What we said in September — and what happened
Ten claims graded, two of them against ourselves
On 27 September we published Oracle at $137.10 under the title "Too Hard — Half the Backlog Is One Customer". We said the database business was a toll road and the AI cloud a leveraged bet on one customer's ability to pay; that the shares were not expensive but could not be priced; and that around $100, where we calculated the data centres would come at about 60 cents for each dollar they cost, we would look again. The shares closed at $115.06 on 17 November. Here is the account.
| What we said on 27 Sep | What happened by 17 Nov | Grade |
|---|---|---|
| 'The pace of financing surprises investors each time' | It happened again, more starkly. On 28 October Oracle committed to limit new capital raises through fiscal 2030 to $20bn; on 2 November it raised the whole $20bn in one day. | ✓ Right |
| 'A decade of buybacks reversed' | 86.2 million more shares sold on 2 November at $116. Shares outstanding: 2,880m in May, about 3,110m now, +8%. | ✓ Right, and worse |
| FCF −$23.7bn in FY26, worse in FY27 | Confirmed, and the first quarter's −$5.4bn was flattered: $11.4bn of operating cash flow was customer prepayments that carry a financing charge. Without them, about −$16.8bn. | ✓ Right |
| 'Debt including leases: $156bn' | Understated. Counting all leases on the balance sheet, $169bn at 31 August; about $179bn after the November notes. Our figure left out current and finance leases. | ✗ Too low |
| '$162bn of property; AI cloud priced at ~1.3× cost; ~0.6× at $100' | Our error. $162bn included $34bn of leased assets; owned property and equipment was $127.8bn. On the right figure, the AI cloud was priced at ~1.7× cost at $137, and would be at ~0.9× at $100, not 0.6×. Part XI redoes the sum. | ✗ Wrong input |
| '1.9% above the 52-week low' | Wrong, inherited from our feed: the 52-week closing low then was $114.99 (24 July), so $137.10 was 19% above it. | ✗ Wrong |
| 'Roughly half of RPO is OpenAI (as reported)' | Still not disclosed by Oracle. Its order book was $664bn at 31 August; the next figure comes with the second-quarter results, expected around 10 December. | ◆ Untestable |
| 'Chips outlive contracts badly' | Early evidence: Microsoft is reported to be winding down a 2023 deal to rent Oracle capacity built on an older generation of NVIDIA chips (Bloomberg, 6 November). Oracle says demand 'remains strong'. | ◆ Early signs |
| Founder 'owns two-fifths' | Incomplete. The proxy filed two days before our report shows 38.2%, of which 413 million shares are pledged against personal borrowing, and a June plan to sell up to 50 million shares, cancelled on 12 September with none sold. | ◆ Missed detail |
| '$100: we would look again' | Not reached: the low close since was $115.06 (17 November), the intraday low $109.21 (6 November). The level itself needs correcting; see Part XI. | ◆ Not reached |
The grade that matters most is the one against ourselves on the property figure. Our September case for $100 rested on an arithmetic claim: that at that price the database business would pay for most of the share, and the data centres would come at about sixty cents on each dollar they had cost to build. The claim used a number from our own data feed that included leased assets with the ones Oracle owns. With the correct figure, $100 would buy the data centres at about ninety cents on the dollar. That is a discount, but not the one we promised. We correct it here, publicly, and we move the level.
★★ The cap that lasted five days
Investor Day's promises, and what Oracle did the following Monday
On 28 October, at its investor day in Las Vegas, Oracle set out the plan its shareholders had been asking for. Revenue of at least $225bn in fiscal 2030 (from $67bn in fiscal 2026), adjusted earnings of at least $21 a share, free cash flow positive in the second half of fiscal 2028 and at least $30bn in fiscal 2029, self-funding from fiscal 2029, and net debt below 3.5 times adjusted EBITDA by then. And it made a promise about money: it "committed to limiting additional capital raises through FY30 to $20bn", half in shares and half in debt, excluding refinancing. The shares rose 10.7% that day, to $144.54.
