Build the strongest case that Oracle's AI buildout fails or is materially impaired:

Oracle’s AI buildout can fail, or be materially impaired, without a full AI bust. The mechanism is a duration and credit mismatch: Oracle has signed 15-to-19-year data-center leases and is financing a multi-year chip and power program largely with debt, while roughly half of its contracted backlog sits with one unrated, cash-burning customer whose compute purchase runs about five years and whose own funding plan assumes continuous external capital. Delays do not retire the obligations; they push revenue later and bunch refinancing into the same window. Credit markets are already pricing that path more harshly than equity bulls are.

Concentration is the credit, not a footnote

S&P Global Ratings cut Oracle to BBB− on July 9, 2026—one notch above junk, outlook stable—and named OpenAI the central credit risk. The agency estimated OpenAI accounts for roughly half of the $638 billion remaining performance obligations then on the books, and wrote that if OpenAI cannot pay, the data centers “will sit idle” while operations, maintenance, and depreciation continue, leaving Oracle with “massive datacenter leases that it might be unable to exit.” Cloud infrastructure is expected to rise from 27% of revenue in fiscal 2026 to nearly 60% by fiscal 2028, diluting the sticky database franchise that used to support the rating. Competitors have internal workloads that can absorb idle capacity; Oracle does not. [1] [2] [3]

The anchor contract is the roughly $300 billion, about five-year OpenAI cloud agreement reported in September 2025, starting in 2027 and tied to about 4.5 gigawatts. By August 31, 2026, RPO had risen to $664 billion. Reporting has put OpenAI’s share near 45–50%. Gil Luria of DA Davidson has called the contract “fantasy” and the chance OpenAI will have $300 billion to spend with Oracle “negligible.” OpenAI also has an exit or recommit option after about five years in some accounts of the deal, while Oracle’s property leases run 15–19 years. That is optionality for the customer and obligation for the landlord. [4] [5] [6] [7]

Circular financing tightens the same knot. Nvidia has invested in OpenAI and, per Nvidia’s own filings cited in research notes, entered guarantees capped at $105 billion to support land, power, and shell buildout for an OpenAI affiliate. OpenAI commits spend to Oracle; Oracle buys Nvidia systems to deliver it. Equity stakes do not protect the credit: they lose value in the same scenario in which the contracts fail. [8] [9] [10]

What this means for a competitor or creditor: Diversified hyperscalers can re-tenant or self-consume. Oracle’s residual demand is the open market after one customer’s plan changes. A partial scale-back, not a bankruptcy, is enough to impair returns.

OpenAI’s funding math does not yet cover the bill

OpenAI’s ability to honor the contract is a fundraising story, not an operating-cash-flow story. The Financial Times reported in September 2026 that an investor presentation projects negative free cash flow of about $278 billion from 2026 through 2030, with compute and infrastructure spend of about $856 billion over that horizon, even as revenue is projected to rise from $36 billion in 2026 to $350 billion in 2030. The same reporting said a $122 billion raise closed in March would be exhausted in 2028 on those figures. Separately, OpenAI told investors annualized revenue was approaching $50 billion at the end of September 2026, about $20 billion below the $68–70 billion figure that had circulated weeks earlier. Oracle, Nvidia, and CoreWeave sold off on the disclosure. [11] [12] [13]

At signing, the Wall Street Journal noted OpenAI was generating roughly $10 billion of annual revenue against an average annual Oracle bill on the order of $60 billion. Even at a $50 billion run-rate, the contract still requires revenue growth, new equity or debt, and continued willingness of strategic investors (Microsoft, Nvidia, SoftBank, Amazon) to fund losses. OpenAI has also signed large commitments elsewhere—Microsoft on the order of $250 billion of contracted sales, Amazon around $138 billion—so Oracle is not the only claimant on the same cash. An IPO has been pushed toward 2027; Altman has ruled out 2026. [4] [10] [12]

The bull reply is that OpenAI has already raised more than $100 billion, remains “very well capitalized” per its CFO, and that enterprise demand is real. The bear reply is mechanical: the presentation the FT saw has expenses outrunning a tenfold revenue ramp, and a miss of even one large round forces a renegotiation of compute orders. Oracle cannot compel a private company to keep buying.

