Oracle

IThe business

Oracle has three businesses and reports them as cloud and software, hardware, and services. Cloud and software was 88% of revenue over the last four quarters.³ That label hides what matters. Inside it sit an old annuity, a good software business and a new capital-heavy one, and they have almost nothing in common financially.

Revenue in the twelve months to 31 August 2026 was $71.8B, up 22% on a year earlier C (calculated). Operating income was $23.1B and net income $18.9B C (calculated). TTM on this page means those four quarters. Fiscal years end on 31 May, so FY2030 ends in May 2030.²³¹⁹

What it sells

• Software support, $19.8B in FY2026. Customers who bought a database or application license pay every year for updates and help. Oracle says substantially all of them renew. This line has been flat for three years and fell 1% last quarter.²³
• Cloud applications, $15.9B. Fusion for finance and HR, NetSuite for smaller companies, and industry suites, including Oracle Health. Up 10% last quarter, with Fusion up 14%.²⁴⁹
• Cloud infrastructure, $18.1B in FY2026 and $7.4B last quarter alone, up 121%. This is rented computing. It includes AI capacity sold to model builders, and it also includes the cloud database, which management says is growing well above 50%.²⁴⁸
• Software licenses, $4.7B, falling. Hardware, $3.1B. Services, $5.7B.²

The backlog

Remaining performance obligations, or RPO, are contracted revenue not yet earned. They were $46.6B in May 2022, $137.8B in May 2025 and $664B in August 2026.²⁰³ More than 90% is infrastructure.⁸ That is 9.3 times TTM revenue C (calculated). Oracle calls it one of the largest in the industry. It cuts both ways: it is an order book, and it is also a very large promise by a small number of counterparties.

Read that table slowly, because it answers a question the headlines skip. FY2030 runs from month 34 to month 45 of this schedule. The two windows around it average $123B and $113B a year C (calculated). Even if the backlog peaks there, it probably covers something like $115B to $150B of that year E (estimate). Oracle’s target is $225B or more.³⁶ The rest has to come from three places: support, applications and other business that sits outside a long contract, which I put near $55B E (estimate); contracts not yet signed; and any repricing under the cost-plus terms. The target is reachable. It is also more than the backlog alone delivers.

How it got here

For a decade this was a business that returned cash. Revenue barely moved between FY2016 and FY2021, and Oracle spent $101.7B on buybacks in those six years C (calculated).²⁰ The change came in steps. In June 2022 it closed Cerner. In FY2025 capital spending jumped to $21.2B. In September 2025 Safra Catz stepped aside, and Clay Magouyrk, who ran the cloud, and Mike Sicilia, who ran industries, became co-CEOs. In the same month the RPO reached $455B on what press reports described as a roughly $300B agreement with OpenAI, a figure Oracle’s filings do not state.^2015324

The stock went with it, both ways. It closed at $328.33 on 10 September 2025, fell to $114.99 on 24 July 2026 and closed the last session at $130.67. It rose 10.7% on 28 October, the day of the Investor Day, and gave back all but 14 cents of that in two sessions.¹

IIThe moat

The honest way to read this business is in two halves, and they deserve different grades.

The old half

An Oracle database holds the records a bank or an airline cannot lose. Applications are built on it over decades, and moving off means rewriting them and accepting the risk of being wrong. Support customers pay a fixed percentage of what they once paid for the license. That is why $19.8B a year renews with so little effort, and why Oracle spends relatively little to keep it.²

The same holds for the ledger. A company’s finance system has its close calendar, its audit trail and its tax logic inside. Management said on stage that about half of its ERP customers are still on-premises, and that moving one to its cloud brings three to four times the revenue. It also claims AI agents cut migration time by 50% to 70%. I treat those as company claims, but the direction fits the 10% to 14% growth the filings show.⁸⁹

What protects this half is switching cost. What threatens it is slower. Software support fell 1% in the latest quarter, licenses fell 15%, and SaaS growth has settled near 10%.⁴ An owner has to ask what a world of AI agents does to seat-based software. Oracle argues agents run on top of its systems of record. I think that is more likely right than wrong for finance and for the database, and less sure for everything else. I cannot judge this with precision, so my cases keep this half at 4% to 5% growth a year and change little across them.