The next Monday, 2 November, Oracle raised all of it. It sold 86.2 million new shares at $116, 11% below the previous Friday's close, and $10bn of notes at coupons of 7.00% to 8.65%, between 2.4 and 3.35 percentage points above US Treasuries. The shares fell 8.8%.
| What Oracle promised on 28 Oct | Where it stands on 17 Nov |
|---|---|
| New capital raises through FY30 capped at $20bn | Used in full on 2 November. Anything else must come from leases, vendor financing, customer prepayments or refinancing, all outside the cap and partly outside the debt line. |
| Gross debt broadly 'flat' after the new $10bn | Pro forma borrowings $135.3bn. Maturities of $7.2bn in fiscal 2027 and $10.1bn in fiscal 2028 must be refinanced. |
| Positive free cash flow in H2 FY28; ≥$30bn in FY29 | No interim free-cash-flow guidance for fiscal 2027. Capital spending of $90–95bn this year is reaffirmed. |
| Investment-grade ratings first, then the dividend, then growth | Ratings unchanged since July: Moody's Baa2 (negative), S&P BBB− (stable), Fitch BBB (stable). New bonds priced like a stressed investment-grade credit. |
| FY30: revenue ≥$225bn, EPS ≥$21, operating margin ≥40% | Unchanged; the first test is the second-quarter report, expected around 10 December (date not yet confirmed by Oracle). |
Two ways to read it. The charitable reading is that Oracle told investors exactly how much it would raise and then raised it promptly, while the window was open, removing the uncertainty in one stroke. The other reading is that a company which needs to say how little it will raise, and then raises all of it within a week at a discount and at 8% coupons, is telling you how expensive its money has become. We hold both readings. What we take from it is narrower: promises about funding from this management should be judged by what is done, not by what is said, and the capital markets have started to charge Oracle like a borrower rather than a software company.
★ The debt you see, and the commitments you don't
$288 billion of leases not yet started, and cash flow helped by customers' prepayments
An owner of Oracle has to read three numbers that do not appear in the usual debt figure. The first is the leases signed but not yet started: data centres that developers are building for Oracle to rent for 15 to 19 years, from this quarter to fiscal 2029. They were $260bn at the end of May and $288bn at the end of August, $28bn more in a single quarter. When they start they will join the balance sheet as lease liabilities, and their rent will be a fixed cost whether or not the tenants underneath Oracle pay.
The second is customer prepayments. Oracle's operating cash flow of $23.1bn in the first quarter included $11.4bn of payments from customers that Oracle treats as carrying a significant financing component: in effect, customers lending Oracle money against future service, with interest accreting on it. That is a sensible way to fund a build, and better than a bond. But it means the reported cash flow was about $11bn better than the cash the business itself produced. Free cash flow without those prepayments would have been about −$16.8bn for the quarter, not −$5.4bn.
The third is the projects financed by others. Developers are borrowing tens of billions to build Oracle's sites, with Oracle as the tenant: reports put the loans at about $38bn for sites in Texas and Wisconsin, $16bn in Michigan and $18bn in New Mexico. None is Oracle's debt. All of it depends on Oracle's rent. In September Oracle sent a force-majeure notice to the developer of the New Mexico project over delays in getting power, reportedly to defer rent rather than to leave.
What the rating agencies watch. S&P has said, as reported by Reuters, that it would downgrade if leverage stayed above 4.5 times; Fitch's threshold is reported at 3.5 times EBITDA, or 4 times including rents. On our feed's figures, cleaned of a one-off gain, net debt is about 3.5 times EBITDA before the leases still to start. The next rating action is the single most important number for the shareholders, because it decides what Oracle's next dollar of debt costs.
| Agency | Rating and outlook (2 Nov 2026) | Reported trigger for a downgrade |
|---|---|---|
| Moody's | Baa2, negative | Leverage could 'approach 5×' temporarily (as reported by Reuters) |
| S&P | BBB−, stable (one notch above junk) | Leverage above 4.5× for long; S&P sees a peak near 4.4× (Reuters) |
| Fitch | BBB, stable | Above 3.5× EBITDA, or 4× including rents (Reuters) |
| The market | New notes at 7.00–8.65%, +240 to +335bp | Five-year credit-default swaps hit 215bp in August, an 18-year high (Reuters); no November figure found |
What moved the shares
From $137 to $115 in seven weeks, and the intraday low of $109
| Date | What happened | ORCL that day |
|---|---|---|
| 24–25 Sep | Force-majeure notice to the developer of the New Mexico site over power delays (Reuters). | $137.10 |
| 29 Sep | A director, Stephen Rusckowski, buys 25,000 shares at $139.35, the only open-market purchase by an insider. | $137.79 |
| 14 Oct | A hot September inflation report lifts bond yields; worries about the New Mexico project. | −8.8% |
| 28 Oct | Investor Day: fiscal-2030 targets, free-cash-flow timetable and the $20bn funding cap. | +10.7% |
| 2 Nov | All $20bn raised: shares at $116 and notes at 7.00–8.65%. | −8.8% |
| 4 Nov | Oracle reported (FT) to have cancelled a five-year chip lease worth about $7bn with Tencent after US officials raised concerns. | −1.9% |
| 5 Nov | Market-wide sell-off in AI shares after the midterm elections. | −6.9% |
| 6 Nov | Microsoft reported (Bloomberg) to be winding down its 2023 deal to rent Oracle capacity on older chips; intraday low $109.21. | $116.10 |
| 11 Nov | CoreWeave's results ease fears about AI-infrastructure debt. | +8.8% |
| 16–17 Nov | AI and chip shares sell off ahead of NVIDIA's results. | −5.0%, −3.4% |
The pattern is that Oracle's shares now move on three kinds of news: its own financing, its customers' commitments, and the market's mood about AI spending. The first two are specific to Oracle and both went against it. The third is shared with every company on this list. An owner should notice that the only good company-specific day in the period was a day of promises, and that the promises were tested within a week.