The balance sheet is already the constraint

Fiscal 2026 capex was $55.7 billion against operating cash flow of about $32 billion, producing free cash flow of −$23.7 billion. Oracle raised about $43 billion of debt and $5 billion of equity that year. Trailing free cash flow through the August 2026 quarter was about −$28.7 billion. In fiscal Q1 2027 (ended August 31), capex was $28.5 billion and free cash flow −$5.4 billion, of which $11.36 billion of operating cash flow was customer prepayments classified as having a significant financing component. Management guided fiscal 2027 reported capex of $90–95 billion and roughly $40 billion more of debt and equity. S&P’s July action widened its fiscal 2027 free-operating-cash-flow deficit forecast to nearly $42 billion and expects adjusted leverage in the mid-4x area, above its 4x threshold for a plain BBB. [14] [15] [2] [16] [17]

Borrowings were about $125 billion at August 31, 2026. Bloomberg has described long-term debt as more than $160 billion after nearly doubling in two years, making Oracle one of the largest non-financial borrowers in the U.S. corporate market. Off the balance sheet, additional lease commitments that had not yet commenced were $288 billion at August 31, 2026, up from about $260 billion at May 31, substantially all data-center arrangements, terms of 15–19 years, expected to start between fiscal 2027 and 2029. Divided across those terms, fully commenced rent is on the order of $15–19 billion a year—before the GPUs inside the buildings are refreshed. [18] [19] [20] [10]

Depreciation has not caught the spend. Fiscal 2026 depreciation was $7.6 billion, about 14% of that year’s capex, after Oracle extended server and networking lives (peers have moved toward five to six years). Michael Burry’s estimate is that hyperscalers will understate depreciation by about $176 billion from 2026 to 2028 if economic life is closer to two or three years than six, and that Oracle’s earnings could be overstated by roughly 27% by 2028. A JPMorgan Asset Management sensitivity cited in industry research found restating GPU and networking capex since 2023 onto a three-year life cuts 2025 EPS by about 17% at Oracle, more than at Amazon, Meta, Microsoft, or Alphabet. Paul Kedrosky’s point is sharper than the accounting debate: 15-to-19-year leases against six-year book lives imply at least two hardware replacements inside a single lease, and unlike fiber or rail, GPUs can be economically worthless in five to ten years. [15] [21] [22] [23] [18]

Secondary-market prices currently sit above straight-line book value—an H100 still near 58% of purchase price after four years in one Silicon Data series—so bulls say six years is conservative while scarcity lasts. Burry’s rebuttal is that those “retained value” charts are discounted rental models, not clearing prices, and that scarcity from power and memory, not durability, is holding rents up. Once Vera Rubin-class supply arrives, legacy rental rates can fall fast. [24] [25]

Power delays slip the monetization clock, not the debt clock

On September 24, 2026, Oracle sent a force-majeure notice to Blue Owl’s STACK Infrastructure on Project Jupiter, a roughly 2.45 GW New Mexico campus intended for OpenAI, preserving the right to delay rent if the site misses a 2028 online date. Oracle says the schedule is unchanged and that such notices are routine. The fuel path is not. An Energy Transfer gas lateral was denied across state trust land and retargeted to February 2027; an air permit for the Bloom fuel-cell microgrid faced a November 23, 2026 decision deadline. Bloomberg reported Oracle was weighing trucked compressed natural gas so early phases could start. About $18 billion of bank loans on the campus have been quoted around 89–91 cents, and banks including Santander and Jefferies have struggled to distribute the paper. [26] [27] [28] [29] [30]

Wisconsin’s Project Lighthouse (about 1.3 GW) had its transmission review restarted after regulators withdrew a completeness finding, wiping months of process. Independent estimates have put meaningful power as late as mid-2028 and full load toward 2029, versus Oracle guidance of initial customer delivery in the second half of 2027. Morgan Stanley credit analysts Lindsay Tyler and Benjamin Guerrero reject the “delay helps cash” story: late buildings mean late revenue and late proof of concept, while GPU purchases bunch into a narrower window just as construction loans and lease commencements hit. Abilene construction loans of about $9.4 billion are interest-only and mature in 2028, when they must be taken out with longer paper priced off Oracle’s credit. If Wisconsin and Jupiter slip into the same year, capex and refinancing concentrate rather than smooth. [31] [32] [33] [34]