The new half

The AI halls are a different kind of asset. Oracle does not own land or buildings; it leases the shell and the power, and it buys the chips, memory and network, which it runs for a customer on a contract averaging five years.⁷ Management says effectively all large AI contracts are cost-plus or similar. If memory prices rise, revenue rises and margin holds.⁸ That protects the margin. It also tells you who holds the power in the deal. A cost-plus contract is how a contractor is paid. A franchise sets its own price.

Is there any lasting edge? Three things are real. The network design lets Oracle split a cluster among many tenants or run it as one. It runs 70 multicloud database regions inside other companies’ clouds, so the database goes where the customer already is, and that business grew 353% last quarter.⁹ And a customer that has run on Oracle’s database for 20 years will often buy Oracle’s compute next to it. That last link is the best argument for the new half. It applies to enterprises, though, and most of the backlog comes from model builders, who are the most capable buyers on earth and will move for price when their contract ends.

So the table below is my view. The company earns very good returns on the old half, and on the new half it earns a contract return, which is the subject of the valuation section.

Oracle’s slide shows operating margins closing to a similar level across the three. That is true of a fully ramped project, before the charge for the capital tied up in it. A business that earns 34% on revenue with $2.30 of capital per dollar of revenue is a very different thing from one that earns 34% on almost no capital.⁷

IIIThe competitive advantage

Where Oracle beats its competitors today is execution. It delivered 850 MW of AI capacity in the first quarter, which it says is 73% of everything it delivered in the prior fiscal year, and more than 300,000 GPUs. OpenAI accepts capacity at Abilene within about 24 hours, and availability is above 99%.⁴⁹⁸ GPU utilization was 97.9%.⁹ Oracle has told investors these figures for several quarters, and revenue has followed: cloud infrastructure growth went from 55% to 68%, 84%, 93% and 121% in five quarters.⁴

The advantage on the customer side is also real. Oracle can price a cluster as a site and fit a customer’s own chips into it. It says it runs AMD, Intel, NVIDIA and Ampere processors today, and analysts at the Investor Day asked about its news on hosting Google’s TPUs.⁸ Bring-your-own-hardware deals mean a customer pays for the GPUs and Oracle provides the hall, the network and the operation. In June it said prepaid and customer-supplied hardware in its large contracts totaled $75B.⁵

Where it loses is the price of money. Compare the same dollar of capital spending at three kinds of owner. A hyperscaler funds a data center from its own cash flow and sells the capacity to itself first. A smaller specialist borrows against its contracts. Oracle sits between them, with investment-grade bonds at the lowest rung, a 6.70% coupon on 30-year notes issued in February, and its entire free cash flow going the other way.¹⁷²¹ In a business where the product is similar and the customer negotiates hard, the lowest cost of capital tends to win over a long enough period. That is a structural disadvantage, and I would not expect execution to erase it.

The 10-K states that no single customer accounted for 10% of revenue in FY2026.² That was true of the revenue already earned. The backlog looks nothing like the revenue, and Oracle’s filing adds that in certain infrastructure offerings it is more concentrated among a number of large customers.² Management said it has diversified the backlog in each of the past four quarters and that it has more than 1,000 AI customers.⁷⁸ It does not publish the shares. S&P’s figure of about half is an estimate, repeated widely, and I use it for that reason only. Oracle has never confirmed it.²¹

IVManagement

Three people run Oracle, and one of them owns it. Larry Ellison founded the company in 1977, is Executive Chair and Chief Technology Officer, and holds 38.2% of the shares. Clay Magouyrk, 40, came from Amazon’s cloud team in 2014 and built Oracle’s. Mike Sicilia came through the industry applications. They became co-CEOs in September 2025. Hilary Maxson joined as CFO in April 2026 from Schneider Electric, so the funding plan presented this month is the work of someone six months in the seat.¹²¹⁵¹⁶

How they are really paid

Follow the incentive. The performance options test one number: FY2028 non-GAAP revenue. At $100B, 80% vest. At $125B, 100% vest. At $250B, 200% vest.¹² Oracle’s own plan last October was $130B for FY2028, and this year it guides at least $90B for FY2027.¹⁰⁴ So the full target sits below the company’s published plan, and the threshold sits about 11% above this year’s guidance.

