Phase Space AI

Valuation

Oracle Corporation [ORCL]

Oracle Corporation [ORCL] — Valuation (RE-UNDERWRITTEN 2026-07-29, v3)

This document supersedes archive/ORCL_Valuation_2026-07-29_superseded.md. What changed and why is in §0. The prior conclusion is preserved in archive/ for comparison, per project convention.

v3 verification pass (2026-07-30)

v2 was written by an agent killed by an API 529 before it could emit ORCL_analysis.json. This pass verified v2's figures against EDGAR XBRL primary rather than rebuilding them — every balance-sheet line, the full income statement, the opex bridge, all nine RPO periods, capex/OCF/FCF, the filed share count and EPS, and the reproduced point-in-time multiple percentiles tie to source. Three arithmetic slips were corrected (§12).

One test v2 did not run changes its headline conclusion. The reverse DCF was run terminal-only. DATA_DEFECTS.md records that the terminal-only instrument overstates required CAGR in proportion to cash generation and instructs passing fcf_margin. Oracle's FCF margin is negative (−35.2%), so that bias inverts: terminal-only UNDERSTATES Oracle's required CAGR by 5.96pp — 12.22% becomes 18.18%, and v2's +5.13pp valuation margin becomes −0.83pp. §4 is rewritten accordingly and the Valuation verdict moves from PASS to PASS WITH ARGUMENT.

Spot $117.85 (2026-07-29 latest trade; prior close $119.95). Diluted shares 2,914.0m (FY2026 10-K); basic outstanding 2,880.5m (cover page, 2026-06-12). Market cap (diluted) $343,416m.

Net debt $135,538m. EV $478,954m. THIS EV IS LEASE-INCLUSIVE — it contains

finance leases and operating lease liabilities. The ex-operating-lease figure is $448,764m and the borrowings-only figure is $441,063m. All three are stated in §2 because the project has a documented defect in which lease obligations were silently excluded from EV and invalidated a seven-name ladder built on a "6.6x" anchor that was really 7.95x lease-inclusive.

---

0. What changed from the superseded version, and why

Classification of each change (per the standing "classify changes by alpha evidence" rule): the net-debt rebuild, the lease disclosure, the point-in-time multiple history, the opex-bridge terminal margin and the identity test are objectively-positive corrections — each replaces an unreproducible or mixed-basis number with one derived from a primary source. The choice of 30.0% rather than 28% or 32% within the bridge, and the 12-month multiple anchor, are unproven methodology judgements.

---

1. COMPANY STATE — declared first, because it determines the instrument

**STATE A on profitability and margin stability; the State-A test FAILS on its

business-model-transition clause. Terminal margin is therefore BUILT (State-C construction),

not anchored on trailing. Evidence grade: B.**

valuation.md STATE A requires: "profitable; operating margin low-variance across ≥5 years; no structural regime change (no transformative acquisition, no business-model transition, no accounting-basis break)."

Oracle is a mature, highly profitable company mid-transition into capital intensity. It is not State B (no exogenous commodity cycle drives it — the driver is its own contracted order book), not State C (it earns a 30% operating margin), and not State D. The four states do not cover this case, and that is a framework gap worth logging rather than forcing.

Consequence, and it is the operative point: the State-A licence to anchor the terminal margin on the trailing actual is void, because the trailing actual was earned by a business whose gross margin is falling ~3.5pp a year. The margin must be built forward through the opex bridge, which is what §3 does. And because the exit multiple is not independent of ROIC, §4 runs the identity — which is where this name's real answer lives.

---

2. NET DEBT AND EV — rebuilt from the balance sheet, line by line

Every figure below is from EDGAR XBRL primary (FY2026 10-K, filed 2026-06-22, period 2026-05-31), not from Alpha Vantage, because AV's debt fields are mislabelled on this name (§7).

The three EVs

Why lease-inclusive is the primary convention here, argued rather than asserted. Oracle's operating leases are not office space. They are data-centre capacity contracted to serve the RPO, and the 10-K discloses $53.3bn of undiscounted lease payments against them. Oracle's own risk factor treats them as an unavoidable obligation: "we could be locked into multi-year commitments for excess data center space and related capital expenditures, as well as associated financings, without receiving corresponding revenue." An obligation the company itself describes as a lock-in belongs in enterprise value. Excluding it understates EV by $30,190m, or 6.7% — and understating EV is the direction that manufactures a PASS.