Local politics is a second constraint money cannot buy through. Research tracking local actions has counted pauses or moratoria in a rising share of U.S. counties, with Texas, Wisconsin, Georgia, and Michigan as live regulatory risks for specific campuses. [35]

The financing window is narrowing before the rating breaks

Five-year CDS closed at a record about 261 basis points in early October 2026, a level some desks translate into an implied five-year default probability above 20% while the issuer is still investment-grade. That is far above the sub-40 to sub-60 basis-point area of a year earlier, and above the roughly 150 basis-point level Morgan Stanley had flagged in late 2025 as the near-term break. Long Oracle bonds have traded with yields around 7.5–8.1%, roughly 2.5 points over Treasuries in one October reading, and some issues have been quoted near 85 cents. Goldman has described a surge in single-name AI CDS volume, with Oracle at the front. [36] [37] [38] [39] [18]

Morgan Stanley’s credit team (Tyler, and earlier David Hamburger) moved from a basis trade—long CDS, long bonds—to outright long CDS, citing the funding gap, growing balance sheet, capex and obsolescence risk, ratings pressure, and counterparty risk. Their adjusted net-debt path, which includes leases, was already described as possibly too low. They have also warned that even another ~$40 billion of customer prepayments may still leave two large parent bond deals needed by the end of fiscal 2028. Barclays has run a sensitivity in which cash is depleted by late 2026 if spending stays elevated. [39] [40] [41] [34]

Syndication is already sticky. Efforts to sell Jupiter-linked loans stalled on Oracle credit concerns. A separate bank group has spent months trying to distribute about $38 billion of construction debt on Texas and Wisconsin sites leased to Oracle, with some lenders willing to discount to non-bank buyers. Oracle is in talks with Apollo and Goldman Sachs on a structure in which outside investors fund an SPV that buys chips and leases them to Oracle—explicitly to avoid more unsecured parent debt. Morgan Stanley treats that as confirmation that chip-lease SPVs will be marked as leverage anyway. A second unsecured bond deal inside calendar 2026 would contradict management’s earlier “one-time” issuance framing. [29] [42] [19] [43]

The ratings trigger is binary for the buyer base. S&P’s downgrade trigger language includes leverage staying above 4x, a further deterioration in free cash flow, or evidence that OpenAI cannot perform. One more notch from a major agency would push a very large bond stock out of investment-grade indexes. Moody’s is already at Baa2 with a negative outlook; Fitch has been the more stable of the three. Wisconsin regulators have also tied financial-guarantee rules for data-center developers to credit ratings, which is why the July cut had a local legal sequel. [1] [44] [45] [29]

Short interest has risen into the decline—nearly doubling over a year to about 38 million shares in one S3 dataset—while five-year CDS widened in parallel. Michael Burry has disclosed shorts and argued that prepayments turn “the customer into the lender,” so a slice of future cloud revenue is a financing gross-up rather than usage. Hedge funds have used Oracle equity and CDS as a liquid proxy for less hedgeable neocloud and OpenAI credit risk. [46] [47] [48]

Demand can stay “strong” while pricing destroys the underwriting

Token prices have deflated at a speed that does not match 15-year leases. An a16z reading of Goldman data put an LLM price index near 5 by September 2026 versus 100 in March 2023, about a 95% decline. A Silicon Data token-expenditure index hit $0.97 per million tokens in August 2026, more than 50% below a May peak. Paul Kedrosky has described tokens as a hyper-deflationary commodity and argued that at very steep annual price declines a frontier buyer needs several-fold unit growth just to hold revenue flat. A Wall Street Journal essay in October 2026 framed the endgame as prices plunging by factors of ten as Chinese open-weight models already price around 90% below some U.S. frontier rates, even while volumes keep rising. [49] [50] [51] [52]

