ED ZITRON: Oracle has tied its financial health to AI data center rentals, mostly for OpenAI, and its debt, binding leases, and weak core growth leave it little room to retreat

The Gist

Oracle bet heavily on renting AI data centers, mostly to OpenAI, and borrowed a lot to do it. It reportedly can't walk away from the leases, its credit rating is one step above junk, and the rest of its business isn't growing much. If OpenAI can't pay, Oracle has few places to turn. This packet steelmans the arguments made by Ed Zitron and Dean Baker in their best light for analysis and is not an endorsement of their conclusions.

Conclusion

Oracle has made AI data center rentals, concentrated in OpenAI, its main source of growth and has financed them with heavy borrowing and leases it reportedly cannot terminate, so it cannot easily retreat if that demand falters, and its credit position already reflects that exposure.

Premises

1. In September 2025 Oracle signed a reported $300 billion, roughly five-year computing contract with OpenAI that begins in 2027, and that contract drives much of its data center construction.
2. Oracle's remaining performance obligations reached about $664 billion in its fiscal first quarter of 2027, and third-party estimates attribute roughly half of that backlog to OpenAI, although Oracle does not disclose the split.
3. Oracle sent a force majeure notice over power delays at the New Mexico Project Jupiter campus, but reporting indicates the lease is hell-or-high-water, Oracle cannot terminate it, and the notice at most defers full rent while Oracle still carries debt costs.
4. Many of these campuses are owned by project companies funded by banks and private credit, and at the Michigan campus Oracle's lease payments are guaranteed by Oracle Corporation itself.
5. S&P cut Oracle to BBB-minus in July 2026, one notch above junk, its long bonds have traded around 8 percent, above average yields for lower-rated issuers, and its five-year credit default swap spread reached a record.
6. Outside cloud infrastructure, Oracle's revenue growth over the past 15 years has been modest and partly acquisition-driven, and it cut about 21,000 jobs in 2026, so there is little else in the business to offset an AI shortfall.

Assumptions

• Oracle's nominal revenue roughly doubled from fiscal 2011 to fiscal 2026; adjusted for inflation, real growth was modest rather than flat, and much of it came from acquisitions such as Cerner.
• A further downgrade would push about $120 billion of Oracle bonds out of investment-grade indexes; selling by holders is driven mainly by mandates and capital rules rather than a legal ban, and research finds such sales are real but often limited.
• Blue Owl said the force majeure notice did not change commitments to the project, and a brokerage expected minimal short-term impact on Oracle.
• The claim that Oracle laid off staff responsible for database license renewals rests on the speaker's sources and was not independently verified.
• Oracle's backlog is contracted revenue and could convert to cash if its customers pay as scheduled.
• Oracle's long bonds traded around 8 to 8.3 percent in late September 2026, slightly below the figure cited in the conversation.

Analysis

Argument type: Inductive.

Potential Fallacies

• Hasty generalization (P3 and P4 to the conclusion ('cannot easily retreat')) — Non-terminable lease terms are documented for Project Jupiter and a parent guarantee for the Michigan campus, but the conclusion applies 'cannot easily retreat' to Oracle's AI data center commitments as a whole. Terms, recourse, and termination rights can differ across project vehicles. The argument needs a bridging premise that these sites are representative.
• Conflating exposure with probability of loss (Conclusion and title ('little room to retreat')) — Fixed obligations and leverage magnify the loss if demand falls, but they say nothing about how likely that is. The premises establish severity, not probability. Without evidence on OpenAI's ability to pay or on re-letting options, the argument shows a large exposure rather than a likely failure.
• Non sequitur on offsetting business (P6) — Modest historical growth and layoffs do not show the core database and applications business cannot cushion a shortfall. A mature franchise can have slow growth and still generate substantial cash. Layoffs may also reflect restructuring to fund capex rather than weakness.
• Market pricing treated as independent confirmation (P5 and the clause 'credit position already reflects that exposure') — Elevated yields and record CDS spreads reflect the market's reading of the same public facts, and Oracle CDS is also used as a proxy hedge for AI sentiment. They show that the market sees risk, not that the risk is justified or will materialize, and they can overstate Oracle-specific distress.
• One-sided framing and selective emphasis (Title and conclusion) — The framing casts binding contracts as a trap and treats retreat as something Oracle will need. It leaves the mirror-image view in the background: the backlog is contracted revenue from a customer obligated to pay. Mitigating evidence (Blue Owl's statement, the brokerage's view, limited forced selling) sits in the assumptions rather than in the headline.
• Reliance on secondhand and single-source evidence (P1, P2, P3) — The contract terms, the OpenAI share of the backlog, and the lease characterization come from press reports, third-party estimates, and one commentator with a known bearish stance. The premises should not be treated as fully independent corroboration of one another.

