Oracle Debt and the OpenAI Contract: Leases, Loans and Ratings in 2026

Oracle has never put a customer name or a $300 billion figure on the OpenAI contract in an SEC filing. The number the market treats as fact is a press report that OpenAI later echoed, while Oracle’s own books show only an unnamed contract worth more than $30 billion a year starting in fiscal 2028, sitting inside a $664 billion backlog that management will not split by customer.

What is actually disclosed versus what is inferred

The only Oracle filing that sizes the economics is the June 30, 2025 8-K. Safra Catz said Oracle had signed multiple large cloud services agreements, “including one that is expected to contribute more than $30 billion in annual revenue starting in FY28,” and warned that amounts ultimately recognized “may vary.” The customer was not named. [1]

Ten weeks later, the Wall Street Journal reported that OpenAI had signed a contract to buy about $300 billion of computing power from Oracle over roughly five years, starting in 2027, requiring 4.5 gigawatts, and that a majority of the new contracted revenue Oracle had just disclosed came from that deal. Oracle declined to comment. [2] [3]

OpenAI then did what Oracle would not. On September 23, 2025, OpenAI wrote that the July agreement to develop up to 4.5 gigawatts of additional Stargate capacity “represents a partnership that exceeds $300 billion between the two companies over the next five years.” That is a counterparty disclosure, not an Oracle RPO line item. Clay Magouyrk is quoted in the same post on delivery, not on dollars. [4]

The two official numbers are reconcilable only if the contract ramps. A flat $30 billion a year for five years is $150 billion. A $300 billion total over five years is a $60 billion average. The Journal’s account was that yearly revenue rises as sites come online. In October 2025 Magouyrk told CNBC that OpenAI can “of course” pay $60 billion a year — an affordability comment that accepts the higher run-rate framing without stating total contract value. [5]

“Stargate” is a brand, not the purchase contract. The January 21, 2025 announcement was a new company intending to invest $500 billion over four years, with SoftBank, OpenAI, Oracle, and MGX as initial equity funders, SoftBank holding financial responsibility and OpenAI operational responsibility. [6] The $300 billion figure is OpenAI’s purchase obligation to Oracle, not the joint venture’s equity check. Later reporting described the JV itself as having been sidelined as an operating company, with “Stargate” reused as an umbrella for compute deals. That is press characterization, not a restatement by Oracle. [7] [8]

Contract table

Contract	Counterparty	Reported size	Start / end	Source	Flag
Unnamed cloud services agreement	Not named in the filing; press later identified OpenAI	More than $30 billion of annual revenue	Starts in Oracle FY2028 (year beginning June 1, 2027). No end date in the filing	Oracle 8-K, June 30, 2025 [1]	Company disclosure of economics; customer is not a company disclosure
Stargate project announcement	SoftBank, OpenAI, Oracle, MGX (equity); Arm, Microsoft, NVIDIA as technology partners	Intended $500 billion of investment over four years, $100 billion to be deployed immediately; 10 GW goal	Announced Jan. 21, 2025; four-year horizon	OpenAI announcement [6]	Company/partner announcement of an intention, not a signed purchase obligation and not Oracle RPO
Additional Stargate capacity	OpenAI / Oracle	4.5 GW additional; with Abilene, over 5 GW under development, “over 2 million chips”; jobs estimate over 100,000	Announced July 22, 2025; no payment start date in this post	OpenAI [9]	Company (OpenAI) disclosure of capacity, not dollars
Same July partnership, dollarized	OpenAI / Oracle	“Exceeds $300 billion” over the next five years; Oracle-linked sites described as able to deliver over 5.5 GW (Shackelford County, TX; Doña Ana County, NM; Midwest later identified as Wisconsin with Vantage; plus up to 600 MW near Abilene)	Five years from the Sept. 23, 2025 framing; capacity phased	OpenAI, Sept. 23, 2025 [4]	OpenAI disclosure. Oracle has not repeated the dollar figure in a filing
Cloud purchase contract	OpenAI / Oracle	About $300 billion over roughly five years; 4.5 GW; majority of that quarter’s new Oracle backlog	Starts 2027; roughly five years	WSJ, Sept. 10, 2025, people familiar; Oracle declined comment [2] [3]	Press estimate. This is the figure markets use
Michigan campus inside the 4.5 GW partnership	OpenAI / Oracle (developer named in the post as building the site)	More than 1 GW; Stargate then described as over 8 GW and more than $450 billion of investment over three years	Announced Oct. 30, 2025; construction expected early 2026	OpenAI [10]	OpenAI capacity/investment claim. The $450 billion is a program total, not an Oracle contract
Meta cloud agreement	Meta	Existence confirmed; size not disclosed by Oracle. Talks reported at about $20 billion, multi-year	Timing not disclosed. Named as a contributor to the Q2 FY2026 RPO increase (quarter ended Nov. 30, 2025)	Existence: Magouyrk, Oct. 16, 2025, via CNBC [11]. Size: Reuters/Bloomberg, Sept. 2025, person familiar [12]. Name in backlog: Oracle Q2 FY2026 release [13]	Size is a press estimate of talks, not a closed-deal disclosure. No GW or term has been published by Oracle
Non-OpenAI booking wave	Four customers, seven contracts; Meta confirmed as one; the other three not named	$65 billion of new cloud infrastructure commitments in 30 days	Spoken Oct. 16, 2025, covering the then-current quarter	Magouyrk at Oracle AI World, reported by CNBC [11]	Company statement via a conference, not an SEC filing. Individual contract sizes not given
NVIDIA as an OCI customer	NVIDIA	No dollar size disclosed. Named alongside Meta as a source of new Q2 FY2026 commitments	Not disclosed	Oracle Q2 FY2026 release [13]; Catz named NVIDIA among AI cloud customers on the Sept. 9, 2025 call [14]	Relationship is a company disclosure. Size is not. Separate from Oracle’s purchases of NVIDIA chips
xAI cloud relationship	xAI	No current size disclosed. A potential ~$10 billion multi-year expansion was reported in May 2024; talks were reported ended in July 2024. Catz still listed xAI as a customer in Sept. 2025	Not disclosed	Ellison had previously confirmed a training deal without a value (Bloomberg, 2024). $10 billion figure and the lapse are press [15] [16] [14]	Current dollar commitment is not a company disclosure. Do not treat $10 billion as a live contract
AMD	AMD	No customer-contract size. Separately, OCI is to offer a 50,000-GPU MI450 supercluster starting calendar Q3 2026	Supplier deployment, not an AMD purchase of OCI	AMD/Oracle announcement, Oct. 14, 2025 [17]; Catz listed AMD as a cloud customer [14]	The 50,000-GPU item is a chip-supply partnership, not RPO from AMD as a buyer
Tencent GPU access	Tencent	About $7 billion over five years; ~100,000 accelerators in Southeast Asia; ~30% upfront	Reported Oct. 2026; five-year term	Financial Times, via Network World; neither company confirmed [18]	Press estimate. Unconfirmed
Customer-funded hardware inside large AI contracts	Multiple, unnamed	$75 billion of prepaid or customer-supplied GPU value inside large AI contracts	Cumulative as of May 31, 2026; most of the Q3 and Q4 FY2026 RPO increases were this structure	Oracle FY2026 Q4 release [19]	Company disclosure. Not allocated to OpenAI or anyone else
Data-center leases not yet commenced	Landlords, not cloud customers	$288 billion of additional lease commitments, substantially all data-center arrangements, terms of 15 to 19 years	Expected to commence between Q2 FY2027 and FY2029	Oracle 10-Q for the quarter ended Aug. 31, 2026 [20]	Company disclosure of Oracle’s obligation, not of customer revenue

How much of RPO is OpenAI

Oracle does not say. The August 31, 2026 10-Q states remaining performance obligations of $664 billion, of which about 13% is expected as revenue over the next twelve months, 37% in months 13–36, 34% in months 37–60, and the rest thereafter. It names no customer. [20] The FY2026 10-K says no single customer was 10% or more of revenue in fiscal 2026, 2025, or 2024. That sentence is about recognized revenue. The backlog that created the concentration had barely started to convert. [21]

The backlog path, all company-reported:

• May 31, 2025: about $138 billion.
• August 31, 2025: $455 billion, up $317 billion in one quarter. Catz: four multi-billion-dollar contracts with three customers. She also said Oracle had signed with “the who’s who of AI, including OpenAI, xAI, Meta, NVIDIA, AMD, and many others” — a customer list, not an allocation of the $317 billion. [22] [14]
• November 30, 2025: $523 billion, up $68 billion, “highlighted by new commitments from Meta, NVIDIA, and others.” [13]
• February 2026 quarter-end: $553 billion.
• May 31, 2026: $638 billion, up $85 billion. Most of the Q3 and Q4 increases were large AI contracts in which the customer prepaid the GPUs or supplied them; that hardware portion totaled $75 billion. [19]
• August 31, 2026: $664 billion. More than $30 billion of additional AI cloud contracts in the quarter, structured so they did not change Oracle’s capital plan. Q1 customer prepayments with a significant financing component were $11.4 billion. [23] [20]

Attribution is arithmetic on other people’s numbers, and the estimates disagree by vintage:

• At the $455 billion print, Journal sources said the majority of the new revenue was OpenAI. A sell-side note the next day treated RPO excluding OpenAI as $155 billion, which implies roughly $300 billion, or about two-thirds of the then-backlog. [2] [24]
• Commentary around the $523 billion and $553 billion prints still described OpenAI as more than half, or “more than $300 billion,” of RPO. The widely circulated “more than half of $523 billion” line is tied to the December 2025 quarter, not to the current book. [8] [25]
• Against the current $664 billion, if the reported contract is still about $300 billion and still almost entirely unearned — consistent with revenue starting in FY2028, and with only 13% of total RPO due inside twelve months — OpenAI is roughly 45%. That is the figure secondary write-ups were using in early October 2026. [26] It is a lower bound only if OpenAI’s own “exceeds $300 billion” language is taken literally and none of it has been recognized or cancelled. Oracle has not confirmed either assumption.

The diversification management points to is real but unallocated. In the 30 days before October 16, 2025, Magouyrk said Oracle contracted $65 billion across seven contracts from four customers, “none of those customers are OpenAI,” and confirmed Meta was one of the four. [11] That $65 billion, plus the later unnamed AI bookings, is why OpenAI’s share of the book should have fallen from the roughly two-thirds implied in September 2025 toward something closer to half or less — if the $300 billion figure did not itself grow. The filings do not let an outsider prove that.

Why the duration mismatch is the actual risk

Oracle’s customer contracts and Oracle’s buildings do not have the same clock. As of August 31, 2026, additional lease commitments not yet on the balance sheet were $288 billion, substantially all data-center arrangements, generally 15 to 19 years, expected to start between the second quarter of fiscal 2027 and fiscal 2029. [20] The OpenAI purchase obligation, as reported, runs about five years and does not start in size until 2027 or FY2028. Oracle does not own most of the shells. Abilene, the flagship Stargate site, is built and financed by Crusoe; Oracle furnishes and operates it. [27]

That is the mechanism behind the counterparty debate. If OpenAI pays, the leases are a funding bridge against contracted revenue. If OpenAI resizes or slips, Oracle is the tenant for 15–19 years against a customer whose obligation, on the reported terms, is much shorter. Oracle has said capacity can be reallocated quickly and that GPU utilization is extremely high; those are management claims about demand fungibility, not contractual take-or-pay terms published in a filing. KeyBanc’s September 2025 estimate that the OpenAI build could require on the order of $100 billion of new debt over four years was an analyst projection, and later contract structures (customer-supplied and prepaid GPUs) were explicitly designed to cut Oracle’s own chip funding. [28] [19]

What this means if you are underwriting the backlog

Treat $300 billion as OpenAI’s number and the Journal’s number, not Oracle’s. The auditable Oracle facts are: one unnamed contract above $30 billion a year from FY2028; RPO of $664 billion with no customer split; no customer at 10% of recognized revenue through May 2026; $75 billion of large-AI-contract hardware already funded by customers as of May 2026; and $288 billion of long data-center leases Oracle has committed to start. Anyone who publishes “OpenAI is 45% of RPO” or “more than half” is doing arithmetic on a press figure Oracle has never confirmed and has, when asked, declined to comment on. The competitive implication is that Oracle’s disclosed moat is capacity delivery and contract structure (prepay and bring-your-own-hardware), not a public, customer-level backlog that a rival or a creditor can diligence from the filings.