Oracle is exposed on both sides of the GPU-price move. SemiAnalysis’s Dylan Patel has argued Oracle locked large deals at essentially fixed margins just as cloud GPU prices rose—CoreWeave raised rates 25% in July 2026 and again since—so Oracle cannot reprice the windfall. If rental rates later collapse as new silicon floods a power-constrained market, long contracts help only if the customer still pays and still wants that generation. Kedrosky’s failure-mode arithmetic is that many low-probability breaks (power, credit, policy, price, a single tenant) compound into a high probability that at least one hits. [53] [18] [51]

A Financial Times calculation in November 2025, after the stock had given back the September deal pop, put the market’s implied value of Oracle’s side of the $300 billion agreement at about −$74 billion. That is a mark, not a cash loss, but it shows equity had already written off the incremental value of the contract while the lease and debt obligations remained. [54] [55]

The fiber analog applies to the financing, not to the scrap value

The late-1990s telecom buildout is the right mechanism and the wrong residual asset. After the 1996 Act, carriers laid on the order of 80 million miles of fiber, funded by debt, IPOs, and capacity swaps that inflated reported demand. A widely cited forecast that traffic would double every 100 days was far too high. By the early 2000s, a large majority of the fiber was dark—often cited below 5% lit—bandwidth prices fell as much as 90%, and Global Crossing, WorldCom, and a long list of regional builders went through bankruptcy or near-total equity loss. The physical plant was later bought cheaply and became the backbone of cloud computing. The winners were the second owners, not the original equity and unsecured creditors. [56] [57] [58] [59]

Lessons that transfer:

• Simultaneous rational build by every well-funded player produces collective oversupply even when the technology is real.
• Vendor financing and circular commitments (capacity swaps then; Nvidia–OpenAI–Oracle–Nvidia now) make demand look more solid than cash.
• Debt-funded builders with a single story customer fail before cash-rich incumbents do. Oracle is the name in this cycle that most resembles a 1999 carrier: negative free cash flow, rising leverage, backlog concentrated in one counterparty that itself must keep raising.
• Forecast error does not have to be total. A two- or three-year slip in utilization, or a 50% price cut on a fixed silicon generation, is enough to break covenants and takeout financing.

Lessons that do not transfer cleanly:

• Dark fiber retained decades of useful life and was repriced, not scrapped. GPUs and the power gear sized to a given generation do not. Kedrosky’s distinction is the one that matters: railroads and fiber were long-lived after the bust; this plant depreciates through the bust. [18] [60]
• Most of today’s capex at Microsoft, Amazon, Google, and Meta is still funded by operating cash flow. Oracle is the exception, which is why its CDS, not Meta’s, has become the market’s AI-credit proxy. [61] [62]
• Power, permits, and transformers are binding in a way fiber-laying was not. That can prevent a pure glut—or it can strand half-built campuses whose debt has already been drawn. Both outcomes impair the original sponsor.

The bull case, stated at full strength

The operating data are not fictional. Cloud infrastructure revenue rose 121% to $7.4 billion in the August quarter. Oracle said it delivered 850 MW and more than 300,000 GPUs, with GPU utilization around 97.5–97.9% and renewals at about a 20% premium. RPO reached $664 billion, with management expecting about half to convert within 36 months and about 13% within 12 months. More than $30 billion of new AI contracts in the quarter, and about $75 billion earlier, were structured as prepayments or bring-your-own-hardware and, management said, required no incremental Oracle capital and no margin degradation versus other contracts. Q1 free-cash-flow burn of $5.4 billion beat a roughly $9.6 billion consensus loss. Co-CEO Clay Magouyrk has said unused capacity is reallocated “very quickly.” Named anchors beyond OpenAI include Meta, xAI, Nvidia, AMD, and TikTok; a reported Tencent capacity deal would add another prepaid buyer. [63] [64] [65] [66]

Deutsche Bank has argued that even a scenario with no incremental OpenAI revenue would cut long-range EPS by a bounded amount and still leave substantial free cash flow, so the stock already embeds little credit for the relationship. Equity targets in October 2026 still spanned roughly $110 to $400, with a Street average well above the price. The investment-grade rating is intact, the outlook at S&P is stable, and an SPV chip lease could slow the pace of unsecured issuance. Core database cash flow has not disappeared, even if software revenue dipped. [67] [7]

Those points answer a different question than the bear case. High utilization today is evidence of a shortage, not of 15-year pricing power. Prepayments improve near-term cash and simultaneously increase dependence on the same customers’ continued ability to advance funds—the point Burry labels “the customer is now the lender.” A 20% renewal premium can reverse if token prices keep falling faster than volume grows. And “no incremental capex” contracts still sit inside a company that has already signed $288 billion of leases it must commence or exit.