Counterarguments

These challenges are not ranked.

• Conclusion (High impact) — The same hell-or-high-water leases and guarantees that bind Oracle are matched by a multi-year contracted backlog that binds OpenAI. This is ordinary infrastructure finance: obligations matched to contracted inflows. The exposure becomes dangerous only if the counterparty fails to pay, and the argument never estimates that probability or OpenAI's funding capacity.
• Conclusion (High impact) — Even if existing leases are fixed, 'cannot easily retreat' overlooks mitigation options. Oracle can sublease or re-let capacity to other AI customers, slow new commitments, sell assets, raise equity, or renegotiate. Lenders and lessors also have strong incentives to restructure rather than enforce, so contractual rigidity does not settle the outcome.
• Premise 3 (Medium impact) — The lease characterization rests on press reports rather than the lease documents, and Blue Owl's statement that commitments are unchanged and a brokerage's 'minimal impact' view suggest the force majeure episode is routine slippage, not a trap. Other campuses may have different termination, assignment, or recourse terms.
• Premise 6 (Medium impact) — Oracle's database, license-support, and applications franchises are sticky, high-margin, and cash-generative. Slow growth is not the same as inability to cushion a shortfall, and layoffs may reflect restructuring to fund AI capex.
• Premise 5 (Medium impact) — Record CDS spreads and 8% yields partly reflect Oracle's use as a liquid proxy for AI-bubble hedging, higher rates, and heavy debt supply. BBB-minus is still investment grade, and forced index selling is real but often limited, so the signals are consistent with a priced-in, manageable risk premium.
• Premise 2 (Low impact) — The OpenAI share of the backlog is an outside estimate and Oracle's non-disclosure may reflect customer confidentiality. The concentration claim therefore rests on inference, and the other half of the backlog is unexamined.

Suggested Improvements

• Counterparty risk — Add evidence on OpenAI's revenue, cash burn, funding runway, and backers, along with the contract's take-or-pay, ramp, and termination provisions. Whether Oracle's rigidity matters depends on whether the counterparty can pay. This is the crux and is currently unaddressed.
• Generalization across the portfolio — Quantify the share of campuses and lease obligations that are non-cancellable or carry Oracle recourse, using filings and project offering documents. This would replace extrapolation from Jupiter and Michigan with portfolio-level evidence.
• Quantifying 'little room to retreat' — Model Oracle's free cash flow, interest coverage, capex funding gap, and debt maturities under scenarios of delayed or reduced OpenAI payments, including the period before the 2027 contract start. The argument asserts inflexibility without measuring whether legacy cash flows could service obligations through a shortfall.
• Asset redeployability — Assess whether campuses and GPU capacity could be re-let to other AI tenants and at what discount. This determines whether inability to exit means manageable losses or ruin.
• Source diversification and balance — Corroborate key claims with primary documents and independent analysts, benchmark against hyperscaler peers, and integrate counter-evidence (Blue Owl, limited index selling) into the main reasoning rather than leaving it in the background. This reduces dependence on one commentator and makes the conclusion more proportionate to the evidence.
• Credit-signal attribution — Compare Oracle's spreads with BBB-minus peers and sector controls to test whether the premium is AI-specific. This would separate Oracle-specific risk from general AI sentiment and rate effects.

