Battery Park, Manhattan

research@frothcity.com

← All reports
ORCL Short AI / Data Centers

Oracle: The Software Franchise Is Funding the Build

Oracle still earns its interest. The short is what happens next. Cash capital spending already exceeds the cash the operations produce, half the backlog is not revenue until late 2029, and the leases that will carry that backlog are not on the balance sheet yet.

Executive summary

We are short Oracle. This is not the CoreWeave short. CoreWeave’s June quarter lost money from operations and still owed $640 million of interest. Oracle’s August quarter produced $6,728 million of operating income and $4,760 million of net income. The software and support franchise is real, no customer was 10% of fiscal 2026 revenue, and the dividend was held at $0.50. A thesis that needs Oracle to miss a coupon this year is the wrong thesis.12

The short is the build layered on top of that franchise. In the quarter ended August 31, 2026, Oracle’s own free cash flow, operating cash flow minus capital expenditures, was $(5,396) million. That operating cash flow already included $11,363 million of customer prepayments. Cash capital expenditures were $28,499 million, against $8,502 million a year earlier. Fiscal 2026 free cash flow was $(23,686) million. The company says fiscal 2027 capital spending will be higher than fiscal 2026’s $55,663 million. S&P, describing what it calls Oracle’s guidance, puts that year at $90 billion to $95 billion.125

Remaining performance obligations were $664 billion. Oracle expects to recognize 13% of that as revenue over the next twelve months and 37% over the following two years. Half is not expected until September 2029 or later. Notes payable and other borrowings were $125.3 billion, not more than $130 billion. Interest expense was $1,428 million, up 55% from $923 million. Off the balance sheet, additional lease commitments, substantially all data centers, were $288 billion. S&P cut the issuer rating to BBB− on July 9, 2026. It also estimates that OpenAI is roughly half of the $638 billion of RPO at May 31, 2026. Oracle has not disclosed that split.15

How we rank the short. Likelihood and impact are Froth City’s judgment. The dollars under them are from Oracle’s filings, except the OpenAI share of RPO, which is S&P’s estimate.
# Reason Likelihood Impact Timing
1 Capital spending already exceeds operating cash, and the company says the year will be heavier High Severe Now, through fiscal 2027
2 Half the backlog is not revenue until late 2029. Depreciation is current. High Severe Costs now; revenue through 2031 and after
3 $125.3 billion of borrowings, interest up 55%, and $288 billion of data-center leases not yet recorded High Severe Interest now; leases commence fiscal 2027–2029
4 The concentration is in the backlog, not in this year’s revenue Medium-high Severe Contract life, multi-year
5 Shareholders are already writing the check High High Done in the August quarter; more if the build continues
Q1 free cash flow
$(5.4)B
Q1 cash capex
$28.5B
RPO
$664B
Due after Aug 2029
50%
Borrowings
$125.3B

Contents

  1. 1 Capital spending the operations do not fund
  2. 2 Half the backlog waits until late 2029
  3. 3 Borrowings, interest, and the leases not on the books
  4. 4 Concentration sits in the backlog
  5. 5 Shareholders are writing the check
  6. 6 What the price assumes
  7. 7 Risks to the short
  8. 8 Other considerations

01 Capital spending the operations do not fund

Oracle defines free cash flow as cash from operations minus capital expenditures. It is a non-GAAP measure, and the company is explicit about that. On that definition, free cash flow was $(362) million in the quarter ended August 31, 2025, $(23,686) million for the year ended May 31, 2026, and $(5,396) million in the quarter ended August 31, 2026. The latest quarter is not a one-off against a cash-generative base. The year before it was already negative, and the year that contains it was negative by $23.7 billion.12