The business, and what Oracle says a gigawatt earns
The first quarter, the second-quarter guide and the Investor Day economics
The first quarter of fiscal 2027 (to 31 August), which we covered in September, brought revenue of $19.3bn (+30%) and adjusted EPS of $1.92. For the second quarter Oracle guided revenue growth of 30–34%, cloud growth of 65–71% and adjusted EPS of $1.85–1.93; reported growth in EPS looks negative only because the same quarter last year included a $2.7bn gain on Oracle's stake in Ampere. Results are expected around 10 December.
What a gigawatt earns, by Oracle's account. At Investor Day, management described a typical one-gigawatt AI project, according to the transcript: about $37bn of capital spending, just under half of the contract's value, a gross margin of about 35% and an operating margin of about 34% over the life of the contract. Vendor financing could cut the upfront capital by about half. Oracle expects about 10 gigawatts by the end of fiscal 2028. If those economics hold and the customers pay, the investment earns a respectable return. The two "ifs" are the whole question, and Investor Day did not answer either of them.
★ The customers
One undisclosed giant, one reported retreat, one cancelled lease
OpenAI. Oracle still does not say how much of its $664bn order book belongs to OpenAI. Its annual report says no customer was 10% or more of revenue in fiscal 2024–2026, which is true and beside the point: the risk is in contracts that have not yet become revenue. Analysts still estimate about half. At Investor Day, Oracle's cloud chief said OpenAI had accepted capacity at the Abilene site "in less than 24 hours", a sign of demand. We found no news between our report and today of OpenAI changing what it has committed to pay.
Microsoft. On 6 November Bloomberg reported, via Reuters, that Microsoft is winding down a 2023 agreement to rent AI capacity from Oracle, because the capacity, built on an older generation of NVIDIA chips, no longer met its needs. Oracle replied that demand "remains strong", including for older chips; at Investor Day its finance chief said four-year-old GPUs were being renewed at premiums of about 20%. Both statements can be true. But the September report's worry, that chips become obsolete faster than the contracts that pay for them, now has a first data point.
Tencent. On 4 November the Financial Times reported that Oracle had cancelled a five-year lease of AI chips to Tencent, worth about $7bn, after US officials raised concerns. It is a reminder that some of Oracle's demand depends on Washington's permission as well as the customer's money.
| Customer signal | Since our report | What it tells an owner |
|---|---|---|
| OpenAI share of RPO | Undisclosed; ~half by analysts' estimates | The concentration is unchanged and still unpriced by Oracle's own disclosure. |
| Microsoft 2023 rental deal | Reported wind-down (6 Nov) | Older capacity can lose its tenant before it is paid for. |
| Tencent ~$7bn lease | Reported cancelled (4 Nov) | Policy can remove demand overnight. |
| New Mexico (Project Jupiter) | Force-majeure notice over power (Sep) | Power, not chips, may set the pace of the build. |
| New contracts in Q1 | More than $30bn booked | Demand is real and still growing; the question is who pays, and when. |
The moat, rechecked — and the competition
The toll road is intact; the airport has more neighbours
The database moat is as wide as it was. Licence and support revenue shrank 3% in the first quarter, as customers move from licences to cloud subscriptions, but nobody is leaving: the multi-cloud database, Oracle running inside Microsoft's, Google's and Amazon's clouds, grew fastest of all, and management said at Investor Day that its "AI database" was growing more than 200%. This is the part of Oracle we would happily own at the right price, and nothing since September has narrowed it.