---

Recent Findings Supplement (October 2026)

Oracle has not disclosed customer-specific breakdowns of its Remaining Performance Obligations (RPO) in SEC filings or earnings releases. All attributions of portions of the $664 billion RPO (as of August 31, 2026) to OpenAI or others remain press or analyst estimates. Oracle filings and calls emphasize large AI infrastructure contracts (many with customer prepayments or bring-your-own-hardware structures) but name no counterparties or exact sizes for the bulk of the backlog.[1][2]

Recent developments (post-April 13, 2026) center on confirmation and expansion of the OpenAI relationship, infrastructure progress on Stargate sites, diversification signals via other named or reported customers (e.g., Meta, Tencent), and updated RPO/capacity delivery figures. No material changes to previously reported core terms (such as the ~$300 billion OpenAI figure) have emerged from company disclosures.

OpenAI Contract and Stargate Updates

The core ~$300 billion multi-year cloud compute agreement (reported as spanning roughly five years starting in 2027 and requiring ~4.5 GW of capacity) received prominent confirmation in a July 2026 Wall Street Journal report citing people familiar with the matter. Oracle had previously hinted at >$30 billion in annual revenue from a major cloud deal beginning in 2027. OpenAI has described the Oracle arrangement as part of its broader Stargate AI infrastructure platform.[3]

In an October 6, 2026 announcement, OpenAI detailed an agreement with Oracle for an additional 4.5 GW of Stargate data center capacity in the U.S. Combined with the Abilene, Texas flagship site, this brings planned Stargate capacity under development with Oracle to over 5 GW (supporting >2 million chips). This advances OpenAI’s White House-era commitment to U.S. AI infrastructure investment.[4]

Infrastructure progress (newer tracker and site updates): Abilene campus reached ~618 MW accepted by Oracle as of late August 2026 (with ~75% of campus capacity delivered per a September 2026 tracker update); GPT-6 Astra was pretrained on >100,000 Grace Blackwell GPUs there. Construction began on the Saline Township, Michigan Stargate campus in June 2026 (targeting ~1 GW).[5][6]

Oracle expanded the commercial relationship in its Q1 FY2027 earnings period (reported September 2026) to include OpenAI API access, ChatGPT for work, and related models via the Oracle Marketplace.[7]

RPO, Capacity Delivery, and Attribution

Oracle’s Q1 FY2027 results (quarter ended August 31, 2026; reported September 10, 2026) showed RPO rising to $664 billion (+$209 billion YoY; +$26 billion QoQ), with >$30 billion in new AI cloud contracts booked (vast majority structured as prepay or bring-your-own-hardware, requiring no incremental Oracle capital). Cloud infrastructure revenue grew 121% YoY to $7.4 billion. The company delivered 850 MW of incremental data center capacity and >300,000 GPUs to AI customers in the quarter (nearly 3x Q4 FY2026 delivery).[1][8]

An October 2026 analyst compilation estimated that the reported $300 billion OpenAI agreement accounts for roughly 45% of the $664 billion RPO. Oracle provides no official split or customer concentration details beyond general risk-factor language about concentration among a few large customers; no single customer reached 10% of FY2026 revenue.[9][2]

Prior-quarter context (Q4 FY2026, reported June 2026): RPO reached $638 billion, with most sequential growth from large-scale AI contracts involving prepayments or customer-supplied GPUs (totaling $75 billion across recent periods).[10]

Other Large AI Customers and Deals

• Meta: References to a multi-year AI cloud deal (reported ~$20 billion in 2025 discussions) persist in recent coverage. Oracle executives have verbally highlighted diversification, including a prior statement on $65 billion in new commitments across four customers (none OpenAI, with Meta named). No new sizing or official Oracle confirmation in 2026 filings.[11][11]
• Tencent: New October 2026 reports (FT/Reuters, citing people familiar) detail a five-year, ~$7 billion lease for access to ~100,000 advanced AI chips across Oracle data centers in Southeast Asia (~30% upfront payment). This is Tencent’s largest overseas cloud commitment with a U.S. provider and helps navigate export restrictions.[12][13]
• xAI and others (Nvidia-linked, etc.): Mentioned as anchor OCI/AI customers with ongoing demand (e.g., additional GPUs for Grok models). No specific new contract sizes reported. Broader RPO growth includes deals with “various AI developers.”[14]

Contract Summary Table (Reported Terms, Post-April 2026 Sources Only)

Note: All specific dollar sizes and attributions beyond aggregate Oracle disclosures are press/analyst estimates or third-party reports (flagged below). Oracle filings name no counterparties for the bulk of RPO growth. Dates reflect reported signing/announcement or revenue start where available.

• OpenAI Cloud Compute (Stargate-related): Counterparty OpenAI; Reported size ~$300 billion over ~5 years (starting ~2027); Capacity 4.5 GW; Source WSJ July 28, 2026 (people familiar with the matter) — press report/estimate.[3]
• OpenAI Stargate Additional Capacity: Counterparty OpenAI (via Oracle partnership); Reported size Part of broader Stargate (no incremental $ beyond above); Capacity Additional 4.5 GW (total >5 GW with Abilene); Source OpenAI company announcement, October 6, 2026.[4]
• Meta AI Cloud Deal: Counterparty Meta; Reported size ~$20 billion (multi-year); Capacity Undisclosed; Source Reuters 2025 discussions + recent references (people familiar) — press estimate.[11]
• Tencent AI Chip Lease: Counterparty Tencent; Reported size ~$7 billion over 5 years (~30% upfront); Capacity ~100,000 advanced AI chips (SE Asia Oracle DCs); Source FT/Reuters October 2026 (people familiar) — press estimate.[12]
• Aggregate Oracle RPO (AI-heavy): Counterparty Various (incl. OpenAI, Meta, Nvidia, xAI, others); Reported size $664 billion (end Aug 31, 2026); Capacity Significant GW-scale AI infra (e.g., 850 MW + 300k+ GPUs delivered Q1 FY2027); Source Oracle Q1 FY2027 earnings/8-K, September 10, 2026 (company disclosure; no customer split).[1]

Implications for competitors or entrants: Oracle’s model relies heavily on customer-funded or prepaid structures to mitigate capex intensity, with high utilization (~98%) and renewal premiums supporting margins. However, concentration risk (especially OpenAI exposure per estimates) and long-duration lease commitments versus shorter customer contracts remain investor concerns in recent coverage. Diversification via deals like Tencent and Meta references provides some offset, but lack of granular disclosure limits precise benchmarking. New capacity delivery and Stargate expansions signal execution momentum but also highlight permitting/power and financing challenges.

The $50B figure traces to an FT report and a CNBC confirmation. Next I’ll pull the primary pieces and the credit commentary on Oracle’s backlog.
I have the headline numbers. I’m opening the primary articles and the credit commentary that ties the gap to Oracle’s backlog.
One more pass on the cash-burn presentation and whether either company actually pushed back on the $50 billion figure.
The October 8 reports did not show OpenAI losing $20 billion of sales. They showed that the late-September “nearly $70 billion” figure was mostly an investor-built gross comparison with Anthropic, while OpenAI’s own investor presentation put net annualized revenue near $50 billion at the end of September. That distinction did not stop the market from marking down Oracle, because Oracle’s backlog is a claim on OpenAI’s cash, not on a partner-grossed headline.

What was actually reported, and by whom

The Financial Times broke it on October 8, 2026, in “OpenAI annualised revenues $20bn less than previously signalled.” Citing financial documents shared with investors, the FT said OpenAI had recently told investors the figure was nearing $50 billion in September, about $20 billion below the nearly $70 billion widely reported the week before. [1] [2]

CNBC confirmed the same day with a person familiar with an investor presentation: roughly $50 billion in annualized revenue at the end of September, 77% growth in the total run rate during the third quarter, and 107% growth in the enterprise run rate. That person said the roughly $68 billion figure that had circulated included gross revenue from OpenAI’s partners, so investors could compare OpenAI more directly with Anthropic. [3] Reuters, also citing a person familiar, said September annualized revenue was almost $50 billion and that the gap “mainly arose from an attempt to produce a direct comparison” with Anthropic, which includes sales through cloud partners such as Amazon Web Services and Google Cloud. OpenAI does not. [4] Bloomberg reported the roughly $50 billion pace the same day. [5]

Annualized revenue here is a run rate: a recent period’s sales stretched to a year. It is not cash collected, and it is not full-year 2026 revenue. The FT had already reported, on September 18, that OpenAI’s own projections put 2026 booked revenue at about $36 billion, rising to $350 billion only in 2030. [6] [7] A separate account put expected 2026 revenue near $35 billion, about half the year-end run rate the company was aiming for. [8]

What the $70 billion figure was built on

The higher number entered the market on September 29, not as an audited result. Axios, citing sources familiar with the financials, said annualized recurring revenue was nearing $70 billion, up more than 70% since the start of the third quarter, with business-to-business revenue up more than 100% and third-quarter consumer revenue larger than all of 2025. [9] Reuters matched that with its own source the same day. Oracle rose 5.3% on the report. Gil Luria of D.A. Davidson said OpenAI was about half of Oracle’s compute backlog, and that faster OpenAI growth reinforced Oracle’s ability to fill contracted capacity. [10]

The October correction describes how that $70 billion was assembled. The FT’s account, repeated by TechCrunch and Dow Jones, is that it came from “attempts by OpenAI’s own investors to produce a direct comparison with Anthropic’s annualised revenues.” Anthropic counts sales made through cloud partners. OpenAI books only its own share. [11] [2] Bloomberg’s October 9 follow-up put the same split more sharply: Anthropic counts the full gross value of sales through partners such as Amazon; OpenAI books only its cut of proceeds from partners such as Microsoft. [12]

The arithmetic, as secondary write-ups of the FT described it, was a base error compounded by a real growth rate. Investors had restated an earlier OpenAI figure onto Anthropic’s gross method, producing reports of about $40 billion in August. OpenAI then told investors revenue had grown more than 70%. Applied to the inflated base, that growth produced roughly $70 billion. OpenAI’s own presentation put the September pace near $50 billion. One account of the same documents said the company’s own July run rate was about $30 billion, not the gross-adjusted $40 billion. [13] [14] CNBC’s source framed the circulated number as about $68 billion of partner-grossed revenue, not a company forecast. [3]

The next day Bloomberg reported a different $70 billion: a year-end target, not a September result. People familiar said OpenAI expects to reach or exceed $70 billion in annualized revenue by the end of 2026, driven largely by enterprise, from the roughly $50 billion September base. [15] [16] On a like-for-like net basis, that is about a 40% increase in the run rate in one quarter, not $20 billion of revenue already earned. [17]

Shay Boloor of Futurum Equities read the gap as channel economics, not lost customers: the $20 billion is the value of sales flowing through AWS, Google, and Azure. [18] That is why the two numbers can both be “real” and still not be comparable. Gross recognition records the customer’s full payment and the cloud partner’s cut as an expense. Net recognition records only OpenAI’s share. Anthropic has argued it is the principal in those marketplace sales, so gross is the right book. OpenAI has historically kept partner sales out of its own headline. Both treatments can be GAAP-compliant. They are not the same denominator for a valuation multiple or for a supplier’s credit model.