Oracle free cash flow, in millions, as the company defines it Quarter ended Aug 31, 2025 $(362) Year ended May 31, 2026 $(23,686) Quarter ended Aug 31, 2026 $(5,396)
Figure 1. Free cash flow as Oracle defines it: operating cash flow minus capital expenditures. Bars are proportional. Sources: Form 10-Q and Form 10-K.
Table 1. Cash from operations, capital expenditures, and free cash flow, in millions. The August 2026 operating figure includes $11,363 million of customer prepayments.
Quarter ended Aug 31, 2026 Quarter ended Aug 31, 2025 Year ended May 31, 2026
Cash from operations23,1038,14031,977
Capital expenditures(28,499)(8,502)(55,663)
Free cash flow(5,396)(362)(23,686)
Customer prepayments in operating cash11,363None4,592

Read the latest quarter without the prepayments and the hole is larger. Operating cash flow of $23,103 million included $11,363 million of deferred revenue from customer prepayments that Oracle says contain a significant financing component. There were none of those prepayments in the year-ago quarter. Strip them out of the latest quarter, which is our subtraction and not a second company measure, and cash from the rest of the business was about $11.7 billion against $28.5 billion of capital spending. The prepayments are real cash. They are also customers financing Oracle’s equipment, not the software business throwing off surplus cash.1

Management’s own words on the direction are narrower than the tape suggests, and we will not dress them up. The 10-Q says cash capital expenditures rose because of data-center expansion, and that fiscal 2027 capital expenditures are expected to be higher than fiscal 2026. S&P Global Ratings, on July 9, 2026, said Oracle had guided fiscal 2027 capex to $90 billion–$95 billion, against S&P’s prior forecast of $60 billion, and that S&P now expects a free-operating-cash-flow deficit of nearly $42 billion in fiscal 2027. That $90 billion to $95 billion figure is S&P’s account of company guidance. It is not a line in the 10-Q. One quarter of the year has already used $28.5 billion.15

Cloud infrastructure is where the spending is showing up in revenue, and it is still the smaller part of the company. Cloud infrastructure revenue was $7,388 million in the quarter, up 121% from $3,347 million, and was 38% of total revenue of $19,345 million. For fiscal 2026 it was $18,101 million, 27% of revenue of $67,357 million. Cloud revenue overall was 60% of the latest quarter, up from 48% a year earlier. The legacy line is still there, and it is not growing: software revenue was $5,550 million, down from $5,721 million. The earnings release calls that a 3% decline and attributes it to customers moving from on-premises software to the cloud.123

Likelihood is high because the cash gap is a reported result, not a forecast, and the company has said the capex year will be larger. Impact is severe because the funding has to come from debt, from customers paying early, or from shareholders. Timing is this fiscal year, which ends May 31, 2027.

02 Half the backlog waits until late 2029

Remaining performance obligations were $664 billion at August 31, 2026, up from $638 billion at May 31, 2026 and $138 billion at May 31, 2025. Oracle expects to recognize about 13% as revenue over the twelve months through August 31, 2027, 37% from month 13 through month 36, 34% from month 37 through month 60, and the rest after that. Thirteen plus thirty-seven is fifty. The second half of the backlog is not expected to become revenue until September 2029 or later. The earnings release says RPO rose $209 billion year over year, and that the company booked more than $30 billion of additional AI cloud contracts in the quarter.123

Share of Oracle remaining performance obligations by recognition window Next 12 months 13% · $86B Months 13–36 37% · $246B Months 37–60 34% · $226B Thereafter 16% · $106B
Figure 2. Recognition windows for $664 billion of RPO at August 31, 2026. Dollar amounts are 13%, 37%, 34%, and 16% of that total, rounded. The red bars start in September 2029. Source: Form 10-Q.