Revenue here can be bought with capital. Nothing in the plan looks at return on that capital, free cash flow, debt or dilution. The bonus is the same: operating income growth, which is untouched by interest cost, and which leaves out stock pay.¹² I am not saying anyone has behaved badly. I am saying that if you wanted a pay plan that pushed managers to build as fast as they could with other people’s money, it would look a lot like this one. The cushion is the option strike. At $308.46 the CEOs do badly unless the stock more than doubles, and that makes them care about the share price, which depends on the funding.

How much of their own money is in

Ellison’s stake is worth about $151B at the close C (calculated). That is alignment on a scale few companies can show, and it is why I take his judgment seriously. There are two details. He has pledged 413M shares, 36% of his holding, against personal loans, and the board says he could repay them without selling.¹² And on 22 June 2026, with the stock at $175.07, he adopted a plan to sell up to 50M shares. On 12 September Oracle announced that he had cancelled it, that none had been sold, and that he has no other plan to sell.³¹³ The stock was at $150.28 on the last close before that. I read the cancellation as a reassurance to the market, and the adoption as evidence that even he wanted some liquidity in June. I do not know why he adopted it.

The others own far less. The proxy lists Magouyrk at 898,530 shares and Sicilia at 467,499, each including options due to vest, and Maxson at none, as she was new. That is about $117M and $61M at the close, against option awards several times larger C (calculated).¹² Directors and officers as a group, excluding Ellison, are under 0.3%. With 38.2% of the vote, Ellison cannot be outvoted in practice. The two directors who retired in January were 86 and 87.¹⁴

What they did with capital over a full cycle

Oracle’s record gives a real test, because it has now done both sides of the trade. From FY2019 to FY2022 it spent $92.6B to retire 1.61B shares, at an average of about $57.50 each C (calculated). Diluted shares fell from 4,238M in FY2018 to 2,786M in FY2022.²⁰ This was a very good use of money for anyone who stayed. In the first quarter of FY2027 it sold 141M shares through the market at about $142.³ Buying at $57.50 and selling at $142 is a fine result on those shares, and the buyers of those shares are now under water.

The other side of that record is the balance sheet it left. The buybacks were partly paid with debt, which rose from $60.6B in FY2018 to $75.9B in FY2022, and shareholders’ equity went negative, at −$5.8B in FY2022.²⁰ Then came Cerner for about $28B in cash and more debt.²⁰ So Oracle entered the largest investment in its history having already spent its cushion. It stopped buying back stock in FY2023. Had it kept half of that $92.6B, the funding question that took the stock from $328 to $115 would be much smaller. Hindsight is cheap, but this is what a full cycle means.

Oracle has cut, too. Employees fell from about 162,000 to about 141,000 in a year, and the restructuring plan has grown to about $2.8B.¹⁸²³ Management’s stated order of priorities is the investment-grade rating first, the dividend second and growth third.⁷ The dividend costs about $6B a year while the company raises outside money, so in practice part of it is paid with borrowed cash. Also, in June it said it did not expect to issue more debt in calendar 2026, and in October it said the $10B of new debt could come as early as calendar 2026.⁵⁸ A small change, and worth watching in a management team asking for trust on a much larger number.

The promise made on 28 October is checkable, which I give them credit for: $20B of new capital and no more through FY2030. Maxson called that a real commitment.⁸ The record will be whether it is kept when a large new contract arrives.