Scenario Tests

• OpenAI raises sufficient capital and pays on schedule from 2027, converting the backlog to cash. (Challenges) — The inability to retreat becomes irrelevant, and the leveraged build-out looks like a successful, if aggressive, strategy. The argument's real contribution is to flag the downside tail.
• OpenAI's funding falters and it seeks to renegotiate or reduce its commitment. (Supports) — Oracle would still owe fixed lease and debt payments on binding structures while revenue shrinks, pressuring cash flow and credit rating toward the junk threshold. Lender-led restructuring might soften the outcome.
• Data center capacity can be re-let to other AI tenants at near-market rates if OpenAI scales back. (Challenges) — Exit costs would be recoverable losses rather than fixed ones, weakening the 'cannot easily retreat' claim.
• S&P or another agency downgrades Oracle to junk, triggering index exclusions on about $120 billion of bonds. (Supports) — Funding costs would rise further as Oracle needs capital for the build-out, making the credit constraint real. The magnitude of forced selling is uncertain, since such sales are often limited and partly anticipated.
• Other campuses prove to have termination rights, non-recourse structures, or customer-side offsets unlike Project Jupiter and Michigan. (Challenges) — The generalization from two sites fails, and Oracle's flexibility is greater than the conclusion implies.
• Credit spreads normalize as sector sentiment improves while Oracle's fundamentals are unchanged. (Challenges) — This would show that the premium was partly a sentiment and proxy-hedging effect rather than a durable verdict on Oracle's exposure.

Coherence & Relevance

The argument is a convergent, cumulative case with four reasonably independent lines: contractual concentration, binding obligations and debt, market-priced credit risk, and a weak offsetting base. Together they coherently support a conditional claim of exposure and reduced flexibility, and the conclusion's hedges ('reportedly', 'easily', 'if demand falters') match the evidence. Coherence is weakest at the step from documented rigidity at a few sites to portfolio-wide inability to retreat, and at the missing probability term for demand failure. The stated assumptions supply important counterweights: contracted backlog, Blue Owl's reassurance, and limited forced selling. These temper the title's more categorical tone but do not contradict the core claim.

• In September 2025 Oracle signed a reported $300 billion, roughly five-year computing contract with OpenAI that begins in 2027, and that contract drives much of its data center construction. (Strong) — Establishes concentration and scale, but the contract's take-or-pay and termination terms and OpenAI's capacity to pay are missing. Costs are incurred now while revenue starts in 2027.
• Oracle's remaining performance obligations reached about $664 billion in its fiscal first quarter of 2027, and third-party estimates attribute roughly half of that backlog to OpenAI, although Oracle does not disclose the split. (Strong) — Supports 'main source of growth' and concentration, but backlog is also contracted revenue. The OpenAI share is an estimate and the composition of the remainder is unknown.
• Oracle sent a force majeure notice over power delays at the New Mexico Project Jupiter campus, but reporting indicates the lease is hell-or-high-water, Oracle cannot terminate it, and the notice at most defers full rent while Oracle still carries debt costs. (Strong) — Directly supports the lack of an exit, but only for one campus. It needs a bridging premise that these terms are representative. The delay is an execution problem, not a demand problem.
• Many of these campuses are owned by project companies funded by banks and private credit, and at the Michigan campus Oracle's lease payments are guaranteed by Oracle Corporation itself. (Moderate) — Shows how project debt can feed back to Oracle's balance sheet, but the guarantee is documented for one site and 'many' is unquantified. It does not show the borrowing is imprudent.
• S&P cut Oracle to BBB-minus in July 2026, one notch above junk, its long bonds have traded around 8 percent, above average yields for lower-rated issuers, and its five-year credit default swap spread reached a record. (Strong) — Directly supports 'credit position already reflects that exposure', but does not separate AI-specific risk from rates, leverage, or sentiment, and shows priced risk rather than likely realization.
• Outside cloud infrastructure, Oracle's revenue growth over the past 15 years has been modest and partly acquisition-driven, and it cut about 21,000 jobs in 2026, so there is little else in the business to offset an AI shortfall. (Weak) — Growth rate is not the same as cash-generating capacity, and layoffs are ambiguous. The legacy business's free cash flow relative to debt service is not quantified, so 'little else to offset' is asserted rather than shown.

View this argument on LogicFirst.ai