The cost does not wait for that schedule. Depreciation was $3,156 million in the quarter, up from $1,351 million. Property, plant and equipment, net, was $127,845 million, up from $99,957 million at May 31. Cloud and software expenses, which hold the cost of delivering those offerings, were $6,400 million, up from $3,607 million. Cloud and software revenue, the cloud lines plus software, was $17,157 million against $12,907 million a year earlier. On those reported lines the blended margin fell from about 72% to about 63%. That is our arithmetic. Oracle does not publish a gross margin for cloud infrastructure alone, and we will not invent one. The blend still contains software support. If infrastructure is the piece that got more expensive, the infrastructure margin is worse than 63%, not better. We stop there.1

CoreWeave’s backlog converts faster than this and still does not fund CoreWeave. At June 30, 2026, CoreWeave had $103.7 billion of unsatisfied RPO and expected 41% of it within 24 months. Oracle’s $664 billion is a different order of magnitude, and only 13% is inside a year. Both companies are spending before the revenue arrives. Oracle can point to a software profit that CoreWeave does not have. The timing mismatch is the same shape. Our CoreWeave note ranks that company’s principal schedule first because the equity there is thin. Here the timing mismatch is second, behind the cash already going out the door.9

Two company disclosures keep the backlog from being read as a pure unfunded promise. At the end of fiscal 2026, Oracle said the prepaid and customer-supplied hardware portions of its large AI contracts totaled $75 billion, which reduces the capital Oracle itself must raise. In the September 10, 2026 earnings release it said the new AI contracts booked in the quarter do not change its plans for raising capital, that it delivered 850 megawatts of additional data-center capacity in the quarter, and that it had delivered more than 300,000 GPUs since the end of the prior quarter. Those are reasons the short can be early. They are also evidence that capacity, and the depreciation that follows it, is arriving now, while half the RPO is still years out.34

03 Borrowings, interest, and the leases not on the books

Notes payable and other borrowings were $7,625 million current and $117,712 million non-current at August 31, 2026. The total is $125,337 million. At May 31, 2026 the same lines were $7,199 million and $122,342 million, or $129,541 million. The balance fell in the quarter. Cash repayments of senior notes, term loans, and other borrowings were $4,202 million, and the cash-flow statement for the quarter does not show a new bond issue. The claim that borrowings are already above $130 billion is not what this balance sheet says.1

The increase is the year before that. At May 31, 2025, current and non-current notes payable were $7,271 million and $85,297 million, or $92,568 million. Fiscal 2026 brought $46,093 million of proceeds from senior notes, term loans, and other borrowings, net of issuance costs. Interest expense in the August 2026 quarter was $1,428 million, against $923 million, an increase of 55%. Operating income of $6,728 million covered that interest about 4.7 times. CoreWeave’s interest exceeded its operating result. Oracle’s does not. The rate of change is still the point: the coupon book grew with the fiscal 2026 issuance, and it is in the income statement now.12

Table 2. Notes payable and other borrowings, in millions. The total fell in the latest quarter. The increase was fiscal 2026.
Aug 31, 2026 May 31, 2026 May 31, 2025
Notes payable, current7,6257,1997,271
Notes payable, non-current117,712122,34285,297
Notes payable, total125,337129,54192,568

The larger number is not the bond line. At August 31, 2026, Oracle had $288 billion of additional lease commitments, substantially all related to data centers, generally expected to commence between the second quarter of fiscal 2027 and fiscal 2029, on terms of fifteen to nineteen years. They are not on the balance sheet and not in the lease-maturity table. Recorded operating lease liabilities were $34,621 million and finance lease liabilities were $9,185 million. Operating-lease right-of-use assets obtained in the quarter were $4,903 million. Unconditional purchase obligations of a year or more, mostly components for cloud assets and power, were $34,150 million. S&P’s July note used $260 billion of additional lease commitments, the figure then in front of it. The August 10-Q is $288 billion.15

On July 9, 2026, S&P lowered Oracle’s long-term rating to BBB− from BBB, and the short-term and commercial-paper ratings to A-3 from A-2. The outlook is stable. BBB− is still investment grade. It is the lowest investment-grade rung. S&P said adjusted leverage is expected to stay above 4 times for two years, against a BBB downgrade trigger of 4 times, and that another cut is possible if leverage is expected to stay above 4.5 times or if S&P stops believing free operating cash flow turns positive by fiscal 2029. Moody’s, on February 2, 2026, affirmed Baa2 and the P-2 commercial-paper rating and kept a negative outlook. Moody’s said the build should not be funded only with debt and debt-like leases, and that the largest AI infrastructure counterparty is, in Moody’s expectation, OpenAI. We do not describe either action as a loss of investment grade. We do describe the constraint as already binding.56