VFinancials

Every number on this page starts from three facts about the last twelve months. Oracle earned $18.9B. It collected $46.9B of operating cash. It spent $75.7B on equipment.²³

Revenue grew 6.2% a year over the decade, and 10.5% a year since FY2023 C (calculated). Operating margin was 34.0% in FY2016 and 30.6% in FY2026 C (calculated). The last line is what used to make the story work: EPS rose 10.9% a year from FY2016 to FY2026 C (calculated), with a third fewer shares doing it.²⁰

Cash, and why reported free cash flow misleads in both directions

The reported −$28.7B overstates the damage, because most of that spending is growth. It also understates it, because operating cash flow includes $16.0B of prepayments that are really customers lending Oracle money against future service.³⁴ They count as cash now and revenue later, with no cash then. On the balance sheet, deferred revenue doubled in one quarter to $30.8B.³ So I rebuild the number.

What an owner can really take out

Ask one question: if the building stopped today, how much cash could the owner carry away each year and leave the business as it stands? Take operating cash flow, remove the prepayments, charge stock pay as the cost it is, and charge depreciation as the floor for replacing what wears out.

So reported earnings of $18.9B are not far from what the business could pay out if it stopped growing, $16.7B. What it actually pays out is nothing. Outside money covers the dividend and the build. The stand-still figure uses a $9.4B depreciation that lags: on the latest quarter’s run rate it would be $12.6B C (calculated). If that is the true cost of standing still, the number is closer to $13.5B E (estimate). I show $16.7B as the top of a range I cannot narrow further.³

Debt, leases and what sits off the balance sheet

The line that matters most is the one that does not appear on the balance sheet. Oracle has signed $288B of leases on data centers, for 15 to 19 years, to serve contracts averaging five years.³⁷ That is about $17B of rent a year once they start C (calculated). It went up $28B in a single quarter. The contract with the customer ends long before the contract with the landlord.

Interest cover is fine today. EBITDA was about 6.6 times TTM interest of $5.1B, against the 3.0 times in the new $10B credit line C (calculated).²³ Interest expense rose 55% on a year earlier in the last quarter, to $1.4B.³ Bondholders have already moved: the notes were at 89 cents in May and 85 cents in August on the 10-Q’s fair-value disclosure C (calculated).²³

VIValuation

Start with what would ruin it

Before any value, I want to know how an owner loses most of the money here. Three ways stand out, and the valuation is built backward from them.

First, the largest customer cannot pay.

S&P estimates OpenAI at roughly half of the $638B of RPO it measured in May, as the press reports it; Oracle does not state the share.²¹ Applied to August's $664B, half is about $330B C (calculated), and NPR reports a promise of about $300B over five or so years.²⁴ The Financial Times reported OpenAI's annualized revenue at about $50B at the end of September, a figure Eisman cited.²³ A presentation reported by the Financial Times shows $856B of compute and $278B of negative free cash flow from 2026 to 2030, on the assumption that revenue reaches $350B in 2030. It raised $122B in March, which the FT reports would last until 2028.²⁵ These are press accounts of a private company, and I cannot audit them. The plain arithmetic is that one customer with $50B of revenue owes one supplier something like $60B a year at peak E (estimate), and that same customer has commitments to several others. If it renegotiates, Oracle holds the rent.

Second, the funding door shuts before the cash turns.

The worst stretch is the next 18 months. Press accounts say S&P cut Oracle to BBB- on 9 July, with a stable outlook, and said it could cut again if leverage stays above 4.5 times on its measure, or if Oracle is not on a path to positive free operating cash flow by FY2029.²¹ That is secondhand: I read it in the press and did not open S&P's report. Moody's and Fitch also rate Oracle, and I left their ratings out because I did not open their own releases. A cut below investment grade would raise the cost of debt, and the 10-K warns it could affect data center leases and bring collateral calls.² Oracle’s $20B cap depends on vendor financing, prepayments and customer-owned hardware doing about half the work.⁷ Those three depend on suppliers and customers that have their own balance sheets to mind.

Third, a five-year contract meets a nineteen-year lease.