The cloud moat is narrower than ever, because the business has become a contest of capital. Amazon, Microsoft and Google are growing their clouds quickly from far larger bases and financing them from operating cash flow; Oracle finances its build with debt and leases. Newer rivals such as CoreWeave, whose third-quarter revenue grew 168% (as reported), compete for the same AI laboratories with even more borrowed money. Microsoft, meanwhile, has reportedly tightened its oversight of the smaller "neocloud" providers it rents from. In a market where customers are getting choosier about old capacity, Oracle's advantage is speed and willingness to build; its disadvantage is the cost of its money.
| The claim | Evidence since September | Verdict |
|---|---|---|
| Database switching costs | Licence and support −3%, multi-cloud database and 'AI database' growing fastest; no sign of customers leaving. | ✓ Wide, intact |
| OCI price and performance | More than $30bn of new contracts in a quarter; GPUs renewed at ~20% premiums (management). | ◆ Real, but rented |
| Speed of build | ~10 GW targeted by end-FY28; power delays in New Mexico. | ◆ Fast, constrained |
| Cost of capital | New notes at 7.00–8.65%; Moody's outlook negative; hyperscalers fund from cash. | ✗ The weak point |
- That the large AI customers, OpenAI above all, pay for the capacity they have ordered over the next five years, so that most of the $664bn order book becomes revenue on schedule.
- That Oracle can fund the gap to positive free cash flow in the second half of fiscal 2028 through leases, prepayments and refinancing without a further large share issue and without losing its investment-grade rating.
- That fiscal-2030 earnings come in near Oracle's own target of $21 a share, so that today's price is about 5.5 times that number, and that the market then pays at least 12 times it.
Management, the founder and governance
Pledged shares, a cancelled sale plan, new grants, and one director who bought
The founder's shares. Our September report described Ellison as owning about two-fifths of Oracle. It did not mention two facts in the proxy filed two days before we published: 413 million of those shares are pledged against personal borrowing, and in June, during the worst week for the shares since 2001, a plan was adopted to sell up to 50 million of them. The plan was cancelled in September before any were sold. We do not know the terms of the loans, and Oracle's policy rules out margin accounts, so a forced sale is unlikely. But an owner should know that the largest shareholder's wealth is partly borrowed against the shares an owner is buying, and that the founder considered selling at much higher prices than today's.
Capital allocation. We score management and capital allocation at 3, down from 4. The reason is not ambition; building Oracle into an AI cloud may prove a brilliant decision. It is the sequence: a funding cap announced and exhausted within a week, shares sold at $116 after a decade of buying them back at higher prices, and dividends maintained with borrowed and newly raised money. Governance: the annual meeting is today, 18 November; shareholders are voting on a proposal about preserving their own right to file proposals. A securities class action in Delaware alleges that Oracle misled investors about its AI financing; the claims are unproven, the motion to dismiss will be fully briefed on 18 December, and a separate suit by bondholders over the 2025 notes is pending in New York.
The numbers
Cheaper on earnings, worse on cash, and six faults in our feed — two of which misled us in September
On earnings, Oracle is cheaper than in September: 14.1 times the fiscal-2027 estimate of $8.15 and 10.4 times fiscal 2028's $11.10, against 16.8 and 12.4 then. On cash, it is worse: trailing free cash flow of about −$28.7bn, a yield of about −8%, before counting the prepayments that flattered it. Interest cover is about 4.5 times. Cash and securities were $37.1bn at 31 August, plus about $19.8bn raised in November, most of which will be spent in the current quarter.
| What the screen shows | What is wrong, and the number to use |
|---|---|
| Property and equipment $161.8bn | Includes $34.0bn of leased (right-of-use) assets. Owned property and equipment was <b>$127.8bn</b>. Our September valuation used the wrong figure (Part I). |
| Total debt $155.9bn | Borrowings $125.3bn plus only part of the leases. All leases on the balance sheet are $43.8bn, so debt with leases was <b>$169.1bn</b>; about $179bn after November. |
| Market value $347.9bn | Uses a stale share count. With the 86.2m shares issued on 4 November, about <b>$358bn</b> at $115.06. |
| EV/EBITDA 12.4×; net debt/EBITDA 3.2× | EBITDA includes a ~$2.7bn one-off gain on Ampere. Cleaned, about <b>13.8×</b> and <b>3.5×</b>. |
| Larry Ellison: 8,194 shares, 41.8% (Feb 2024) | <b>1,158,375,174 shares, 38.2%</b> at the September record date, 413 million pledged (2026 proxy). |
| Dividend summary empty | Four payments of $0.50 in the last twelve months (<b>$2.00</b>). |
The dividend
$2.00 a year, flat, and paid with money raised for other purposes — tested
Oracle pays $0.50 a quarter, $2.00 a year, a yield of 1.7%. It has not been raised since April 2025. In the first quarter Oracle paid $1.57bn of dividends while free cash flow was −$5.4bn; in effect the dividend was paid from the $20bn raised by selling shares. Oracle ranks the dividend second in its priorities, after keeping its investment-grade ratings and before growth.