How impairment actually shows up

The strongest failure path is not a dramatic Oracle default in 2026. It is a sequence:

1. Power or permit slips push 2027–28 revenue out of the underwriting window while 2028 construction-loan maturities and lease commencements arrive together.
2. OpenAI’s next raise comes in smaller, later, or with strings, and the five-year take-or-pay is resized. Oracle keeps the leases.
3. CDS and new-issue spreads gap further; an SPV chip deal prices at a cost that signals the unsecured market is closed on acceptable terms.
4. S&P or Moody’s cuts through BBB−/Baa3. Index exclusion forces selling into a book already trading like low-BBB or high-BB.
5. GPU rental rates on the installed generation fall as Rubin-class supply and cheaper models reset the clearing price, so re-tenanting the buildings does not cover rent plus refresh capex.

Any two of those, not all five, are enough to force a cut to the 2030 revenue ambition, equity dilution at lower prices, and a rump cloud business whose returns sit below the cost of the debt and leases used to build it. That is material impairment even if the lights eventually come on and someone else owns the shells.

The historical lesson to keep is not “infrastructure always wins.” It is that the first owners of a debt-financed build, matched to a demand forecast and a customer who must keep raising, are the ones who discover whether the forecast was early, wrong, or merely shorter than the lease.

---

Recent Findings Supplement (October 2026)

Oracle’s AI buildout shows mounting execution and funding stresses in October 2026, centered on OpenAI concentration, power delays, negative free cash flow amid heavy debt issuance, and hardware economics.[1][2]

Recent disclosures and analyst commentary (primarily September–October 2026) highlight specific mechanisms of impairment risk that were not fully visible earlier in the year.

OpenAI Concentration and Funding Ability

Oracle’s $664 billion remaining performance obligations (RPO) as of August 31, 2026, include roughly $300 billion tied to a five-year OpenAI cloud contract (announced 2025, ~45% of backlog).[3][4] Early October 2026 reports indicated OpenAI’s annualized revenue run-rate at approximately $50 billion—about $20 billion below prior late-September estimates—raising questions about its capacity to fund the full commitment.[3][5]

• DA Davidson’s Gil Luria described the OpenAI contract as a “fantasy,” assigning negligible probability to the full $300 billion materializing and warning Oracle might need to scale back its revenue ambitions (e.g., quadrupling by 2030).[6][6]
• OpenAI continues fundraising talks (e.g., $30 billion at ~$1.4 trillion pre-money valuation reported late September 2026), but the revenue shortfall directly pressures the largest single offtake agreement anchoring Oracle’s infrastructure spend.[7]

Implication: A scaled-back or delayed OpenAI spend would leave Oracle with long-duration lease obligations (15–19 years) against shorter or uncertain revenue, amplifying refinancing and utilization risks.

Power and Construction Delays

On September 24, 2026, Oracle issued a force-majeure notice for Project Jupiter, its 1.4–2.45 GW New Mexico campus (via Stack Infrastructure/Blue Owl) intended to serve OpenAI. The notice cited delays in a gas pipeline (pushed back ~6 months by regulatory denials and pending air-quality permits).[8][9][10]

• Morgan Stanley’s credit team noted that power bottlenecks convert construction delays into revenue, cash-flow, and credit-timing problems; delayed GPU purchases provide near-term capex relief but risk bunching spending as lease liabilities and maturities approach.[2]
• Alternatives such as trucked compressed natural gas are significantly more expensive (roughly 4x hub prices in some analyses), eroding margins on AI chip rentals.[11]
• Ares Management’s Fall 2026 newsletter flagged force-majeure invocations by offtakers as a key risk in the ~$573 billion web of AI infrastructure debt/leases involving Oracle and peers.[8]

Implication: Physical and regulatory constraints on power supply create a hard limit that capital alone cannot accelerate, directly impairing the timeline for monetizing the backlog.