Neither company disputed the $50 billion figure

OpenAI is the source of the lower number, not a critic of it. It declined to comment to the Financial Times. [2] [13] It did not respond to Reuters. [4] It declined to comment to Bloomberg on the year-end projection. [12] TechCrunch said it had asked for comment; no on-the-record rebuttal appeared in that report. [11]

What OpenAI did not do, on the public record, was deny the late-September $70 billion stories while they were circulating. A write-up of the FT said the company did not deny those reports at the time, then later showed investors a presentation with the net figure near $50 billion. [14] That is silence, not an endorsement.

No Oracle statement disputing the $50 billion figure turned up in the October 8–13 coverage. Oracle’s response was the share price and a parallel financing story, not a press release. The stock closed at $135.69 on October 8, down 5.48%, on more than twice the prior day’s volume. Nvidia fell about 3% and CoreWeave nearly 8%. [3] [19] The same day, the Wall Street Journal reported Oracle was in talks with Apollo and Goldman Sachs on a structure in which outside investors fund a vehicle that buys chips and leases them to Oracle. [20] [19]

Why a definitional gap still hits Oracle’s ability to pay and the RPO

Oracle’s remaining performance obligations were $664 billion at the end of August. The five-year cloud agreement with OpenAI, reported in September 2025 at about $300 billion, is roughly 45% of that backlog. [19] [21] CFO Hilary Maxson has said Oracle expects about half of RPO to convert to revenue over the next 36 months. [22] [21] That conversion schedule is an accounting expectation about signed contracts. It is not evidence the customer can fund the invoices.

S&P had already made that the credit thesis, before this revenue episode. On July 9, 2026, S&P cut Oracle to BBB-, one notch above junk, and called OpenAI a key credit risk. It estimated OpenAI was roughly half of the then-$638 billion RPO. If OpenAI could not pay, S&P said, Oracle could be left with data-center leases it might be unable to exit, or could re-lease only on worse terms. OpenAI’s ability to meet the contracts, S&P wrote, depends on continued AI demand, model leadership, and external financing. [23] [24]

Gil Luria’s standing view is harsher than a 20% haircut to a run rate. In a Bloomberg column, he called the OpenAI contract “fantasy” and said the likelihood OpenAI will have $300 billion to spend with Oracle is “negligible at this point.” He made that comment before Oracle’s earnings, so it is not a direct reaction to the October 8 restatement. It is the credit frame the restatement dropped into. [25]

The mechanism is a mismatch of clocks. A $300 billion, five-year commitment is on the order of $60 billion a year of compute purchases once it is fully ramping, against a company whose own projections put 2026 booked revenue near $36 billion and whose September net run rate is about $50 billion. Even the year-end $70 billion target, if hit on OpenAI’s net basis, is a pace of sales, not cash available after compute costs. The FT’s September presentation showed negative free cash flow of $278 billion from 2026 through 2030, and about $856 billion of compute and infrastructure spending by the end of 2030, with the March funding on track to be exhausted in 2028. [6] [7] Partner-grossed revenue does not pay Oracle. Net cash and fresh capital do.

That is why credit researchers treat RPO credibility as a funding question, not a bookings question. Oracle has been trying to shrink the balance-sheet piece of the risk. On the fiscal first-quarter call, management said the vast majority of new contracts came through prepay, bring-your-own-hardware, or similar structures that do not require incremental Oracle cash. [26] The Apollo-Goldman talks are the same idea applied to chips: move the purchase into a vehicle Oracle leases from, so the debt does not all sit on Oracle’s books. [20] That changes who holds the risk. It does not shrink the capital the build-out consumes. [19]

The balance sheet is already stretched for an investment-grade name. Bloomberg puts long-term debt above $160 billion, nearly double two years earlier, which would make Oracle the fifth-largest borrower in the U.S. corporate bond market. Capital spending in the June-to-August quarter was $28.5 billion, against $2.3 billion in the same quarter two years earlier. [20] [19] Credit-default-swap prints around early October were widely described as stressed, but the published levels do not agree: one account put five-year CDS near 261 basis points and at a record around the revenue headlines; another put early-October spreads near 151 basis points, a level last seen in 2009. [27] [28] [29] The direction is the usable fact. The precise spread is not.

The bull reply, made by some Oracle holders rather than by credit analysts, is resale. If OpenAI does not take all the contracted capacity, the racks are generic enough to lease to other labs and enterprises, possibly at better rates, because the constraint is still power and GPUs, not OpenAI specifically. [30] S&P’s downgrade note is the counter: unlike Amazon, Google, and Microsoft, Oracle does not have a giant internal cloud workload to absorb a vacant campus, and its leases run longer than the OpenAI contract. [23] [25]

For someone underwriting Oracle, the October reports change the denominator, not the contract. A $50 billion net run rate makes the $300 billion commitment look further from self-funding than a $70 billion gross comparison did. Backlog conversion over 36 months remains credible only if OpenAI keeps raising, or if Oracle can re-tenant the capacity without a loss. Signed RPO is not the same as collected revenue.

Fundraising capacity, and why the multiple moved even if demand did not

OpenAI closed a $122 billion round in March at an $852 billion post-money valuation. CFO Sarah Friar told CNBC, in the week before the correction, that the company was still “very well capitalized.” [3] It is also back in the market. CNBC and Bloomberg describe early talks for about $30 billion or more, driven by investor demand, with no term sheet finalized, at a pre-money valuation around $1.4 trillion. The IPO has been pushed to 2027. [3] [12] [12]

The $50 billion figure changes the ask, not the cash need. At $1.4 trillion pre-money, $50 billion of net annualized revenue is about 28 times sales. The same valuation on the $70 billion figure was about 20 times. The two bases are not comparable, which is the point: anyone who underwrote the round on the gross comparison was using a number OpenAI does not book. [13]

The cash need is set by the burn, not by the run rate. The September investor presentation the FT saw projects $278 billion of negative free cash flow from 2026 to 2030, revenue of $36 billion this year scaling to $350 billion in 2030, and $856 billion of compute spending by 2030. On that path, the March raise is exhausted in 2028. [6] [7] A $30 billion round covers a slice of one year’s infrastructure bill. It does not fund a $300 billion Oracle commitment or an $856 billion compute plan. Those get funded by repeated raises, customer prepayments, supplier financing, and, eventually, an IPO the company has already delayed.

Capacity to raise is still there, but the October episode showed the terms. Sovereign and strategic money has already been willing to fund losses: Amazon completed a $50 billion investment, including a contingent tranche, for roughly a 5% stake. [31] The constraint is not whether a check can be written. It is whether the next check clears at $1.4 trillion once investors insist on OpenAI’s net number, and whether that check arrives before Oracle, Broadcom, and the neoclouds have to decide how much of OpenAI’s rent they will carry themselves. Broadcom, the same week, was reported to be arranging more than $50 billion of debt so OpenAI can buy Broadcom chips, with Apollo and Blackstone among lenders approached. [32] The revenue correction and the financing circular are the same story: the customer’s sales do not yet cover the infrastructure, so suppliers and their lenders are being asked to bridge the gap.

The competitive implication is narrower than “OpenAI missed.” Anthropic’s $65 billion July run rate is not an apples-to-apples lead if it is gross and OpenAI’s $50 billion is net. [3] [12] Anyone pricing a lab, a chip order, or a cloud backlog off headline ARR without asking who is principal on the marketplace sale will mis-size both the multiple and the credit. The number that matters for Oracle is the one that can be turned into cash after compute costs. On the documents OpenAI actually showed investors in late September, that number was near $50 billion annualized, against a compute plan measured in the hundreds of billions.

---

Recent Findings Supplement (October 2026)

Financial Times (FT) on October 8, 2026, first reported that OpenAI told investors its annualized revenue run rate was approaching $50 billion as of end-September 2026, roughly $20 billion below the ~$70 billion figure that had circulated in late September.[1][2] Bloomberg followed on October 9 confirming the $50 billion end-September figure while adding that OpenAI shared with investors an expectation to reach or exceed $70 billion by year-end 2026, driven primarily by enterprise growth, as part of discussions for a $30 billion+ fundraising round at a $1.4 trillion pre-money valuation.[3][4]

The earlier ~$70 billion figure originated in late September reports (Axios on Sept. 29, picked up by Reuters and others) citing sources familiar with the matter; it stemmed from investor attempts to align OpenAI’s metrics with rival Anthropic’s rather than from OpenAI itself.[5][6]

• FT and subsequent reporting (CNBC, Bloomberg) attribute the gap to methodology: Anthropic books the gross value of sales through cloud partners (e.g., AWS, Google Cloud), while OpenAI books only its net share (primarily via Microsoft). A person familiar with the documents stated the higher number “did not come from OpenAI.”[7][2]
• OpenAI provided supporting growth details in investor materials: 77% total run-rate growth in Q3 and 107% enterprise segment growth; it started 2026 at ~$20 billion run rate.[8]
• OpenAI declined to comment on specifics in multiple reports; no public correction or dispute from the company.[4]

No direct public dispute or correction has come from OpenAI or Oracle. OpenAI has continued sharing the clarified figures and forward projection with investors during fundraising, and the process appears ongoing without reported disruption.[9]

Oracle’s exposure creates the most immediate market reaction and credit scrutiny. Oracle holds a reported five-year, ~$300 billion cloud/compute agreement with OpenAI (roughly 45% of its $664 billion remaining performance obligations or RPO/backlog as of late August/early September 2026).[10][10] Oracle shares fell ~5% on the October 8 FT report (and broader AI stocks including Nvidia declined), with analysts highlighting concentration risk—OpenAI representing around half of Oracle’s compute backlog per prior S&P commentary.[5]

• Gil Luria (D.A. Davidson) has called the OpenAI contract “fantasy” and assessed the likelihood of full $300 billion spend as “negligible,” citing OpenAI’s loss-making status and the scale of commitments relative to its revenue trajectory; this predates but aligns with post-report sentiment.[11][11]
• Oracle faces heavy capex ($90–95 billion guidance for FY2027) and has nearly doubled long-term debt to over $160 billion in recent years; it is reportedly in talks with Apollo and Goldman Sachs for SPV-style financing to fund chips while limiting direct borrowing.[12][12]
• RPO conversion remains a point of focus: ~13% expected in the next 12 months and roughly half within 36 months in recent disclosures, with some prepayments or customer-supplied hardware mitigating near-term cash needs but long-dated leases creating potential stranded-cost risk if OpenAI scales more slowly.[13]

The revenue clarification raises questions about Oracle’s backlog credibility and conversion assumptions without invalidating underlying demand. The $664 billion RPO is the largest in software and grew rapidly on AI contracts, but lacks public customer-level detail or cancellation terms; the OpenAI concentration amplifies sensitivity to any recalibration of the customer’s paying capacity.[14]

• Credit researchers and analysts note that while OpenAI’s growth remains steep (from ~$20 billion start-2026 to $50 billion), sustaining payments on a ~$60 billion annualized commitment requires continued rapid scaling plus external capital raises.[15]
• Oracle has pointed to diversification (e.g., Meta, Nvidia, AMD, xAI deals) and high GPU utilization (~97–98%), plus the ability to reallocate capacity, but investor reaction shows the OpenAI narrative still dominates sentiment.[11]

OpenAI’s own funding needs and capacity appear resilient but face higher scrutiny on multiples and trajectory. The company is seeking at least $30 billion in new capital at a $1.4 trillion pre-money valuation (implying a higher revenue multiple on the $50 billion base versus the prior $70 billion figure).[4]

• Private-market discussions continue, suggesting investors view the update as an accounting recalibration rather than a demand miss; enterprise momentum (107% Q3 growth) and the year-end $70 billion+ projection provide a forward bridge.[9]
• Broader implications include pressure on AI infrastructure economics: if application-layer revenue grows more slowly than infrastructure commitments, suppliers like Oracle face tighter margins or higher financing costs amid already elevated debt loads and capex.[16]

For competitors or new entrants in AI infrastructure/cloud: The episode underscores the value of diversified customer bases and transparent backlog disclosure to mitigate single-counterparty risk. Oracle’s experience shows how even accounting-aligned revenue figures can trigger sharp repricing when concentration is high; firms emphasizing net-revenue transparency, prepayment structures, or multi-tenant utilization may gain relative credibility. Long-term lease and power commitments remain a structural vulnerability if AI labs’ cash generation or fundraising pace slows.