The prior figure

The superseded document carried $102,601m. It cannot be reproduced:

• borrowings − cash = $97,647m (off by $4,954m)
• borrowings + finance leases − cash = $105,348m (off by $2,747m)
• AV shortLongTermDebtTotal − cash = $124,295m (off by $21,694m)

No documented convention yields $102,601m. The nearest is the ex-operating-lease figure, which the prior document understated by $2,747m while also omitting $30,190m of operating leases. Logged as defect 1. Note also that the prior figure is not the $133.9bn error the audit found in the screen's net_cash — that error was upstream, in the XBRL pipeline; the memo had already partly repaired it and introduced a smaller, unreproducible one of its own.

---

3. TERMINAL EBIT MARGIN — 30.0%, built through the opex bridge

3.1 What Oracle has actually demonstrated

Computed as operatingIncome / totalRevenue from AV normalized annual statements, cross-checked against EDGAR OperatingIncomeLoss. AV's ebit field is NOT operating income and was not used — on ORCL FY2026 it reads $24,194m (35.92%) against a true $20,606m (30.59%), an error of $3,416m and 5.07pp of margin (defect 4).

The single most important fact in this section: gross margin has fallen 13.9pp in four years and operating margin has RISEN 4.9pp over the same span. Operating leverage on the expense base has more than offset the depreciation loading into cost of revenue. Any terminal margin that ignores this is modelling half the mechanism.

3.2 The bridge, reconciled to EDGAR primary

FY2026, all lines from EDGAR (revenue $67,357m). AV's sellingGeneralAndAdministrative field contains G&A ONLY and silently omits sales & marketing entirely — $1,618m against a true $8,331m of S&M, i.e. 12.37pp of revenue missing (defect 3). The bridge below is therefore built from EDGAR, not AV.

Bridge residual FY2026: reported EBIT exceeds the bridge by $361m (0.54pp) — Oracle's "acquisition-related and other" presentation. The bridge reconciles to within 0.54pp, which is the reproducibility standard the brief asks for.

Hard constraint: m_EBIT,T (30.0%) ≤ m_gross,T (58.0%) — SATISFIED with 28.0pp of headroom.

3.3 Is 30.0% below the trailing actual, and if so why is that not an error?

Yes — by 0.59pp against the 30.59% FY2026 actual. The brief's test is "if your terminal margin is below the company's own trailing actual, that is almost certainly an error — say why it is not."

Here is why it is not:

1. It is 0.59pp, not 5.6pp. The superseded 25.0% was −5.6pp and the screen's 21.9% was −8.7pp; those were the errors. A 0.59pp step-down is inside the noise of the AV-vs-EDGAR disagreement on the very same line ($172m = 0.26pp).
2. The cause is named, quantified and company-stated: management has guided FY2027 gross margin down, and the mechanism is $90–95bn of FY2027 reported capex depreciating into cost of revenue. That is valuation.md Rule 2's "explicit causal bridge", not a haircut.
3. It is offset, not ignored: the bridge gives back 7.1pp of opex leverage against 7.2pp of gross margin. The near-perfect offset is not a coincidence — it is what the last four years actually printed, and extrapolating it is the least-assumption forecast available.

Sensitivity is run at 25.0% (the superseded figure, retained for comparability), 28.0% and 32.0% in §4.

---

4. THE IMPLIED-PATH TEST — and the identity that breaks the multiple

assets/reverse_dcf.py. Terminal value is 100% of EV in this instrument, so the reverse DCF is mandatory as the primary long-horizon output.

Held fixed and named: terminal EBIT margin 30.0%; exit multiple 23.05x EBIT (today's own, on AV EBIT, for comparability with the historical series); WACC 11.6%; horizon 5 years; TTM revenue $67,358m; EV $478,954m.

4.0 The interim-FCF correction — the test v2 did not run, and it moves the headline

DATA_DEFECTS.md records that reverse_dcf.py was originally terminal-only, that this overstates required CAGR in proportion to cash generation (+1.22pp at a 5% FCF margin, +6.04pp at 26%), and that the fix is to pass the company's own demonstrated fcf_margin. v2 did not pass it.