At year five the customer can leave. The hardware is old, the rent runs for another decade or more, and a rival may have newer chips in a cheaper hall. Management argues the opposite: in its view, today’s data centers will look cheap in five years, because replacing them will cost more.⁸ I think that is plausible for power and shells, and less sure for GPUs.

Each of the three is a way to lose most of the equity, and they share a cause, so they are likely to arrive together. Anything that hurts OpenAI also hurts the funding and the second contract. Keep that in mind when the numbers below look tidy.

What a gigawatt really earns

On 28 October Oracle published the economics of one gigawatt, which is the most useful page it has released in years. About $80B of contract value over five years, $37B of capital spending in the first two years, $16B of revenue a year once ramped, 35% gross margin, 34% operating margin, and what it calls a return on investment of about 28%.⁷

That 28% needs reading. The footnote defines it as after-tax operating profit plus depreciation, divided by capital spent. Adding back depreciation means counting the return of your money as a return on it. On those figures, after-tax operating profit is about $4.5B a year on $37B, or 12.1% C (calculated). The 28% is a cash payback rate: $10.4B a year pays back the $37B in about 3.6 years C (calculated). Management’s reason for adding depreciation back is that upkeep spending is limited until the hardware is replaced, which is true, and that is exactly the point. The replacement comes.⁷⁸

Here is the simple version. The first contract alone returns about 9% E (estimate). That is above what Oracle pays on its 30-year debt and below what I ask of the equity. So the whole question for an owner is the sixth, seventh and eighth years. If the hardware earns for six years the project clears my 11.0% rate. Eight years gives 17.2% E (estimate). The best evidence is on Oracle’s side: it said GPUs up for renewal in the first quarter, the majority four years or older, renewed at a 20% premium.⁹ The evidence is thin, though, since those are small fleets from a shortage, and the first very large contracts do not come up until after 2030. Nobody can know this yet, so I show the range instead of one number.

What the price implies

At $130.67 my base case returns 10.3% a year, a little below the 11.0% I require. The bear returns 1.1% and the bull 15.4% E (estimate). I compare each with my own rate first and the 4.94% ten-year Treasury yield second.¹²⁸

The 6.2 times is the number bulls quote, and it deserves an answer. Earnings in FY2030 would come from hardware bought in FY2027 and FY2028, in contracts that end around FY2032 to FY2033. They carry a depreciation charge that understates the next cycle’s bill if prices rise, and they sit on debt and leases that come first. A multiple of 6 on those earnings is low. Compare it with the 4.0% stand-still yield today and the picture changes: the buyer is paying for four years of a build that has to go right, and then a low multiple on what comes out. If the $21 arrives on time, the price will have been low for those earnings, and my bull value of $195 reflects that. On about $8B of yearly interest and today's diluted share count, my bull earnings come to about $20.90 a share in FY2030 and $16.00 in FY2029, which match Oracle’s published path E (estimate).¹⁰ So the disagreement is about the odds of $21, and only a little about what it is worth.

The reverse DCF starts from the $16.7B of stand-still owner cash. First I take off $150B, the present cost of what my base spends above depreciation through FY2029. Then I solve for one growth rate for ten years, at 11.0%, with 3.0% growth after that. The price needs 18.0% a year C (calculated). Stand-still cash grew 2.4% a year from FY2016 to FY2026 C (calculated), and my base grows 14.7% a year from the same start E (estimate).²⁰³

The discount rate

I use 11.0%, above the 10.5% on Broadcom and Amazon and the 10.0% on Alphabet. Oracle’s beta against the S&P 500 is 1.29 on ten years of month-end returns, 1.78 on five and 2.94 on two C (calculated). The capital asset pricing model, with the 4.94% ten-year Treasury and Damodaran’s 4.20% implied premium, gives 10.4% to 12.4% on the first two.²⁸²⁹³⁰ The recent number measures a stock caught in a funding scare, and I give it little weight. I use one rate on unlevered cash flows and deduct net debt. Oracle’s capital is about 80% equity at market, so a cost of equity near 12.4% and after-tax debt near 5.7% blend to about 11% E (estimate). One point on the rate moves the base by 17% to 23%, so treat this as the largest single judgment on the page.