Oracle is funding an AI data-center build that already outruns the cash the business produces. In the quarter ended August 31, 2026 (Q1 FY2027), operating cash flow hit a record $23.1 billion, but $11.4 billion of that was customer prepayments, and capital spending was $28.5 billion. Company free cash flow was therefore −$5.4 billion even after the strongest operating quarter Oracle has reported. Management’s full-year FY2027 plan keeps the same structure: reported capex of $90–95 billion, of which Oracle’s own net cash outlay is guided to no more than $70 billion, with the gap filled by prepayments, customer-supplied hardware, and external capital.

Fiscal year ends May 31. As of October 13, 2026, the latest reported period is Q1 FY2027 (ended August 31, 2026; released September 10, 2026). FY2026 (ended May 31, 2026) is the last completed year. “Current fiscal year” below is FY2027.

Capex: reported vs. net cash outlay

Oracle now guides two different capex numbers because a large share of the build is no longer funded by Oracle’s own cash. Reported capex is what hits the cash-flow statement. “Net cash outlay” subtracts customer prepayments with a significant financing component and certain short-term financing tied to capex. That second number is the cash Oracle itself expects to write.

• FY2026 reported capex: $55.66–55.7 billion, up from $21.2 billion in FY2025 and above the roughly $50 billion guidance then in force. FY2026 net cash outlay for capex was presented at about $48 billion.
• FY2027 company guidance (reaffirmed on the September 10, 2026 call): reported capex $90–95 billion; net cash capex not more than $70 billion. The $20–25 billion gap is customer prepayments and timing, not a cut in the physical build.
• Q1 FY2027 reported capex: $28.499 billion. Company supplemental math put net cash outlay for the quarter near $18.0 billion after $11.363 billion of customer prepayments with a financing component and an $830 million repayment of short-term financing related to capex.
• Analyst estimates: S&P Global Ratings, after the June guidance jump, forecasts a fiscal 2027 free operating cash flow deficit of nearly $42 billion (inclusive of prepayment benefits, aligned with the ~$70 billion net cash outlay). A pre-print sell-side snapshot on September 10, 2026 put the full-year FY2027 free-cash-flow loss near $45 billion. Q1 free cash flow of −$5.40 billion beat the LSEG consensus of about −$9.56 billion.

What this means for anyone modeling the stock: the $90–95 billion figure is the capacity Oracle is installing; the ≤$70 billion figure is the cash hole shareholders and lenders have to fill. If prepayments slip, reported capex does not fall—Oracle’s cash requirement rises.

Debt, cash, and the lease stack

On the balance sheet, borrowings fell slightly in Q1 because Oracle repaid debt and did not issue new bonds. The real leverage increase is in leases that have not started.

As of August 31, 2026 (Form 10-Q):

• Notes payable and other borrowings: current $7.625 billion + non-current $117.712 billion = $125.337 billion.
• Cash and cash equivalents $36.369 billion + marketable securities $708 million = $37.077 billion (MD&A).
• Company-style net debt (borrowings minus cash and marketable securities, leases excluded): $88.26 billion.
• Operating lease liabilities $34.621 billion (current $4.027 billion) + finance lease liabilities $9.185 billion (current $749 million) = $43.806 billion of recognized lease liabilities.
• Additional lease commitments not yet commenced: $288 billion, substantially all data-center leases expected to start between Q2 FY2027 and FY2029, with terms of 15–19 years, and not on the balance sheet. At May 31, 2026 that figure was about $260 billion.
• Undiscounted payments on leases already running: about $48.1 billion operating + $15.2 billion finance.
• Purchase obligations rose to about $34.2 billion at August 31, 2026, from $13.3 billion at May 31, 2026.

May 31, 2026 comparison: borrowings $129.54 billion ($7.199 billion current + $122.342 billion non-current); cash and marketable securities $31.89 billion; net debt before leases about $97.7 billion; recognized lease liabilities about $37.9 billion.

Borrowings + recognized leases + leases not yet started were about $457 billion at August 31, 2026, against remaining performance obligations of $664 billion. RPO is contracted future revenue, not cash already collected; the 10-Q expects about 13% of that $664 billion to be recognized within 12 months.

Six-quarter cash-flow trend

Oracle defines free cash flow as operating cash flow minus capital expenditures. That is the series below. Q1 FY2027 operating cash flow includes $11.4 billion of customer prepayments with a significant financing component, so the “record” operating figure is not all earnings converting to cash.

Quarter ended	Operating cash flow	Capex	Free cash flow	Leases not yet commenced (period end)
May 31, 2025 (Q4 FY2025)	~$6.2B	$9.08B	−$2.92B	$43.4B
Aug 31, 2025 (Q1 FY2026)	$8.140B	$8.502B	−$0.362B	$99.8B
Nov 30, 2025 (Q2 FY2026)	~$2.06B	$12.03B	−$9.97B	$248B
Feb 28, 2026 (Q3 FY2026)	~$7.16B	$18.64B	−$11.48B	$261B
May 31, 2026 (Q4 FY2026)	~$14.62B	$16.49B	−$1.87B	~$260B
Aug 31, 2026 (Q1 FY2027)	$23.103B	$28.499B	−$5.396B	$288B

Trailing twelve months to August 31, 2026: operating cash flow about $46.9 billion, capex $75.7 billion, free cash flow −$28.7 billion. Full FY2026: operating cash flow $31.98 billion, capex $55.66 billion, free cash flow −$23.7 billion (versus −$0.39 billion in FY2025).

The burn is not steadily worsening quarter by quarter—Q3 FY2026 was the trough at −$11.5 billion, and Q4 improved sharply—but every one of the last six quarters was free-cash-flow negative, and the full-year FY2027 plan implies a much larger annual deficit than FY2026.

Financing, dividends, and buybacks

Oracle has not completed a new public bond deal since February 2026. It has, however, already used the equity side of its 2026 funding plan, and press reports in early October 2026 said it was exploring off-balance-sheet chip financing.

• September 2025: $18.0 billion senior notes (six tranches, due 2030–2065).
• February 2–4, 2026: $25.0 billion fixed-rate notes across eight tranches (4.550% to 6.850%, due 2029–2066) plus $500 million floating-rate notes due 2029. Together with the September deal, fiscal 2026 senior-note issuance was about $43 billion of par.
• February 2026: $5.0 billion of 6.50% Series D mandatory convertible preferred stock (depositary shares trade as ORCL-PRD). Preferred carrying value was $4.954 billion at both May 31 and August 31, 2026. Preferred dividend in Q1 FY2027 was $81 million.
• Q1 FY2027: at-the-market common-stock program fully used—141 million shares, gross proceeds of $20 billion, net proceeds $19.909 billion.
• June 10, 2026 guidance: expect to raise about $40 billion of debt and equity in FY2027, including the already-announced $20 billion ATM (not on top of it), and do not expect additional debt issuance in calendar 2026. The ATM is done; the remaining ~$20 billion of that FY2027 plan had not been raised as new Oracle bonds as of the Q1 print.
• October 8, 2026 press: Oracle was reportedly in talks with Apollo and Goldman Sachs on a structure in which investors would fund a separate vehicle that buys chips Oracle would lease—explicitly to avoid more borrowing on Oracle’s own balance sheet. That is not a completed issuance.

Dividend: quarterly common dividend remains $0.50 per share ($2.00 annualized). It was raised from $0.40 to $0.50 in the dividend paid January 2025 and has not been increased since. Latest declaration (with Q1 results): payable October 23, 2026 to holders of record October 9, 2026. Preferred dividend of $1,625 per preferred share payable October 15, 2026. Q1 cash dividends to stockholders were $1.565 billion.

Buybacks: effectively paused. No repurchases in the three months ended August 31, 2026. About $6.3 billion of authorization remains, with no expiration, but the 10-Q says the pace depends on capex and other cash needs. FY2026 share repurchases were only about $0.1 billion, versus multi-billion-dollar programs earlier in the decade.

Company-reported vs. analyst-adjusted leverage

The gap is almost entirely leases that GAAP has not yet put on the balance sheet, plus how rating agencies treat preferred stock and prepayments.

Measure (Aug 31, 2026 unless noted)	Amount	What it includes
Gross borrowings	$125.3B	Notes and other borrowings only
Company-style net debt	$88.3B	Borrowings minus cash and marketable securities; excludes all leases
Recognized lease liabilities	$43.8B	Operating + finance leases already commenced
Net debt + recognized leases	~$132B	On-balance-sheet debt-like claims
Leases not yet commenced	$288B	Signed, mostly 15–19 year data-center leases; off balance sheet
Borrowings + all lease claims	~$457B	One filing-based stack used by independent analysis
S&P adjusted leverage (FY2027 forecast)	mid-4x	Debt adjusted to include the ~$260B of additional lease commitments (May 31 figure) and purchase obligations; rating BBB− (cut from BBB on July 9, 2026), stable outlook
Filing-based net debt / EBITDA, leases excluded (TTM)	~2.5–2.6x	Independent calculation from filed operating income, depreciation, and net debt before leases
Same source, net debt and recognized leases / EBITDA	~3.7x	Still excludes the $288B not-yet-commenced pile

S&P’s FY2027 free-operating-cash-flow deficit of nearly $42 billion is an analyst-adjusted view that already gives Oracle credit for customer prepayments, and it is still far worse than the −$23.7 billion company free cash flow of FY2026. Fitch has been cited at BBB, with a prior threshold that leverage above 3.5x EBITDA—or above 4x if rents are included—would pressure the rating. Moody’s has been cited at Baa2 with a negative outlook. Those agency ratios are not the same as Oracle’s borrowings-minus-cash figure, and they are the ones that govern index eligibility and borrowing cost if the rating slips below investment grade.

Implication: a competitor or creditor who stops at “net debt of $88 billion and 2.6x” is looking at the narrowest definition. The binding constraint is the combination of ≤$70 billion of FY2027 cash capex, $7.2 billion of debt principal due in FY2027, a dividend that still costs more than $6 billion a year at the current rate, and $288 billion of leases that start whether or not the contracted AI tenants pay on the schedule embedded in the $664 billion backlog.