04 Concentration sits in the backlog

No single customer accounted for 10% or more of Oracle’s revenue in fiscal 2026, 2025, or 2024. The 10-K says the receivable book is diversified and that payment terms on most contracts are short. A short that says Oracle’s current sales are an OpenAI story is contradicted by the filing. The same 10-K then says the thing that matters: Oracle enters large, long-term cloud contracts that require significant infrastructure investment, and the returns depend on customer demand and on key customers meeting their obligations.2

Oracle does not name those key customers in the RPO footnote, and it does not give a percentage of RPO for any one of them. S&P does. In the July 9, 2026 downgrade, S&P wrote that it estimates OpenAI makes up roughly half of the $638 billion of RPO. Half of $638 billion is about $319 billion. Half of the later $664 billion would be about $332 billion if the same mix still held, which S&P has not updated in the note we are using. A round “$300 billion” is the right order of magnitude for that estimate. It is not an Oracle disclosure, and we do not treat it as one. If OpenAI could not pay, S&P’s concern is that Oracle would keep data-center leases it might be unable to exit, or would re-lease them on worse terms.5

The same counterparty shows up, at a disclosed and much smaller size, in CoreWeave. CoreWeave’s June 30, 2026 10-Q says OpenAI committed to pay up to about $6.5 billion through May 31, 2031. Leaked 2025 OpenAI accounts, obtained by Ed Zitron and reviewed by the Financial Times, show revenue of $13.07 billion and an operating loss of $20.92 billion. We have not inspected those statements. They are the same figures used in the CoreWeave note, and they are press reporting. They matter here because S&P has identified that buyer as roughly half of Oracle’s backlog, and because the data-center leases that would serve that backlog are expected to commence over the next three fiscal years and then run for fifteen to nineteen years.79

We rank this fourth, not first. Current revenue concentration is low, so the likelihood that this quarter’s profit depends on one lab is low. The likelihood that the backlog does is medium-high, because a rating agency has said so and Oracle has not published a rebuttal figure. The impact, if that estimate is right and the customer stumbles, is severe. It is the scenario that turns the lease commitments into a problem the software franchise cannot carry.

05 Shareholders are writing the check

The August quarter was financed with stock. Oracle issued 141 million common shares under its at-the-market program for net proceeds of $19.9 billion and fully used the $20 billion program. Common shares outstanding were 2,880 million at May 31 and 3,024 million at August 31. The 10-Q cover page counts 3,023,736,000 common shares at September 7, 2026. There was no common-stock repurchase in the quarter. A year earlier the company repurchased 0.4 million shares for $93 million. About $6.3 billion remained authorized. The authorization has no expiry, and the 10-Q says the pace depends on capital spending, the dividend, and debt.1

The dividend was not cut. The board declared another $0.50 per common share in September 2026, payable October 23. In the August quarter the company recorded $1,484 million of common dividends and $81 million of dividends on the 6.50% Series D mandatory convertible preferred. Cash paid to stockholders was $1,565 million. Preferred stock outstanding had a carrying value of $4,954 million, from the February 2026 issue. Free cash flow was $(5,396) million in the same quarter. The dividend is being paid. It is not being paid out of free cash flow.1

That is the shareholder version of the same gamble. Capex and the lease commitments are contracted or already spent. The revenue that would repay them is, on the company’s own schedule, weighted to the end of the decade. Equity raised at the market, a preferred that converts later, and a dividend that continues, are how existing holders fund the wait. S&P said it expects further equity if credit metrics threaten the investment-grade rating. We rank dilution behind the cash, the timing, the leases, and the customer, because issuing stock is the relief valve. It is also how the relief valve works: the share count goes up.