| Test | Evidence | Result |
|---|---|---|
| 1. Cover on free cash flow | None: trailing free cash flow about −$28.7bn against about $5.8bn of dividends. | Fail |
| 2. The trend of that cover | Negative since fiscal 2025 and planned to stay negative until the second half of fiscal 2028. | Fail, for now |
| 3. From operations or from paper? | From paper: new shares, notes and prepayments. | Paper |
| 4. Balance-sheet room | Borrowings ~$135bn pro forma; ~$19.7bn of credit lines available; ratings one or two notches above junk. | ◆ Thin |
| 5. What would force a cut | A downgrade towards junk, or a major customer failing to pay. Oracle has said the ratings come before the dividend. | ◆ Possible |
| 6. The growth rate and its direction | Flat since April 2025, after rises from $0.32 (2022) to $0.40 (2023) to $0.50 (2025). | Stalled |
Our reading. The dividend costs Oracle about $6bn a year against a capital programme of $90–95bn, so it is not what will decide the company's fate. But it is paid from borrowed and newly raised money, and Oracle has told us plainly that it would protect its rating first. Nobody should own Oracle for the income.
★ Valuation — the two Oracles, recomputed
What the database is worth, what is left for the data centres, and where we would look again
The method, unchanged; the input, corrected. In September we valued the database and applications business at about $285bn (about $39bn of annual revenue at a 50% operating margin, taxed, at about 18 times), subtracted about $88bn of net borrowings, and treated whatever the share price left over as the price of the AI cloud. We then compared that with what the AI cloud had cost to build. The comparison used $162bn of property; the right figure is $127.8bn.
Done correctly: at $137.10 in September, the price left about $218bn for an AI cloud that had cost about $128bn, 1.7 times cost, not the 1.3 we published. Today, with 8% more shares, the old Oracle is worth about $63 a share after net debt, and $115.06 leaves about $161bn for the AI cloud, about 1.26 times cost. At $100 it would be about 0.9 times. The 0.6 times we described for $100 now arrives at about $88–92, depending on how much more the AI cloud costs by December.
| Case | What has to happen | Value per share |
|---|---|---|
| The database alone | The AI cloud is worth no more than its debts and leases: a large customer fails to pay, chips age, money stays dear. | ~$63 |
| Our level | The database pays for about two-thirds of the price; the data centres come at about 0.65× their cost. | ~$90 |
| Oracle's plan, delivered | FY2030 EPS of $21, valued at 12× and discounted 3.5 years at 10%, plus dividends. | ~$186 |
Why we still do not name a central value. For most companies the cases in a table like this are points on a continuum, and the middle is a fair estimate. For Oracle the outcome is closer to binary: if the large customers pay and the money stays available, the plan's $186 is reachable and today's price will look cheap; if they do not, the debts and the $288bn of leases decide what the shares are worth, and that is not much more than the database alone. We cannot tell you which, and we are suspicious of anyone who can. That is what "too hard" means.
The level. We said $100 because there, we believed, the data centres would come at about 60 cents on the dollar. Correctly computed, that discount is now at about $90: about 11 times the fiscal-2027 estimate, with the database business worth about 70% of the share price. At that point an investor would be paid for the leverage and the concentration rather than asked to ignore them. We move the level to $90, and we keep the conditions: look at the second-quarter free cash flow and at any rating action first.
Risks, lawsuits & controversies
Verified 18 November 2026 against the Q1 FY2027 10-Q, the 2026 proxy, the November prospectuses and attributed reports
Securities class action. Barrows v. Oracle (District of Delaware) alleges that Oracle misled investors about its AI build-out and its financing; the claims are unproven. Oracle moved to dismiss on 16 September; the motion will be fully briefed on 18 December. A separate suit by bondholders, led by an Ohio carpenters' pension plan in New York state court, alleges misleading statements in the offering documents for the September 2025 notes; we found no later developments. In the Netherlands, a privacy class action was sent back to a lower court by the Supreme Court on 17 July 2026.