The three cases

How I built the base. Oracle says about 10 GW by the end of FY2028, of which five are the large sites.⁷ Fully ramped at about $16B a gigawatt on its own hardware and $11B on a customer’s, against the roughly $9B of revenue per gigawatt that today’s fleet earns E (estimate), that supports something like $120B to $140B of infrastructure revenue E (estimate). Oracle’s $225B needs about $166B, the figure it gave for FY2030 last year, because the other businesses add up to near $59B E (estimate).¹¹ So the headline number needs $16.6B on every gigawatt in the fleet, older ones included, or more gigawatts than the 10 C (calculated). Maxson said some additional growth would come mainly after FY2028.⁸ My base has $190B for FY2030, and it reaches $274B in FY2036.

The margin check is reassuring. In the bull, $58B of other business at a 55.5% margin and $167B of infrastructure at 34.5% give 39.9%, which is the 40% target, so that target needs nothing heroic from the operating lines E (estimate).⁶ In the base I hold infrastructure near 31% in FY2030 because I expect the later contracts to be tighter and some hardware to be idle during ramp-ups. Last quarter’s overall figure was 42%, and my base is 41% for the year and 37% to 38% from FY2028.⁹

The funding check is harder. From 1 September 2026 to 31 May 2028 my base burns $68B before interest. Add about $10B of after-tax interest and $11B of dividends and the need is near $89B E (estimate). Sources: about $22B from cash, taking it to $15B, and the $20B of new capital. That leaves about $47B for vendor financing and prepayments E (estimate), which fits inside the 50% cut Oracle describes on about $220B of spending.⁷ So the plan works on my base. It has limited room if prepayments stop, since one quarter’s prepayment was $11.4B and I cannot see how much more is contracted.³

A word on $30B. Before the event, data compiled by Bloomberg and FactSet showed analysts expecting this measure to stay deeply negative, so the target was a surprise.²⁷ Oracle’s target is operating cash flow less capital spending. Its slide says vendor financing principal is a separate payment; the footnote places the interest in operating cash flow.⁷ If that principal sits in financing, as similar short-term items do today on its cash flow statement, the $30B is before those repayments, and before stock pay of about $6B E (estimate).³ My base gets to about $23B for FY2029 on Oracle’s definition, and $17B after stock pay and interest E (estimate). I cannot size the vendor repayments, because Oracle would not give the split. At one third of about $110B over five years they would be near $7B a year E (estimate).⁸

Shares are the 3,023.7M outstanding on 7 September, plus 34M for stock awards, 31.2M for the mandatory convertible at its maximum rate, and 76.5M for the $10B of equity at today’s price.³² Operating leases are not deducted, because the rent is in the margin. The $288B of unstarted leases are not deducted either, for the same reason; the cases pay that rent as they use the space. In the bear, some of it is rent on idle halls, which is why the infrastructure margin falls to 12%.

How the odds shrink

The strongest case for this stock needs several things to go right at once. Multiply them out, with my own rough odds.

Four good-odds bets multiply to about one in five. That is why I weight the bull at 20%, the base at 45% and the bear at 35%, which gives $109 E (estimate). The price, as noted, pays for 57% on the two-outcome view. The odds are mine, and a reasonable person could say 35% for the bull and 20% for the bear. Do that and the weighted value is about $132, about the price.