06 What the price assumes

We do not publish a price target. The table is arithmetic on figures we can cite. It is not a forecast of where the stock trades if the backlog slips.

Table 3. Valuation arithmetic. The share count is the September 7, 2026 cover figure. The price is the September 18, 2026 close, the last trading day before this note. Preferred stock is included at carrying value, not a market price.
ItemFigureHow it is derived
Common close, Sep 18, 2026$147.61Yahoo Finance historical prices8
Common shares, Sept 7, 20263,023.736 millionForm 10-Q cover1
Common equity value at the close$446,334 million3,023.736 × 147.61
Preferred stock, carrying value$4,954 millionAugust 31, 20261
Notes payable and other borrowings$125,337 million7,625 + 117,712
Cash and marketable securities$37,077 million36,369 + 708
Illustrative enterprise value$539,548 million446,334 + 4,954 + 125,337 − 37,077
Oracle stockholders’ equity, ex-preferred$61,818 million66,772 − 4,954
Book value per common share$20.4461,818 ÷ 3,023.736
Close ÷ book value7.2×147.61 ÷ 20.44

The enterprise-value figure leaves out the $288 billion of leases that have not commenced. Putting that commitment in at face value would more than double the claim ahead of the common stock, and it would also double-count whatever those leases earn if the RPO is collected. We leave it out of the arithmetic and in the thesis. Restricted cash, if any remains inside other lines, is not broken out as a separate cash deduction beyond the cash and marketable securities on the balance sheet. The share count is eleven days older than the price.

At $147.61 the common stock is about 7.2 times book value per share, on a quarter in which company-defined free cash flow was negative $5.4 billion and cloud infrastructure, the line growing at 121%, was still 38% of revenue. Fiscal 2026 revenue was $67.4 billion. An illustrative enterprise value of $540 billion is about 8.0 times that revenue. The price assumes the $664 billion of RPO is delivered on the company’s schedule, that the customer S&P is pointing at can pay, and that the leases commencing through fiscal 2029 earn more than they cost. The software franchise is what keeps that from being a CoreWeave-style residual. It is also what the build is spending.

07 Risks to the short

Oracle is profitable. Operating income of $6.7 billion in a quarter covers $1.4 billion of interest with room. Net income was $4.8 billion. No customer is 10% of reported revenue. The database and application support business is the asset a short seller can be wrong about for a long time, because it throws off cash while the cloud build is underway.12

Customers are already paying ahead, and some are supplying the hardware. The $75 billion figure at year-end, and the company’s statement that the quarter’s new AI contracts do not change its capital-raising plans, are the bull case in the company’s own words. Cash and equivalents were $36.4 billion. The at-the-market program raised $19.9 billion without a failed deal. S&P’s outlook is stable, not negative, and S&P says Oracle is trying to stay investment grade. Capacity is not only a slide: the company reported 850 megawatts delivered in the quarter and more than 300,000 GPUs since the prior quarter.1345

The OpenAI share of RPO can be wrong. It is S&P’s estimate. If the backlog is more diversified than that, reason 4 shrinks. If the six-year wait turns into collected revenue and the leases earn their cost, the fiscal 2027 cash deficit is the trough S&P describes, not the destination. The short is wrong if Oracle delivers the capacity, collects the RPO, funds the gap with a mix of prepayments and equity the market will take, and reaches positive free cash flow by fiscal 2029 without a break in the credit rating or the customer. S&P’s base case is close to that. We think the equity is paying for it as if the base case were the only case.