---

Recent Findings Supplement (October 2026)

Oracle's aggressive AI/cloud infrastructure buildout has driven record capex far exceeding operating cash flow, funded primarily by equity raises, customer prepayments, and prior debt issuances, with negative FCF persisting but showing some moderation in the most recent quarter amid strong backlog growth.[1][2]

Recent data (primarily from Q4/FY2026 results released June 2026 and Q1/FY2027 results released September 10, 2026, plus related 10-Q/10-K filings and analyses) focus on the period after April 13, 2026. Oracle's fiscal year ends May 31 (FY2026: June 2025–May 2026; FY2027 ongoing). No significant new dividend or share buyback policy changes were reported.[3][4]

Capex (Reported and Guidance)

Oracle distinguishes between headline (reported) capex and "net cash outlay" after customer prepayments/financing effects. Guidance emphasizes the latter for funding needs.[5][6]

• FY2027 guidance (issued/updated in Q1 FY2027 reporting): Reported capex $90–95 billion; net cash outlay at most ~$70 billion (after ~$20–25 billion in expected customer prepayments). This is above prior analyst estimates (~$67–68 billion) and follows FY2026 actuals exceeding the company's ~$50 billion target.[1][2]
• Q1 FY2027 (ended Aug. 31, 2026): Reported capex $28.5 billion (already over half of full FY2026 total; ~51% of FY2026 spend in one quarter). Net cash outlay ~$18 billion after ~$11.4 billion customer prepayments.[2][4]
• FY2026: Reported capex $55.66–55.7 billion (up ~162% YoY from $21.2 billion in FY2025).[7][8]

Implications for competitors/entrants: The scale of spending (tied to AI demand outpacing supply) creates a high barrier; rivals must match capacity or differentiate via efficiency/partnerships, while Oracle's prepayment model reduces net cash needs but locks in contracted revenue.[9]

Operating Cash Flow (OCF), Free Cash Flow (FCF), and Quarterly Trends

OCF has risen sharply with prepayments and growth, but FCF remains deeply negative due to capex. TTM figures reflect the cumulative impact.[10][4]

• Q1 FY2027: OCF $23.1 billion (up ~184% YoY; includes $11.36 billion customer prepayments with financing component); FCF –$5.4 billion (better than analyst expectations of ~–$9.56 billion; improved from prior quarter's –$11.48 billion but still negative).[2][2]
• FY2026: OCF ~$32.0 billion (up ~54% YoY from ~$20.8 billion in FY2025); FCF –$23.7 billion (vs. –$0.394 billion in FY2025).[7][8]
• TTM to Q1 FY2027 (four quarters ended Aug. 31, 2026): FCF –$28.7 billion; capex $75.7 billion.[1]

Last six quarters trend (approximate quarterly figures drawn from earnings releases, 10-Q/analyses; quarters aligned to fiscal calendar; prepayments increasingly material in later periods):[4][1]

• Q1 FY2026 (ended ~Aug. 2025): Capex ~$8.5 billion; OCF ~$8.1 billion; FCF ~–$0.4 billion.
• Q2 FY2026 (~Nov. 2025): Capex ~$12.0 billion; OCF ~$2.1 billion (low); FCF ~–$10.0 billion (ex-prepayments).
• Q3 FY2026 (~Feb. 2026): Capex ~$18.6 billion; OCF ~$7.2 billion; FCF ~–$11.5 billion.
• Q4 FY2026 (~May 2026): Capex ~$16.5 billion; OCF ~$14.6 billion (incl. ~$4.6 billion prepayments); FCF ~–$1.9 billion.
• Q1 FY2027 (~Aug. 2026): Capex $28.5 billion; OCF $23.1 billion (incl. $11.4 billion prepayments); FCF –$5.4 billion.

Trend insight: Capex and negative FCF accelerated through FY2026 as AI buildout ramped (peaking in Q3 before partial moderation), with OCF boosted by prepayments in Q4 FY2026 and especially Q1 FY2027. TTM FCF worsened to –$28.7 billion. Prepayments mitigate but do not eliminate cash burn.[4]

Implications: Sustained negative FCF pressures balance sheet and ratings (S&P downgraded to BBB–); success depends on backlog conversion and utilization (reported at ~98% in Q1).[7]

Total Debt, Net Debt, and Lease Commitments

Debt has risen with prior issuances but showed slight sequential decline in Q1 FY2027 due to repayments and equity funding. Off-balance-sheet leases represent a massive future obligation.[11][1]

• Latest (Aug. 31, 2026): Total borrowings/notes payable ~$125.3 billion (down from $129.5 billion at May 31, 2026); lease liabilities on balance sheet ~$43.8 billion; net debt (before leases) ~$88 billion.[12][1]
• Off-balance-sheet lease commitments not yet commenced: $288 billion (primarily data center leases, 15–19 year terms, expected to commence Q2 FY2027–FY2029). Purchase obligations also rose sharply (~$34 billion).[11][4]
• Company-reported vs. analyst-adjusted (lease-inclusive) leverage: Reported metrics focus on on-balance-sheet debt (~$125 billion) plus leases (~$44 billion). Adjusted figures incorporating off-balance-sheet leases would show substantially higher leverage (potentially hundreds of billions in total obligations), contributing to credit concerns and the S&P downgrade.[13]

Implications: High lease commitments amplify long-term obligations beyond reported debt; competitors or new entrants face similar scaling challenges but without Oracle's contracted AI backlog cushion.[7]

Latest Bond Issuances, Equity/Convertible Raises, and Capital Plan

No new bond issuances in calendar 2026 beyond February; equity filled the gap.[6][14]

• Recent raises (post-April 2026 focus): $20 billion at-the-market (ATM) common stock issuance completed in Q1 FY2027 (~141 million shares, ~$19.9 billion net proceeds).[1][4]
• Prior (within scope): $18 billion senior notes (Sep. 2025, multiple tranches 4.45–6.10%); $25 billion senior notes + $500 million floater (Feb. 2026, tranches up to 6.85%); $5 billion 6.50% Series D mandatory convertible preferred (Feb. 2026).[14][1]
• FY2027 plan: ~$40 billion in debt + equity (includes the completed $20 billion equity); no additional debt expected in calendar 2026.[5]

No notable changes to dividends (ongoing payments) or buybacks (minimal/reduced activity noted in some periods).[4]

Table of Key Sources (selected recent primary/analytical sources; full list includes earnings releases, 10-Q/10-K, and analyses):

• Oracle Q1 FY2027 earnings/10-Q (Sep. 2026 filings and releases): Capex, OCF/FCF, debt, leases, RPO, guidance.[4]
• Oracle FY2026/Q4 earnings (Jun. 2026): Historical FY2026 figures and initial FY2027 outlook.[3]
• SEC analyses/filings summaries (e.g., sec-api.io Oct. 2026): Debt, leases, TTM aggregates, quarterly breakdowns.[1]
• News/analyst reports (WSJ, Reuters, Economic Times, etc., Sep.–Oct. 2026): Confirmation of figures, expectations vs. actuals, financing details.[2][5]

Differences between company-reported and adjusted leverage stem mainly from the $288 billion off-balance-sheet leases (not included in standard debt metrics but disclosed in footnotes). All monetary values are in USD. Data is current as of the latest October 2026 sources; further 10-Q details or updates could refine quarterly splits.

Oracle’s Stargate campuses are financed off the company’s balance sheet through developer special-purpose vehicles, with Oracle as the long-term tenant and OpenAI as the subtenant. The debt is large, short-tenor on the bank side, and increasingly hard to distribute. The only widely reported secondary-market level is on the New Mexico loan. Apollo is not a lender on these site financings.

How the structure works

The template is the same across sites. A developer (Blue Owl’s STACK, a Blue Owl–Crusoe joint venture, Vantage, or Related Digital) owns the campus through an SPV, borrows construction money, and leases the finished building to Oracle for roughly 15 years. Oracle’s lease payments are what service the debt. The loans themselves do not sit on Oracle’s balance sheet, which is why banks and insurers underwrote them against a lease rather than against Oracle’s corporate credit directly.

That separation is now the pressure point. Single-tenant and single-counterparty caps filled up as Oracle’s own rating slid to BBB-, one notch above junk, after an S&P cut in July 2026. Banks that arranged the loans at tight spreads cannot always resell them at par. Newer paper has had to clear at wider margins, and one campus was refinanced as a long project bond because banks would not hold another construction loan.

Site-by-site financing

Abilene, Texas — Crusoe, Blue Owl, Primary Digital Infrastructure. This is the original Stargate flagship, on the Lancium Clean Campus.

• January 2025: JPMorgan provided a $2.3 billion construction loan for the first 206 MW, two-building phase. Newmark arranged it. The sponsors had formed a $3.4 billion joint venture. Oracle had a long-term lease on the buildings. [web:17] [web:23] [web:21]
• May 2025: a JPMorgan-led consortium provided $7.1 billion for phase two — six more buildings, taking the campus to eight buildings and 1.2 GW — inside a $15 billion joint venture. Newmark arranged that loan as well. [web:15] [web:10]
• The Financial Times later described Blue Owl as majority owner, with about $3 billion of equity and roughly $10 billion borrowed from JPMorgan, repaid by Oracle’s 15-year lease. A person close to the deal put Blue Owl’s targeted return as high as 25 percent. The Wall Street Journal separately put Abilene construction loans at about $10 billion. JPMorgan has said it originated $9.6 billion across two Abilene construction loans as lead left, sole underwriter, and sole structuring agent. Those figures overlap; they are not additive. [web:9] [web:8] [web:95]

Public pricing on the Abilene loans was not disclosed. An adjacent expansion of roughly 600–700 MW was not financed under an Oracle lease. Bloomberg reported in March 2026 that Oracle and OpenAI walked away after talks dragged on financing and OpenAI’s changing needs; Microsoft then agreed to lease about 700 MW (a later Crusoe statement cited about 900 MW) from Crusoe, with the first building expected powered by mid-2027. Oracle said claims that planned capacity at the existing Abilene site was delayed were inaccurate, and a source told Reuters there were no changes to OpenAI’s existing Oracle agreements. This is a reported tenant substitution on an expansion, not a confirmed construction-loan default or force majeure. [web:125] [web:128] [web:132] [web:134]

Shackelford County, Texas (“Frontier”) and Port Washington, Wisconsin (“Lighthouse”) — Vantage. These are one financing package, not two unrelated loans.

• JPMorgan and Mitsubishi UFJ (MUFG) led a $38 billion debt package, split as about $23.25 billion for the Texas campus and $14.75 billion for Wisconsin, according to people familiar with the matter cited by Bloomberg in October 2025. Wells Fargo, BNP Paribas, Goldman Sachs, Sumitomo Mitsui Banking Corp. (SMBC), and Société Générale were allocated portions after the first underwriter round. [web:62] [web:11]
• Structure, as reported: senior secured, four-year maturity with two one-year extension options, priced around 2.5 percentage points over the benchmark (SOFR). Interest-only during construction, then amortizing once operations begin. [web:62] [web:3]
• The Financial Times reported in February 2026 that the $38 billion Texas/Wisconsin facilities and the $18 billion New Mexico campus — at least $56 billion in total — had received investment-grade ratings, and that more than a dozen banks had lent against Oracle’s lease at SOFR plus 2.5 percent. TD Cowen, in a January 26, 2026 note, said borrowing costs on newer, not-yet-sold Oracle-linked data-center debt had widened to 3 to 4.5 percentage points over SOFR. [web:3]
• Distribution was slow. Business Insider reported in January 2026 that JPMorgan had seen diminishing interest in the tail of the syndication, while a person familiar said the projects were fully financed. The Wall Street Journal reported in April 2026 that Vantage said the loans were largely syndicated in the fourth quarter of 2025, with more than 50 lenders, and were expected to close in the second quarter of 2026. The Financial Times reported in May 2026 that lenders including JPMorgan and MUFG had spent more than six months distributing the $38 billion, that some banks sought to sell pieces at a discount to non-bank buyers, and that banks were sounding out investors on significant-risk-transfer structures to move the riskiest slices. No public secondary-market quote (in cents on the dollar) was found for the Vantage loans specifically. [web:59] [web:8] [web:71]

Frontier is a Vantage campus of about 1.4 GW. Lighthouse is the Port Washington campus, commonly described as roughly 0.9–1.3 GW of load and about $15 billion of project investment, with Vantage guiding to full buildout in 2028. [web:55] [web:43]

Doña Ana County, New Mexico — Project Jupiter, STACK Infrastructure (Blue Owl) and BorderPlex. This is the loan that has actually traded, or at least been quoted, in stressed territory.