A stress with no probability

The bear keeps OpenAI as a partial payer. The stress removes it. Revenue from it stops in FY2029, infrastructure falls to about $50B in FY2030, and the margin there goes below zero for a year because the rent continues. Oracle cuts capital spending to $10B and lives on the older business. It is worth about $33 a share E (estimate). It has no weight in the value. It also leaves out what such an event would do to the rating, the refinancing of $10.1B due in FY2028 and the terms of the leases, which I cannot model. The equity could be worth less. The older half of the business is the reason it is worth something: about $50B of revenue at a margin near 50% E (estimate).²

My margin-of-safety price

My margin-of-safety price is $70, 43% below my base value. A flat 30% does not suit a company where the outcomes are this wide and debt sits on top. So I set it as the price at which the stress case, with the largest customer gone, still returns the 4.94% Treasury yield E (estimate). At $70 the bear returns 7.1% and the base 16.4%. The price sits 87% above it. The lowest close in the past year was $114.99.¹

The next best use of the same dollar

A dollar in this stock has to beat what else it could do. A Treasury pays 4.94% with no risk.²⁸ Oracle’s own bonds are a better comparison than people admit. The 30-year notes issued in February carry 6.70%, and the book’s fair value has fallen since, so the yield today is higher, with a claim ahead of the shareholder.¹⁷³ My base gives the shares only about 3.6 points more than that coupon, and my bear gives about 5.6 points less.

Alphabet is the best comparison on this site, since it rents out the same capacity from its own cash. On my pages its base returns about two points less than Oracle’s, and its bear about four points more. Microsoft’s base is about 7.5%. Oracle pays you two points for a much wider spread of outcomes. Whether that is a fair trade depends on how you feel about the bear.³³³⁴

Quality and price, separately

These are two questions. The quality of the older half is high. The quality of the new half is unproven: a low-margin, capital-heavy contract business with one very large customer and a cost of money above its peers. Taken together I would call this a good business carrying a bet that may or may not be good. The price is a different matter. Against my base it is close, 7% above. It is not the clear overpayment that several pages on this site show, and it is far below where it was. It is also well above the point where you are paid for the risk of the bear. The market may well be right at $130.67, since a price near my base is what I would expect if it had read the same documents.

What I cannot judge

• OpenAI’s ability to pay. I have press accounts of a private company, and they have already been revised. It is the largest input on this page.
• The share of the backlog that is OpenAI. S&P’s estimate is the only number, and Oracle does not give one.
• GPU life. Nobody has run these chips at this scale for eight years.
• How vendor financing appears in the accounts. Oracle’s first quarters of doing this at scale will show it.
• What AI agents do to seat-based software by 2032.

Where this could be in 5 and 10 years

The base stands at five years if the build lands about three quarters late and revenue reaches $190B in FY2030 at a 30.5% infrastructure margin. By May 2031 the heavy spending is behind Oracle, and the base earns about $68B a year of owner cash. That is why it rises from $122 to $201. It stands at ten years if revenue reaches $274B and Oracle keeps spending about $74B a year to keep the capacity current. Cash after FY2036 is 61% of the base value, so year ten rests on a long tail.

The bear and the bull differ on the largest customer and on margin. In the bear, OpenAI pays for only part of what it signed, capacity is re-let for less, and the infrastructure margin falls to 12%. The bear still rises to $73 by 2031, because spending ends in 2029. In the bull, Oracle hits its own targets, $225B in FY2030 at 34.5%. All three use an 11% discount rate and 3% growth after FY2036.

The most likely way the base is wrong is that the largest customer pays less, or later, than I assume. Revenue and margin would both slip, and the answer would move toward the bear.

The return column uses the price of $130.67, the owner cash and the year-ten value, so it differs from the returns above, which use unlevered cash. These are estimates of value, not forecasts of the share price. The market price can sit above or below value for years.

VIIThe case for the price

What most people miss

Three facts cut in Oracle’s favor. The first is that the new contracts are lighter than the old ones. In the first quarter it added more than $30B of AI contracts, and management says the vast majority came through prepayment or customer hardware, so they need no extra capital from Oracle. A customer-hardware gigawatt takes about $6B of capital against $37B, on $11B of revenue and a margin at least as good.⁴⁹⁷ The second is the 20% renewal premium on four-year-old GPUs. It is early, and it goes straight at the biggest doubt about the business.⁹ The third is the cost-plus term. Revenue per gigawatt went from about $10B to $16B in a year, largely on memory prices, with the margin kept. Oracle has passed through that inflation.⁸

