oracle's debt against default had reached its highest level in nearly eighteen years, according to credit-market data compiled after the June results (iXBT citing CDS market data, Aug 6). The equity has absorbed worse: the stock has lost more than a third of its value over the past year despite the record backlog (Foreign Policy Journal, Aug 16).
Here is the bounded historical comparison, and it is the one anyone who lived through it will recognize. Between 1998 and 2001, WorldCom, Global Crossing and Qwest borrowed tens of billions to lay long-distance fiber against contracts signed with customers whose own businesses depended on cheap bandwidth. The networks were real, the contracts were signed, and the borrowers still went broke, because the payments arrived years after the debts came due and the customers proved less durable than the paper promised. The pattern to watch is not whether demand exists. It is who holds the obligation while the concrete cures.
The counter-example argues the other way, honestly stated. Comcast and the cable companies also levered up massively in the same era to rebuild their networks, and they thrived, because their revenue arrived monthly, from millions of households, none of whom could walk away in a single renewal. If Oracle's $638 billion backlog behaves like cable subscriptions, the debt is a rounding error against future cash (Oracle Q4 FY2026 results announcement, Jun 10). But it does not, quite: the backlog leans heavily on a handful of model developers, and OpenAI alone represents a commitment reported at roughly $300 billion (Wall Street Journal reporting via TechRepublic, Sep 2025). Concentrated counterparties are not households. They are other companies burning cash, which is why Nvidia spent late July negotiating to guarantee OpenAI's own leases, standing its balance sheet where OpenAI's unrated credit cannot (Bloomberg, Jul 26).
Walk the consequences forward and they land in a specific order. First, the lenders: a BBB-minus rating raises Oracle's cost of every future dollar, and the February promise not to sell more bonds becomes the first casualty if the build schedule holds. Second, the workforce and the suppliers: Oracle began another round of layoffs in August, cutting tens of thousands even as it pours capital into data centers, paying for silicon by shrinking payroll (The Next Web, Aug 13). Third, the public: twenty-three states have already approved special utility rates that let the largest power users pass infrastructure costs toward ordinary customers, and Oracle's new rating now sits below thresholds some state regulators set precisely for this reason (NationofChange, Aug 18). The buildout is financed in three places at once, the bond market, the employees and the ratepayers.
Who profits is equally specific. The equipment sellers get paid on delivery: Nvidia booked guidance near $91 billion in revenue for the quarter it reports August 26, much of it flowing from these very build programs (company guidance cited by CoinPaper, Aug 21). The power developers and the landowners in places like Pike County, Ohio collect twenty-year leases underwritten by chipmakers rather than tenants (Forbes, Aug 20). The risk sits almost entirely with the party that signed the middle of the chain, Oracle, which owns the buildings, owes the banks and waits on the customer.
Ask the question that breaks the read: what if the backlog converts on schedule and the cash arrives faster than the doubt? Then every lender who priced distress at eighteen-year-widest levels loses, and Oracle refinances its way back up the ratings ladder within two years. That scenario requires OpenAI to keep paying roughly $60 billion a year for computing, an expense OpenAI itself describes as its cost of goods sold, against revenue the same reports put at about $13 billion in 2025 (Yahoo Finance compilation of Reuters and The Information reporting, Sep 2025). The gap between those two numbers is the whole story. Someone funds it, and as of July, the answer is no longer allowed to be Oracle's unsecured bondholders alone.
What confirms the read: watch the next move in the credit, not the stock. If S&P's peers follow down, if the February no-more-bonds pledge is broken, or if Oracle's default insurance widens beyond the August highs, the market has decided the fiber-era comparison applies. What breaks it: a single quarter showing the OpenAI contract billing on schedule with gross margins holding, which would make this the rare borrowing spree that outran only its own reputation.
The judgment this piece earned is uncomfortable but simple. Oracle built the finest orderbook in corporate computing and bought it with the credit profile of a company one downgrade from the junk pile, and both facts are true because they are the same fact. The customer everyone is betting on has not yet earned enough to pay for the machines being bolted together in its name. When the world's richest companies start guaranteeing each other's customers, someone in the chain is doing the lending that used to be done by profits.
Oracle booked the biggest orderbook in its history and paid for it with the credit rating of a company one bad quarter from the junk pile, and those are the same transaction, not two.
Method. This analysis rests on the sources cited below. ARCANE does not publish a proprietary universe, cohort weighting or exclusion list for this piece — the reading is the desk's, argued from the record, not a screened back-test.