• Bloomberg reported on November 7, 2025 that about 20 banks were providing a roughly $18 billion project-finance loan. Administrative agents: SMBC, BNP Paribas, Goldman Sachs, and MUFG. Retail syndication commitments were due in late November 2025. [web:7]
• Reported terms: SOFR plus 2.5 percent, four-year maturity with two one-year extensions. [web:44] [web:3]
• Blue Owl equity is about $3 billion. A person familiar with the terms told Reuters that Blue Owl earns a lower return during construction, stepping up once the campus is complete. Other reporting puts the construction-stage yield at about 9 percent, rising to about 11 percent at completion. That is a different, and lower, figure than the “up to 25 percent” target reported for Abilene. [web:12] [web:36]
• The $165 billion figure attached to Jupiter is a local industrial-revenue-bond tax structure (power, buildings, and tenant equipment), not the construction loan. The cash project financing that has been reported is the $18 billion of bank debt plus roughly $3 billion of Blue Owl equity. [web:36] [web:101]
• Secondary market: the Financial Times, cited by Reuters on September 18, 2026, reported that syndicate banks including Santander and Jefferies were quoting the loans at 89 to 91 cents on the dollar, and that efforts to sell the debt had stalled, leaving those banks holding more than planned. The package had been investment-grade when arranged. No exchange print was published; these are private quotes. [web:67] [web:4]

Saline Township, Michigan (“The Barn”) — Related Digital. This is the deal Blue Owl did not do, and the one that had to be rebuilt as a bond.

• The Financial Times reported on December 17, 2025, citing three people familiar, that Blue Owl would not back a roughly $10 billion financing and equity investment for the 1 GW campus after negotiations stalled. Sources said lenders wanted stricter lease and debt terms because of Oracle’s debt load and AI spending, and that Blue Owl viewed the Michigan leases as less favorable than the ones it had structured elsewhere. A person familiar told CNBC that Blue Owl pulled out over unfavorable debt terms, repayment structure, and the risk that local politics would delay construction. Oracle’s spokesperson said Related Digital “selected the best equity partner from a competitive group of options, which in this instance was not Blue Owl,” and that equity negotiations were on schedule. Related Digital called the claim that Blue Owl walked away “unequivocally false.” Blackstone was reported to be in talks but had not signed. Label: reported withdrawal from talks, with an explicit company dispute over the “walked away” framing. [web:9] [web:5] [web:34]
• The replacement closed in late April 2026 as a $16 billion package, not a bank construction loan. Bank of America sold $14 billion of 144A senior secured notes maturing in 2045, with a 7.5 percent coupon, priced at 98.75 cents on the dollar. PIMCO bought about $10 billion. Blackstone contributed about $2 billion of equity; Related Digital supplied the rest of the equity. Goldman Sachs and Wells Fargo advised Related Digital. Related Digital and Blackstone announced the financing on April 24, 2026, subject to customary closing conditions. One account described six years of interest-only payments followed by amortization. Investors had pushed for more than 100 basis points over Oracle’s 2040 corporate bonds and for removal of a proposed early-call option. Banks had retreated; a bond fund anchored the deal. [web:115] [web:119] [web:35] [web:13]

Apollo, Goldman, and who is actually in the stack

Apollo is not a reported lender or equity sponsor on Abilene, Vantage’s Texas or Wisconsin campuses, Project Jupiter, or the Michigan campus. Its Oracle-related role, as of early October 2026, is separate: the Wall Street Journal reported that Oracle is negotiating with Apollo and Goldman Sachs on an off-balance-sheet vehicle that would buy AI chips for about a gigawatt of capacity and lease them to Oracle. No size or close was reported. Apollo’s $35 billion chip financing with Blackstone, announced in June 2026, backs Anthropic’s lease of Broadcom-related chips, not an Oracle data-center site. [web:90] [web:86] [web:89]

Goldman Sachs is in the site stack as an administrative agent on the $18 billion New Mexico loan and as an allocated underwriter on the $38 billion Vantage package, and as an advisor (with Wells Fargo) to Related Digital on Michigan. JPMorgan led Abilene and co-led the Vantage package. MUFG co-led Vantage and is an administrative agent on New Mexico. SMBC and BNP Paribas are administrative agents on New Mexico; SMBC was also in the Vantage underwriter group. Bank of America structured Michigan. Santander and Jefferies are the banks named as stuck long the New Mexico loan in secondary quotes. [web:7] [web:62] [web:67] [web:115]

Pullbacks, renegotiations, and delays

The pattern is not a wave of defaults. It is a syndication and documentation problem that got worse as Oracle’s credit and local permitting both deteriorated.

• Michigan equity: Blue Owl did not proceed (December 2025, reported, disputed in framing). The project was re-cut from a bank construction loan into a 2045 project bond and closed in April 2026 only after PIMCO took about $10 billion. [web:9] [web:41] [web:119]
• Vantage $38 billion: months of syndication against counterparty caps; some holders reportedly willing to sell at a discount; SRT structures explored. Vantage said distribution to more than 50 lenders was successful and a second-quarter 2026 close was expected. No public cents-on-the-dollar quote was found. [web:8] [web:71]
• New Mexico $18 billion: syndication stalled; private quotes at 89–91 cents as of September 18, 2026. [web:67]
• Abilene expansion: Oracle/OpenAI did not take the additional buildings; Microsoft did. Reported. The financed 1.2 GW campus was not described as cancelled. [web:132] [web:125]
• Pricing regime shift: the early Stargate construction loans cleared around SOFR plus 250 basis points. By late January 2026, unsold Oracle-linked project debt was being talked at SOFR plus 300 to 450 basis points. Michigan’s bond, a different instrument, cleared at a 7.5 percent coupon to 2045. [web:3] [web:119]

Force majeure and construction disputes

Each item below is labeled by how it was established.

Project Jupiter force majeure — reported, not independently confirmed by a filed document, and not denied. On September 24, 2026, Bloomberg reported, citing people familiar, that Oracle sent a force majeure notice to Blue Owl’s STACK Infrastructure. The reported purpose was to preserve the right to defer payments if the campus misses a 2028 online date, not to terminate the lease. Oracle did not confirm the notice in those words. Spokesperson Michael Egbert said the project “remains on our planned schedule,” and that force-majeure notices “are commonplace in developments of this scale” and “do not, by themselves, establish a project delay.” STACK said Blue Owl, STACK, and Oracle “remain fully aligned” and that the notice “does not change the financial commitments.” Reuters, citing a person familiar, said the program is being delayed by a year and that Oracle invoked force majeure because of delays securing power. The Financial Times, citing people familiar, reported that even if rent is deferred, Oracle still owes a “carry cost” covering lender interest and a set equity return, for up to three years if the first phase is not ready by the first half of 2027, and that Oracle cannot terminate the lease. Label: reported (multiple outlets, people familiar; company statements acknowledge the notice category without confirming a delay). [web:47] [web:49] [web:12] [web:50]

The underlying power problem is documented in regulatory reporting, separate from the contractual notice. New Mexico’s State Land Office denied Energy Transfer a pipeline right-of-way twice; the targeted in-service date slipped to February 1, 2027. A state air permit for the Bloom Energy fuel-cell microgrid (up to about 2.45 GW) has a decision deadline of November 23, 2026. Bloomberg reported on October 8, 2026, citing people familiar, that Oracle is considering trucking compressed natural gas to Jupiter so early stages can start before the pipeline is in service, and that it has already used trucked gas for more than a year at a Salt Lake City-area site and for early work at the Shackelford County campus. Oracle did not comment. Label on trucking at Jupiter: reported, not confirmed by the company. [web:53] [web:36] [web:105]

Wisconsin collateral lawsuit — confirmed filing, then withdrawn; not a construction or lender dispute. Oracle sued the Wisconsin Public Service Commission on June 19, 2026, over a rule requiring very large customers rated below A- to post collateral for new generation. Oracle said the rule could require more than $7 billion of security at an annual cost that could exceed $100 million. Oracle voluntarily dropped the suit in August 2026. A separate resident lawsuit over Port Washington’s tax-increment district was dismissed in September 2026. We Energies has said the campus remains on schedule. An infrastructure-data firm, Aterio, published an analysis on September 30, 2026 arguing that a restarted transmission review creates a material risk that meaningful load arrives closer to mid-2028 than Oracle’s second-half 2027 guidance. That is an analyst view, not a force majeure claim or a lender dispute. [web:76] [web:75] [web:57] [web:83]

Abilene financing walk-away — reported, and limited to an expansion. No force majeure notice involving Oracle, Crusoe, Blue Owl, or lenders on the financed 1.2 GW campus was found in public reporting. The dispute that was reported is the failed expansion lease, described above.

Michigan — no force majeure found. The documented fight was over equity and debt terms in late 2025, resolved by a different capital structure in April 2026. Construction was reported underway by the time financing was announced. [web:41] [web:115]

Flooding at Jupiter — reported, parties say no schedule impact. Flooding reached the site around September 30, 2026. Oracle and STACK said in a joint statement that no critical infrastructure was damaged and the schedule was unchanged. [web:54]

What this means for anyone underwriting the next campus

The early deals worked because a lease to an investment-grade Oracle, plus an IG rating on a construction loan, let project-finance banks syndicate at SOFR plus 250 basis points and get the risk off their books. That channel is clogged. Counterparty caps, a BBB- tenant, and local permit risk on power are now priced in three different places: wider talked spreads on unsold loans, a 9 to 11 point discount on the one loan that has been quoted, and a 7.5 percent, 19-year bond that only cleared because PIMCO took most of it.

The contractual fine print matters more than the headline lease. On Jupiter, reported terms say Oracle cannot walk away, still covers debt service via a carry cost during a delay, and may only defer rent — while Blue Owl’s equity yield is lower until completion, so a year of delay hits the sponsor’s IRR even if lenders are kept current. That is why a force majeure notice that both sides call routine still moved the loan quotes and, according to Reuters, chilled other OpenAI-linked financings. Anyone competing for the next Oracle or OpenAI campus should assume bank construction loans at SOFR plus 250 basis points are no longer the base case, that equity partners will reopen lease terms if Oracle’s rating or local politics worsen, and that power permits — not the building loan — are the event that can reopen who pays during a delay.

---

Recent Findings Supplement (October 2026)

Project Jupiter (New Mexico) financing and force majeure developments dominate recent reporting. Oracle issued a force majeure notice in September 2026 to a Blue Owl Capital unit (STACK Infrastructure, the developer) over power supply delays at the ~1,400-acre, up to 2.45 GW campus (part of the Stargate/OpenAI effort). The notice seeks to protect against or defer rent obligations if the project misses its 2028 target due to issues like pipeline rerouting (after state regulator rejections and local opposition), permitting hurdles, and related infrastructure lags.[1][2]

The ~$18 billion syndicated construction loan facility (arranged late 2025/early 2026 by a group of ~20 banks led by BNP Paribas and Goldman Sachs, with MUFG and SMBC also involved) has faced secondary-market pressure. It has been quoted at 89–91 cents on the dollar in mid-to-late September 2026 reporting, with syndication stalling and lenders including Santander and Jefferies reportedly holding larger positions than planned amid concerns over Oracle’s leverage, credit profile (S&P cut to BBB- in July 2026), and project timelines. Blue Owl has approximately $3 billion in equity invested; the broader project has been referenced with planned investment around $165 billion.[3][4][5]

Parties have stated they remain aligned and that the notice does not alter ultimate financial commitments (reported status). Oracle has explored workarounds such as trucked compressed natural gas (CNG) to enable partial early operations, with pipeline in-service now targeted for February 2027. The loans trade at a discount comparable to (or slightly better than) Oracle’s own bonds in some analyses.[6][7]

Texas and Wisconsin Oracle-leased projects show lender distribution challenges. A group of banks including JPMorgan, MUFG, Morgan Stanley, and SMBC has spent more than six months attempting to distribute or offload portions of an approximately $38 billion construction debt package tied to these sites. Some lenders have explored sales to non-bank buyers at a discount to reduce exposure, reflecting broader caution on holding large AI-related construction loans.[8]

Michigan site updates remain largely tied to late-2025 reporting with limited new post-April 2026 developments. Blue Owl declined to back the planned ~$10 billion Saline Township project after debt-term negotiations stalled (concerns cited included shifting market sentiment on AI spending and Oracle’s debt load). Oracle disputed characterizations of a “walk-away,” stating its partner (Related Digital) selected another equity partner from a competitive process; Blackstone was reported in talks as a potential replacement but had not committed as of the reports. No major new financing announcements or renegotiations for this site appear in the most recent sources.[4][9]