The bull case

The case here comes from the company and its brokers. Maxson says the $20B completes the funding, that free cash flow turns positive in the second half of FY2028, and that the mechanics are in place rather than a forecast.⁷⁸ Magouyrk says the data centers coming online now will look very cheap in five years, and that memory and chip supply cap the industry’s build, so a glut is unlikely for three to five years.⁸ After the event, Baird’s Rob Oliver called it “very positive” with a $215 target. Oppenheimer’s Brian Schwartz, at $275, said it “somewhat dents the credit risk concerns”. RBC’s Rishi Jaluria, at $165, called it “incrementally positive rather than thesis-changing”. All three come through TipRanks, so they are secondhand.²⁶ On the other side, Steve Eisman said on 16 October that “half of Oracle’s backlog is just OpenAI”, also secondhand.²³

Against reality

If Oracle delivers 7 GW, OpenAI pays and the funding lasts, $225B and $21 follow, and my model values that at about $195. That is possible. Abilene shows the work can be done: six of eight buildings delivered, and 121% growth. Today’s numbers are a long way off, though. TTM revenue is $71.8B and must triple in under four years. Four sites have yet to deliver their first service, and New Mexico still needs an air permit.⁹⁸ Free cash flow must swing about $59B from TTM −$28.7B to FY2029’s $30B C (calculated). A customer with $50B of revenue must fund commitments many times that.²³²⁵

What a 10x would take

Ten times the price is $1,307 a share, or about $4.1T of diluted market value C (calculated). At 20 times earnings that needs $207B of net income. At a 30% net margin it needs about $690B of revenue C (calculated). My bull has $372B of revenue and $110B of after-tax operating profit in FY2036 E (estimate), so a 10x by then would need nearly twice my bull’s earnings. In the ten years to October 2026, split-adjusted, Apple rose 12.1 times and Microsoft 8.7 times, and both were larger then than Oracle is now; NVIDIA rose 115 times from a much smaller base.³² Oracle rose 3.4 times.¹ It has happened, a few times, to the best companies in history. For a business with this much debt and one customer of this size, inside ten years, I regard it as close to impossible.

What they have to hit

Where I come down

The Investor Day moved me on funding: a hard $20B and a published gigawatt are more than I expected, and they took the funding failure out of my base. It did not move me on the customer or on year six, and those decide the value. Better disclosure of OpenAI’s share, or two more years of renewals at a premium, would raise my values more than any quarter of growth could.

VIIIRisks, and what would change my view

The valuation section covers the three ways this goes badly. These are the others, and the signals I will read.

• Delays and permits. New Mexico needs an air permit for fuel cells, and Wisconsin’s power is being arranged through the grid. Oracle’s targets allow for up to a couple of quarters on any site. Costs mostly move with the build.⁹⁸
• Supply. An analyst at the Investor Day described Oracle as heavily weighted to one chip supplier, and Magouyrk would not discuss any supplier’s financing. Purchase commitments went from $13.3B to $34.2B in a quarter.⁸³²
• Prices for compute. Short-term prices are high now. Analysts at the Investor Day cited 40 GW of industry supply in 2027 and 60 GW in 2028. Magouyrk answered that memory and chip supply will cap what really gets built.⁸
• Stock pay and dilution. 141M shares were issued in one quarter, 5% of the count, on top of $4.8B a year of stock pay, and 77M more are implied by the $10B plan at today’s price C (calculated).³
• Key people. Ellison is 82, and a co-CEO structure is new. His pledged shares tie part of his finances to the stock.¹²
• Litigation. A securities class action over statements about the cloud business names the company and executives; a motion to dismiss would be briefed by 18 December 2026.³
• Government and contract risk. The company holds 15% of the TikTok USDS joint venture and sells to governments across many countries.³

What would raise my base value: published proof of a diversified backlog, renewals that hold, and an extra gigawatt or two funded from cash. What would lower it: a restructured contract, new capital above the cap, or a first renewal at a discount. A lower price would not change my base. It would change the answer to whether the bear is paid for.