Customer and project news in this report (Microsoft's wind-down, the Tencent lease, the New Mexico force-majeure notice, project-finance amounts) comes from press reports by Bloomberg, the Financial Times and Reuters, as cited; Oracle has not confirmed the Microsoft and Tencent reports in its filings.
| Test | Today | What would change our mind |
|---|---|---|
| Q2 FY2027 results, ~10 Dec | RPO $664bn; FCF −$5.4bn (Q1) | RPO up with a named customer under 40% of it, and free cash flow better than −$10bn without prepayments, would make the case less binary. |
| Ratings | Baa2 neg. / BBB− / BBB | A downgrade to junk by any agency would raise the cost of every future dollar and put the dividend in question. |
| Leases not yet started | $288bn | Another large rise without matching new customer contracts would be the wrong direction. |
| Funding outside the cap | Cap used | Any new share issue before fiscal 2030 would break the Investor Day promise outright. |
| Delaware class action, 18 Dec | Motion to dismiss pending | A dismissal removes a cost; survival of the claims means years of discovery about the AI financing. |
| The price | $115.06 | Near $90 with the rating and the customers intact: look again seriously. Near $90 after a downgrade: wait. |
In September I told you Oracle was too hard, and that around a hundred dollars I would look again, because there the old database business would pay for most of the price and the new data centres would come at about sixty cents on each dollar they had cost. The shares have not reached a hundred. But I have looked again anyway, and the first thing I must tell you is that my sixty cents was wrong. I took the cost of the data centres from a figure that included buildings Oracle rents, not ones it owns. Done properly, at a hundred dollars you would be paying about ninety cents on the dollar. The discount I described is nearer ninety dollars a share.
The second thing is the funding. At the end of October Oracle gave its owners the plan they had asked for: free cash flow by 2028, thirty billion dollars of it in 2029, earnings of twenty-one dollars a share by 2030, and a promise to raise no more than twenty billion dollars of new capital before then. On the following Monday it raised the whole twenty billion: half in new shares at a hundred and sixteen dollars, half in bonds paying up to eight and two-thirds per cent. Perhaps that was prudence, taking the money while it was there. But when a company has to tell you how little it will borrow, and then borrows all of it within a week at those rates, the lenders are telling you something, and so is the company.
Give the business its due. The database underneath it all is one of the finest franchises in software; nobody is leaving it, and it is now sold inside every major cloud. The new business is booking contracts faster than almost anyone, more than thirty billion dollars of them in a single quarter. If the customers pay, Oracle's own plan makes today's price look like a bargain, and I will have been too cautious twice.
But the signals since September have mostly run the other way. Leases signed for data centres not yet built rose by twenty-eight billion dollars in one quarter, to two hundred and eighty-eight billion. Part of last quarter's cash flow was customers lending Oracle money. One large customer is reported to be walking away from older capacity, and another contract was cancelled because Washington objected. The founder, who owns more than a third of the company, has a large part of it pledged against personal loans. None of that is fatal. All of it makes the outcome more binary, and I do not know how to put a fair price on a coin toss with debt attached.
So: still too hard. If the shares come to about ninety dollars, with the ratings intact and the customers still paying, the database would be worth about seventy per cent of what you pay and the data centres would come at about two-thirds of their cost. That is a price at which you are paid for the risk rather than asked to ignore it, and I would look again seriously. Until then, I would rather explain a missed opportunity than a leveraged loss.
— The Buffett Lens. When the arithmetic in a bull case depends on one number, check that number twice; we did not, and we have now.
The Cap Lasted Five DaysA superb database franchise financing a leveraged AI build whose funding promise lasted a week. ★ At $115.06, 14.1× FY2027 consensus, the data centres are priced at about 1.26× what they cost, between the database alone (~$63) and Oracle's own plan delivered (~$186). Our September level of $100 rested on a wrong figure; corrected, the discount we wanted is at about $90. Q2 FY27 results ~10 December; watch free cash flow and the ratings.
Articles that take this apart
MarketWho actually gets paid in the AI build-outEnormous sums are moving. Following where they settle, rather than where they are announced, separates the businesses collecting cash from the ones committing it.
ValuationWhen a stock screener lies to youA screener is a very fast way to be confidently wrong. The ratio can be right and the model wrong, the model right and the data wrong — and every failure tends to point the same way.
Companies in FocusCheap, or cheap for a reason? Two companies that look identical from outsideA low price is the easiest thing in the world to find and the hardest thing to interpret. There is one question that separates a bargain from a trap, and it is not about the multiple.