Broader lender and market dynamics reflect tightening terms and secondary-market activity. Across AI data center debt, lenders are demanding higher yields, stricter milestone protections, delay-draw structures, and greater emphasis on permits/community support before committing or advancing funds. Some banks (e.g., references to Société Générale and MUFG) have stepped back from certain deals. Secondary trading of Oracle-linked project debt at discounts has contributed to market ripples, with analysts noting impacts on pricing and sentiment. Oracle has separately pursued new large-scale debt arrangements (e.g., with Apollo and Goldman Sachs) for AI chip financing via off-balance-sheet structures.[10][1]

Crusoe/Abilene and other sites (e.g., Vantage mentions are sparse) have fewer new loan-specific updates. Recent focus has been more on Crusoe equity raises (e.g., portions of a $3.9 billion Series F closing in September 2026) and operational/land filings than on fresh construction loan details or disputes. Earlier financing (such as the JPMorgan-led $7.1 billion Abilene facility) predates or lacks significant new post-April reporting in the results.[11]

No confirmed force majeure-triggered payment defaults or major renegotiations beyond the reported Jupiter notice and general tightening; most items are reported (with some elements like exact trading levels based on banker/investor quotes). Wisconsin-specific collateral disputes (Oracle lawsuit over a potential $7 billion bill) and Oracle’s overall debt levels (~$125–169 billion range across reports) appear in context but add limited new project-finance specifics.[8]

These shifts indicate lenders are reallocating risk more conservatively on Oracle-linked and similar AI infrastructure deals, with secondary-market pricing and force majeure mechanics serving as key signals of caution amid power and permitting bottlenecks.

Oracle is the credit market’s liquid proxy for AI-buildout risk: agencies still rate it investment grade, but 5-year CDS near 260 basis points and long bonds yielding above 8% already price it closer to high-yield. The gap is the cost of funding a cash-flow-negative infrastructure ramp that peers are mostly financing from stronger balance sheets.

Rating actions: one notch from junk at S&P, split among the agencies

S&P is the agency that has already moved. On July 9, 2026 it cut Oracle’s long-term issuer credit rating to BBB- from BBB and the short-term rating to A-3 from A-2, with a stable outlook. That leaves Oracle one notch above speculative grade at S&P. The agency had put the outlook at negative in July 2025 specifically because of the pace of the AI infrastructure buildout. The July 2026 cut reflects a more cautious view of the AI infrastructure industry itself—rising capex, an uncertain path to profitability, a shifting competitive landscape, and high customer concentration—plus weaker free operating cash flow in fiscal 2027 and S&P-adjusted leverage staying above 4x for the next two years (mid-4x in fiscal 2027, improved from a prior 4.8x forecast after equity issuance). S&P has flagged OpenAI concentration in the remaining performance obligations as a central credit risk. It expects fiscal 2027 capex of about $90–95 billion and a free operating cash flow deficit near $42 billion, financed with a mix of debt and equity. [1] [2] [1] [3]

Further S&P downgrade triggers (explicitly stated):
- Sustained leverage above 4.5x (the old BBB threshold was 4x; base case peaks around 4.4x and is not expected to breach 4.5x).
- A view that Oracle is not on a path to positive free operating cash flow by fiscal 2029.
- A more negative view of Oracle’s AI strategy or of the AI industry overall.

An upgrade is described as unlikely over the next two years and would require growth targets plus consistently positive free operating cash flow. [1] [3]

Moody’s is one notch higher and still negative. It rates Oracle Baa2 (equivalent to BBB) with a negative outlook. On February 2, 2026 it affirmed Baa2 and kept the outlook negative around the equity-issuance announcement. In the February commentary explaining that outlook, Moody’s said no other hyperscaler is as leveraged or as cash-flow negative entering this build phase, and that leverage “will increase materially in the near term and could approach 5x temporarily.” Public reporting through early October 2026 still describes the rating as Baa2/negative; no subsequent downgrade has been reported. [4] [3] [5] [6]

Fitch is the most stable of the three. It affirms BBB / F2, outlook stable. That was the stance on October 13, 2025, and again on February 2, 2026 when it rated a new senior-notes deal BBB and reaffirmed the IDR. Fitch treats the AI capex as tied to multi-year contracts (OpenAI, xAI, Meta, Nvidia) that it expects to be repaid inside the contract window, with EBITDA leverage near or above its 3.5x sensitivity in fiscal 2026 and then declining in fiscal 2027–28 as AI compute revenue arrives. It has forecast deeply negative pre-dividend free cash flow (over $11 billion in both fiscal 2026 and 2027 in the October 2025 note; later commentary cited over $26 billion and $18 billion). [7] [8] [9] [10]

Fitch negative-rating sensitivities (February 2026 action):
- EBITDA leverage sustained above 3.5x, or lease-adjusted (EBITDAR) leverage sustained above 4x.
- (Cash flow from operations minus capex) / debt sustained below 15%.
- A shift away from the publicly disclosed balanced debt-and-equity funding mix for calendar 2026.
- AI compute revenue failing to materialize in the first half of fiscal 2027.
- Sustained revenue declines with EBITDA margins below 40%.

Upgrade would require EBITDA leverage sustained below 3.0x (or EBITDAR below 3.5x). [10]

The split matters for index mechanics. Major investment-grade indexes typically use the middle of the three ratings. With Moody’s and Fitch still at the BBB/Baa2 level and only S&P at BBB-, a single further cut does not automatically make Oracle a fallen angel. Market commentary has nonetheless focused on roughly $120 billion of bonds that could leave IG indexes if the ratings converge on junk—an analyst/market estimate, not an agency figure. Morgan Stanley credit analyst Lindsay Tyler has said two low-BBB ratings, possibly by calendar year-end 2026, is a live risk. [3] [11] [12]

What this means for anyone competing with or lending to Oracle: the rating is being held up by time and equity, not by current cash metrics. S&P has said on its own call that today’s credit metrics are not investment-grade and that the rating rests on cash harvest in years three to five of the AI contracts. Miss that path—or let leverage sit above 4.5x—and the next S&P move is to junk. [13] [14]

5-year CDS: from the 50s to a post-crisis high near 261 bp

Oracle’s 5-year CDS is the cleanest public price of that bet. The path over the past twelve months is a steady re-rating, then a break after the S&P cut.

• Late September 2025: Morgan Stanley’s recommended long-CDS entry was 57 bp. Around the same period, protection was still described as below roughly 40 bp earlier in the summer of 2025. [15] [16] [17]
• November 2025: above 105 bp (S&P data cited by Reuters); late November about 125 bp, a three-year high at the time (ICE). [18] [19]
• December 11, 2025: 139 bp, the highest in at least five years after an earnings miss (S&P Global Market Intelligence / Reuters). [20]
• Start of 2026: about 144 bp. [21] [21]
• July 20–21, 2026 (post-downgrade): about 203 bp, the highest in ICE data going back to late 2008, after a prior peak of 198 bp. Late July prints near 212–215 bp. [22] [23] [24] [25]
• August 24, 2026: year-to-date high then of 219.66 bp, having climbed from the 170s after the downgrade. [26]
• September 25, 2026: 232.41 bp (Yonhap Infomax), with a separate late-September print near 227 bp also called a record at the time, around the Project Jupiter force-majeure headlines. [26] [27]
• October 8–9, 2026: fresh highs near 257–261 bp. Multiple reports citing Bloomberg put the close at 261 bp, with an implied five-year default probability above 20%. IFR had 257 bp on the Thursday of that week. An October 8 market screenshot showed 250.5 bp. [28] [29] [30] [31]

Peak in the past 12 months, and the post-2008 high: about 261 bp in early October 2026. Current level (as of October 8–13, 2026): at or just off that peak, roughly 250–261 bp. Levels move daily; these are reported closes and intraday highs, not a single consolidated tape.

That is a roughly four- to five-fold widening from the mid-2025 entry Morgan Stanley used, and more than double the 2008-era reference some desks still cite near 198 bp. At 261 bp, insuring $10 million of Oracle debt costs about $261,000 a year.

Peers: Oracle is the outlier; CoreWeave is a different credit

The same AI-debt wave has lifted protection costs across tech, but the levels are not comparable.

Name	Approx. 5-year CDS	When reported	Context
Oracle	~203 bp, then ~261 bp peak	July 2026; early Oct 2026	BBB-/Baa2/BBB; negative FCF; OpenAI concentration
Meta	~95 bp (from ~56 at end of prior year); ~95 bp on July 30	Late July 2026	AA-tier ratings; heavy AI spend but still a strong cash generator historically
Alphabet	record ~64–67 bp (from ~50 when the contract became more active in March)	Late July 2026	FCF turned negative in one quarter; still high IG
Microsoft	~53 bp (from ~35)	Late July 2026	AAA / Aaa; has not issued bonds since 2017; CDS rose the least
CoreWeave	~855 bp in late July (implied ~50% five-year default odds); later back above 800 bp in early September after a dip to ~672	July–Sept 2026	Speculative-grade AI cloud; not an IG peer

[25] [25] [32] [33] [34] [35]

July 2026 was the moment peer CDS set records too—Meta near 92–95 bp, Amazon near 67 bp, Alphabet near 64 bp, Nvidia near 79–82 bp—while a basket of Oracle, Amazon, Alphabet, and Microsoft was cited around 75 bp, and the same basket excluding Oracle around 49 bp. By early October, Oracle and Broadcom (Broadcom 5-year CDS at a record 136 bp) were the names resetting highs on every financing headline. Credit analysts who have looked through the hyperscaler complex generally conclude the “exploding CDS” story is mostly an Oracle story: Microsoft, Alphabet, Amazon, and Meta still sit in AA–AAA territory with debt-to-equity well below 1x on reported balance sheets, while Oracle’s on-balance-sheet leverage and lease stack do not. [36] [37] [38] [39]

CoreWeave is the wrong comparison for rating analysis and the right one for “how far AI infrastructure credit can gap.” At 800-plus basis points the market is pricing distress odds, not a temporary IG leverage spike. Oracle at ~260 bp is expensive insurance for a BBB credit and cheap insurance relative to a pure-play GPU lessor.

Bond spreads and yields: the market has already downgraded the long end

Cash bonds have moved with the CDS, and the long end is where the cost of capital shows up.

• In July 2026, the spread on Oracle’s 6.7% bonds due 2056 widened about 8 bp to 263 bp; the 5.7% notes due 2036 widened about 9 bp to 205 bp. Ten-year Oracle yields were quoted around 6.4–6.5%, closer to BB than to the BBB index average. [23] [40] [41] [42]
• Around September 24, 2026, after the Project Jupiter force-majeure report, the 6.7% 2056s crossed an 8% yield for the first time. By October 8 they were quoted near 82 cents on the dollar, yield to maturity about 8.3%, G-spread about 268 bp. [11] [6] [29] [31]
• A Morgan Stanley note dated October 7, 2026 cited Oracle’s 10-year dollar bond yield at 7.5% alongside the CDS move from 57 bp to 260 bp. [15]
• As of early October, a 6.85% bond due 2066 was screening at roughly 8.45% yield to maturity (price about 82); a German venue showed a similar 2066 note near 8.7% yield at about 81 cents on October 12–13. Only Oracle and SpaceX, among the hyperscaler complex screened by one October 5 survey, had bonds yielding above 7.5%. [43] [44]
• MarketAxess noted that Oracle 30-year spreads had widened about 150 bp since the September 2025 $18 billion AI-infrastructure deal, and that the bonds were trading wide of liquid BB names (about 25 bp wide of Charter, 100 bp wide of Viatris) after having traded tighter than almost all of those comps in summer 2025. Bid-ask and daily volatility had also migrated toward high-yield behavior. [45]

Part of the yield is the Treasury backup—the 30-year Treasury was cited near 5.63% in early October, up about 39 bp in a month—but the Oracle-specific spread is the larger story. New-issue coupons in the mid-6s on 2025–26 deals are already stale relative to an 8% secondary yield on the long bond. [43]

CDS as the hedge for AI exposure

Analysts do not treat Oracle CDS as a bet that Oracle fails next quarter. They treat it as the most liquid way to short the credit leg of the AI capex boom.