Balance Sheet, Free Cash Flow, and Financing Pressures

Oracle reported negative free cash flow of $23.7 billion for fiscal 2026 and $5.4 billion in Q1 fiscal 2027 (ended August 2026), with quarterly capex at $28.5 billion (guidance $90–95 billion for FY27, net cash capex ~$70 billion).[12][13] Long-term debt has nearly doubled to more than $160 billion over two years; net debt figures range from ~$88–125 billion depending on the source.[14][4]

• S&P downgraded Oracle to BBB− (one notch above junk) in July 2026, citing weak cash flow and rising business risk; some 20-year bonds yielded 8.1% (~2.5 points over Treasuries), with CDS at record wides.[15][12]
• October 2026 reports detailed ongoing talks with Apollo and Goldman Sachs for off-balance-sheet structures to finance GPU purchases via a separate entity that Oracle would lease from, aiming to avoid direct additional borrowing.[4][5]
• S&P projects FY27 FCF deficit could widen to $42 billion; lawmakers (including Sen. Elizabeth Warren) and Ares have highlighted systemic risks from SPV/lease financing structures that obscure full exposure.[12][16]

Implication: Continued negative FCF funded by debt and equity raises (another ~$40 billion planned for FY27, including equity programs) increases leverage and dilution risk, with limited room before ratings or covenant pressure intensifies.

GPU Depreciation, Obsolescence, and Pricing Dynamics

Michael Burry’s early October 2026 analysis estimated Oracle’s earnings could be overstated by ~26.9% by 2028 due to extended useful-life assumptions for servers (hyperscalers shifting from 2–3 to 5–5.5+ years, with Meta citing a $2.9 billion annual depreciation reduction from one such change).[8] GPUs can lose 50–60% of rental value within two years amid rapid Nvidia cycles (Hopper → Blackwell → Rubin).[8]

• Oracle’s long-term data-center leases (15–19 years) require hardware replacement multiple times; fixed-margin structures in major AI cloud deals prevent Oracle from capturing recent GPU rental price increases (CoreWeave raised prices 25% in July 2026 plus another 10%).[17][15]
• Gross margins have compressed (from 79.1% in FY22 to 65.8% in FY26) while debt rose sharply.[17]

Implication: Depreciation and obsolescence create a mismatch between durable lease liabilities and perishable hardware economics, potentially leading to write-downs or margin pressure not fully reflected in current accounting.

Bear-Case Analyst Views, Systemic Risks, and Historical Analogs

Paul Kedrosky (early October 2026) singled out Oracle as among the most vulnerable, citing declining legacy database revenue (software down 3% YoY in one recent quarter), heavy borrowing to bridge to an AI future, and convergence of four bubble forces (technology, real estate, loose credit, government policy). He warned of catastrophic failure if infrastructure loses value post-boom.[1][15]

• Burry highlighted ~$3 trillion in aggregate future commitments across Oracle and peers, comparing dynamics to the 1960s computer-leasing bubble and noting capital-cycle peaks.[18]
• Morgan Stanley credit and Ares reports emphasize execution/force-majeure risks in the interconnected debt web.[2][8]
• Historical parallels (telecom/fiber buildout, railroads) are invoked by Kedrosky: durable physical assets retained value post-crash, but GPUs may not.[15]

Bull counterpoints (for balance): OCI revenue grew 121% YoY to $7.4 billion in Q1 FY27 with high utilization (97.9%) and premium renewals; many new contracts use customer prepayments or bring-your-own-hardware, reducing Oracle’s incremental capex; some analysts maintain high targets (e.g., Stifel $200, Bernstein $325) and note backlog conversion potential.[19][3]

These October 2026 developments—particularly the OpenAI revenue adjustment, force-majeure notice, financing structures, and prominent bear commentary—add concrete, time-stamped data points to impairment risks that were more theoretical earlier in 2026. Continued monitoring of Q2 FY27 results, debt issuance outcomes, power timelines, and OpenAI funding execution will be decisive.