08 Other considerations

Project Jupiter is real, and it is not in the 10-Q. Reuters, on September 18, 2026, reported that about $18 billion of loans tied to an Oracle-leased data center in Doña Ana County, New Mexico, were quoted at 89 to 91 cents on the dollar. Reuters identified the campus as the 1,400-acre Project Jupiter development and described it as part of Oracle’s agreement to provide computing capacity to OpenAI. We searched the August 31, 2026 10-Q and the fiscal 2026 10-K for the project name and did not find it. The loan price is a press report, not a company-reported loss. It belongs here as an example of the syndication risk already visible in the $288 billion of leases that have not commenced. It is not a separate pillar of the short.10

What we corrected in the draft. Borrowings are $125.3 billion, and they fell in the quarter. They are not a balance-sheet figure above $130 billion. Interest is up 55%, and that increase follows fiscal 2026 issuance more than a bond deal in the latest quarter. Current revenue is not extremely concentrated. The OpenAI share of RPO is S&P’s “roughly half,” not an Oracle line, so a flat $300 billion is an order-of-magnitude reading of that estimate. Oracle does not disclose an initial gross margin on new AI capacity. The dividend was maintained, not cut. Buybacks did stop.

CoreWeave, beside this. CoreWeave is the leveraged version of the same build: $35.6 billion of scheduled principal, interest larger than operating income, and a named OpenAI commitment of up to $6.5 billion. Oracle is the investment-grade version: a software profit that still covers the coupon, a backlog an order of magnitude larger, and lease commitments that dwarf CoreWeave’s debt. Applied Digital’s May 31, 2026 10-K and Core Scientific’s June 30, 2026 10-Q show the landlord side of this trade, including CoreWeave as a tenant. Oracle is the tenant and the cloud operator at the other end of the stack. The industry risk is shared. The equity cushion is not. That is why this note is a short on the price of the cushion, not a prediction that Oracle fails to pay its bonds this year.9

09 Notes

  1. Oracle Corporation, Form 10-Q for the quarter ended August 31, 2026. Unless a sentence is footnoted elsewhere, the quarterly figures in this note are from that filing. sec.gov.
  2. Oracle Corporation, Form 10-K for the year ended May 31, 2026. Fiscal 2026 free cash flow, capital expenditures, borrowings, revenue, cloud infrastructure revenue, and the customer-concentration sentence are from that filing. sec.gov.
  3. Oracle Corporation, Exhibit 99.1 to the Form 8-K filed September 10, 2026, first-quarter fiscal 2027 earnings release. The 850 megawatts, the GPU deliveries, the $30 billion of new AI contracts, and the statement that those contracts do not change capital-raising plans are from that release. sec.gov.
  4. Oracle Corporation, fourth-quarter and fiscal 2026 earnings release, June 10, 2026. The $75 billion of prepaid and customer-supplied hardware on large AI contracts is from that release. Oracle Investor Relations.
  5. S&P Global Ratings, “Oracle Corp. Downgraded To ‘BBB−/A-3’ From ‘BBB/A-2’,” July 9, 2026. The OpenAI estimate, the description of fiscal 2027 capex guidance, and the rating action are S&P’s, not Oracle’s. S&P Global Ratings.
  6. Moody’s Ratings, “Moody’s Ratings affirms Oracle’s Baa2 rating on equity announcement; Outlook remains negative,” February 2, 2026. Moody’s.
  7. OpenAI’s 2025 revenue and operating loss are not from an Oracle filing. Quartz, June 16, 2026, on statements obtained by Ed Zitron and reviewed by the Financial Times: revenue $13.07 billion, operating loss $20.92 billion.
  8. Common-stock unadjusted closing price of $147.61 on September 18, 2026, the last trading day before this note, from the historical price table at Yahoo Finance.
  9. CoreWeave figures are from CoreWeave’s Form 10-Q for the quarter ended June 30, 2026, sec.gov, and are discussed in our CoreWeave note. Applied Digital’s contracted lease book is in its Form 10-K for the year ended May 31, 2026. Core Scientific’s CoreWeave colocation contract is in its Form 10-Q for the quarter ended June 30, 2026.
  10. Project Jupiter: Reuters, September 18, 2026, citing the Financial Times, on the loans and the campus.

This note is opinion, not investment advice. Froth City Research and its affiliates may hold positions in securities discussed. Do your own diligence.