Mechanism. Banks and direct lenders financing data-center construction loans—where Oracle is the future tenant, not always the borrower—buy Oracle CDS to hedge takeout and tenant risk. Morgan Stanley’s Lindsay Tyler and David Hamburger have argued that those construction-loan hedges, more than traditional bondholders alone, became a primary driver of CDS demand from late 2025 onward. When loans are slow to syndicate, the hedge stays on. A Bloomberg-reported $38 billion construction financing for Texas and Wisconsin sites leased to Oracle was still being distributed more than six months later, with some banks exploring significant risk transfers because internal AI-exposure limits were binding. [19] [46] [47]

Who is using it.
- Morgan Stanley credit has been structurally long Oracle 5-year CDS since a 57 bp entry on September 28, 2025. By October 2026 the same trade was marked near 260 bp. The October 7 note still recommends buying 5-year CDS and selling the 2035 and 2055 bonds, on the view that cash funding needs and lease liabilities keep deteriorating, and that lender hedging will keep CDS bid. An earlier shift (November 2025) dropped the long-bond leg of a basis trade and went outright long CDS because of the funding gap, balance-sheet growth, capex obsolescence, ratings pressure, and counterparty risk. They had warned CDS could clear 150 bp and approach 200 bp if financing communication stayed thin; both levels were exceeded. [16] [17] [15]
- Goldman Sachs’ credit desk, per Bloomberg as relayed in early-October market reports, said single-name AI CDS volumes were up about 400% year over year, with Oracle the most active name. [29]
- Thematic and macro accounts use Oracle because it is the widest liquid IG tech CDS and the name with the most fallen-angel optionality. Net notional CDS on major tech names was reported at a record $12.5 billion in July 2026, up about 500% since Q2 2025, with Oracle alone about $6.5 billion (LSEG data cited via the Financial Times). [21] [21]
- Protection sellers exist on the other side. In November 2025, Boaz Weinstein’s Saba Capital was reported to have sold CDS on Oracle, Microsoft, Meta, Amazon, and Alphabet to banks that wanted the hedge—evidence that the product had become a two-way market for AI credit risk, not a one-way panic. [18]

Why Oracle, not Microsoft or Alphabet. Microsoft can point to positive free cash flow and AAA ratings; its CDS barely cleared 50 bp even as issuance anxiety peaked. Alphabet and Meta widened, but from a low base and with ratings that do not threaten index exclusion. Oracle combines the largest relevant CDS market, negative free cash flow (about $23.7 billion in fiscal 2026), debt on the order of $125–130 billion, a backlog in which OpenAI is widely described as roughly half of remaining performance obligations (total RPO figures cited around $638–664 billion), and an S&P rating one notch from junk. That is why every new financing headline—bond deal, chip-lease SPV talks with Apollo and Goldman, force majeure on a New Mexico campus—has widened Oracle CDS rather than tightened it. Chip-lease SPVs may keep some debt off the parent’s face balance sheet; Morgan Stanley’s point is that lenders still hedge the parent, so the CDS still widens. [22] [48] [49] [16]

Lisa Shalett, CIO at Morgan Stanley Wealth Management, put the market’s heuristic bluntly in late 2025: the opening screen was Oracle CDS, because if investors start doubting Oracle’s ability to pay, that is an early read on nerves about the whole debt-financed AI trade. [50]

What this does to Oracle’s cost of capital

The transmission is direct. Oracle’s incremental dollar of AI capacity is funded at a spread the agencies have not fully marked into the rating, but the new-issue market will.

• All-in borrowing cost has stepped up. Recent long bonds were issued with coupons in the mid-6% area. Secondary long yields above 8%, and a cited 10-year yield of 7.5%, mean the next unsecured deal clears wider unless spreads reverse. S&P’s own base case still has a fiscal 2027 free-cash-flow hole near $42 billion against capex of $90–95 billion, so Oracle cannot wait out the market. Management has guided to roughly $40 billion of debt and equity funding in fiscal 2027, including a $20 billion common-stock program, after about $43 billion of debt and $5 billion of equity in fiscal 2026. [2] [22] [51] [15]
• Fallen-angel optionality is already in the price. Bonds trading through many BB comps, and a 2056 yield above the average single-B yield in some market commentary, mean IG buyers are demanding a high-yield concession today. If Moody’s outlook converts to a downgrade or Fitch’s 3.5x leverage tripwire is treated as sustained, the technical selling from index accounts would widen that concession further. S&P has already said another cut follows leverage sustained above 4.5x or a loss of confidence that free cash flow turns positive by fiscal 2029. [45] [1] [3]
• Equity is the pressure valve, and it is not free. The February 2026 mandatory convertible preferred ($5 billion) and the planned common issuance are what kept S&P’s leverage forecast from the high-4x area and what Moody’s cited when it affirmed rather than cut. Dilution protects the rating and therefore the debt cost, but it raises the equity cost of the same AI projects. That is a different tradeoff than Microsoft funding fiscal 2027 AI capex from free cash flow, or Alphabet and Meta issuing at spreads a fraction of Oracle’s. [1] [4] [52]
• Off-balance-sheet structures do not fully escape the CDS. Lease commitments that have not yet started have been reported in the high hundreds of billions (one September 2026 figure: $288 billion of mostly data-center leases not yet commenced, on top of $125 billion of borrowings). Chip-lease SPVs with loan-to-cost above 50%, if they close, may reduce parent bond supply. They do not stop lenders from buying Oracle CDS, and rating agencies have said leases reduce financial flexibility because the assets cannot be sold or pledged. Delays concentrate the problem: Morgan Stanley has warned that a Wisconsin power delay into mid-2028 could bunch capex with other campuses and with the refinancing of an Abilene construction loan, so the funding need hits in a short window precisely when the credit spread is already near high-yield levels. [48] [29] [5] [15]
• Relative cost is the competitive point. At ~50–70 bp, Microsoft and Alphabet can add a dollar of AI capacity without the capital structure becoming the story. At ~260 bp CDS and 8% long yields, Oracle’s hurdle rate on the same GPU-hour is higher before a single dollar of OpenAI revenue is collected. If contracted revenue slips—OpenAI annualized revenue was reported in early October well short of earlier figures circulated in the market—the debt service does not slip with it. That asymmetry is why the credit market is the binding constraint on Oracle’s AI strategy, and why its CDS has become the hedge for everyone else who does not want to underwrite that constraint. [49] [28] [31]

Bottom line for capital providers and competitors: Oracle can still issue, and management has said it intends to stay investment grade. The price of that paper is no longer a mid-BBB price. Until free cash flow inflects—or equity and customer prepayments do the deleveraging the agencies are underwriting—the marginal cost of Oracle’s AI capacity will keep being set in the CDS and the long bond, not in the rating letter.

---

Recent Findings Supplement (October 2026)

S&P downgraded Oracle to BBB- (lowest investment-grade notch) on July 9, 2026, with a stable outlook, citing accelerated AI infrastructure risks and cash flow strain; Moody’s holds Baa2 with negative outlook; Fitch maintains BBB stable.[1][2]

S&P’s action reflected heightened business risk from the AI buildout (including customer concentration, with OpenAI representing roughly half of Oracle’s $638 billion contractual backlog), sharply higher capex forecasts ($90–95 billion for FY2027, up from a prior $60 billion estimate), and a projected free operating cash flow (FOCF) deficit of nearly $42 billion in FY2027. Adjusted leverage is expected to reach the mid-4x area in FY2027 (improved from prior forecasts partly due to a $5 billion mandatory convertible preferred issuance in February 2026 and planned equity raises), exceeding S&P’s typical 4x threshold for BBB. The stable outlook assumes scaling AI capacity will eventually generate cash flow in years 3–5 of contracts, with deficits funded via debt and equity mixes. S&P explicitly noted Oracle (alongside SpaceX) as an “outlier” for investment-grade status despite current metrics not fully supporting it.[3][4][5]

Moody’s negative outlook (in place since at least February 2026 and referenced in August 2026 commentary) highlights Oracle’s leveraged position entering the AI build phase relative to peers, with potential temporary leverage approaching 5x. Fitch affirmed BBB/F2 with stable outlook in February 2026 (still referenced as current in October 2026 materials), noting elevated near-term capex but expected leverage decline in FY2027–2028 as AI revenue scales; its triggers include sustained EBITDA leverage above 3.5x (or above 4x including rents). No further agency actions have been reported since July 2026.[6][7]

Oracle’s 5-year CDS has widened dramatically to record levels of ~260–261 bps as of early October 2026 (implying >20% 5-year default probability), up from ~144 bps early/mid-2026 and ~200+ bps in July 2026; this is the widest among major tech/AI peers.[8][9]

Recent peaks reflect successive financing announcements, S&P’s downgrade, data center execution concerns (e.g., force majeure notice on the New Mexico Project Jupiter site in September 2026), and broader AI capex anxiety. Earlier points in the 12-month window included ~57 bps levels referenced in Morgan Stanley trading notes (around late 2025 entry). Peers as of July–October 2026 snapshots: Broadcom at record 136 bps; Nvidia ~80–87 bps (record near 82 bps in July); Meta ~86 bps; Alphabet ~62 bps; Microsoft ~51 bps. Oracle stands out as the most stressed. CoreWeave (a smaller AI infrastructure peer rated in the BB-/B+ range) saw its own CDS spike sharply (e.g., ~855 bps / ~50% implied default probability in July 2026), though direct apples-to-apples comparisons are limited.[10][11][12]

Bond spreads and yields have widened sharply in tandem, with Oracle’s longer-dated paper (e.g., 6.7% due 2056) trading above 8% yield for the first time and 10-year yields around 7.5% by October 2026; spreads now trade wide to many BB names.[13][9]

Yields on 30-year bonds widened ~150 bps since September 2026 amid heavy issuance and rising Treasury yields. This reflects market pricing of near-junk credit risk despite the IG ratings, with increased volatility, wider bid-ask spreads, and higher trading volumes in both bonds and CDS. AI-linked debt issuance across tech has been massive (~$500 billion in 2026 per some estimates), amplifying scrutiny.[8]

Analysts view the CDS widening primarily as hedging against concentrated AI execution and financing risks rather than outright default fears, with Morgan Stanley actively recommending long CDS positions.[9]

Morgan Stanley (October 2026 research) highlights lender hedging, basis trades, and shorts as drivers, alongside structural concerns like potential 2028 capex concentration (power/infrastructure bottlenecks could delay revenue while large construction loans mature) and customer concentration risks. High single-name AI CDS volumes (Oracle often leading, with Goldman noting 400% YoY surge in such trading) indicate investors using CDS to hedge capex burn, negative FCF periods, and uncertain payback on multi-year contracts. This is distinct from equity optimism around backlog growth.[14]

Implications include materially higher Oracle cost of capital, potential “fallen angel” index exclusion risk on any further downgrade (~$120 billion in bonds could be impacted), and pressure toward equity raises, SPV structures (e.g., chip-leasing vehicles), or customer prepayments to bridge funding gaps.[15]

With ongoing multi-billion-dollar annual financing needs, wider spreads directly raise incremental borrowing costs and could constrain flexibility if leverage metrics deteriorate further. Peers with stronger balance sheets (e.g., Microsoft, Alphabet) face milder pressure, highlighting Oracle’s relative vulnerability in the AI infrastructure race. No major new rating or CDS data has emerged in the immediate days before October 13, 2026.

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.