The defect is documented as one-directional. On Oracle it fires in the opposite direction, because Oracle's own demonstrated free-cash-flow margin is negative: −$23,686m on $67,357m = −35.16%. The interim term is therefore a cost to be funded, not a credit to be discounted, and omitting it understates what the price requires.

The correction is 5.96pp. No name in this corpus burns cash at Oracle's rate, so no name carries a larger version of this error. Generalisation worth propagating: the defect note should read "the terminal-only instrument misstates required CAGR in proportion to the magnitude of free cash flow, in the direction of its sign" — on cash generators it overstates, on cash burners it understates.

I hold −35.2% as primary because that is the instruction — the company's OWN demonstrated FCF margin — and because Oracle has committed to no year in which it turns positive (§6). It is nonetheless the harshest defensible reading: capex must eventually stop compounding. The row-by-row table above is the honest output, and the verdict is reported across it rather than at a point.

>>> THE MARKET REQUIRES: a revenue CAGR of 18.18% for five years (FCF-inclusive, primary).

>>> On v2's terminal-only basis it was 12.22%.

>>> Verdict: PASS WITH ARGUMENT — not the clean PASS v2 recorded.

The price now requires more than Oracle has demonstrated on the trailing full year, by 0.83pp. criteria.md admits that only as PASS WITH ARGUMENT, which requires "a specific, evidenced argument" rather than optimism. Here it is, and it is deliberately not "the RPO is large":

$76,560m of current RPO — the contractually booked next-twelve-month slice — is 85.1% of the $90bn FY2027 revenue guide. The growth the price requires is not a forecast for the first of the five years; 85% of it is already signed. Independently, the Street underwrites +46.1% for FY2028 ($130.5bn on 40 analysts), so the required 18.18% is below both the company's guide and the Street's, on a point-in-time consensus snapshot taken at this same $117.85 price.

Two things weaken it and are recorded, not resolved. The multi-year backward CAGRs now fail by ~7.5pp rather than ~1.7pp — they predate the regime change, but they are the only demonstrated five-year figures that exist, and a five-year required CAGR compared against a one-year actual is a duration mismatch in the flattering direction. And the exit multiple fails the identity test (§4.1), so the margin is conditional on a multiple no ROIC can warrant at Oracle's own cost of capital.

Robustness: NOT ROBUST. The verdict flips on the interim-FCF assumption alone, and flips again on the exit multiple. It is a verdict about an instrument as much as about a company, and it is labelled that way.

Implied multiple compression: 0.0x at the base (exit 23.05x = today's 23.05x). Against Oracle's own six-year point-in-time median of 27.61x, the base embeds a 16.5% de-rating that has already occurred.

4.1 THE EXIT MULTIPLE IS NOT A FREE PARAMETER — and Oracle's fails the identity

valuation.md Rule 4: EV_T / EBIT_T = (1−t)(1−g/ROIC) / (WACC−g). Run it.

>>> ORACLE'S ROIC (9.84%) IS BELOW ITS WACC (11.6%). At g = 3.5% and WACC = 11.6% the identity's

CEILING as ROIC → ∞ is (1−t)/(WACC−g) = 0.85 / 0.081 = 10.49x. Oracle trades at **23.05x.

NO value of ROIC warrants the traded multiple at Oracle's own cost of capital.**

This is not a rounding quarrel. Solving the identity for the WACC that would warrant 23.05x at g = 3.5% and infinite ROIC gives WACC = 7.19% — against a name whose own 252-day realised volatility is 65.8% and whose beta is 1.85. The market is pricing Oracle's AI compute contracts as low-risk annuities while its equity trades like a high-beta growth stock. Both cannot be right.

What this does and does not mean. It does not mean the reverse DCF verdict flips to FAIL — that test asks a different question (what growth does the price require) and it clears. It means the exit multiple used in that test is not independently warrantable, so the +5.13pp margin is conditional on a multiple the identity cannot support. Per valuation.md §"declare it unidentified rather than defaulted", the honest label is: exit multiple ANCHORED-BUT-UNWARRANTED. Reported, not resolved.

At the identity-consistent multiples the required CAGR is:

On the FCF-inclusive basis the break-point against the FY2026 actual is an exit multiple of ~24.0x — above where Oracle trades. There is no exit multiple in the lower half of Oracle's own six-year range at which the name clears on demonstrated growth once its cash burn is funded. It clears only against the forward bases (guide +34%, Street +39.2%), which is precisely why the verdict is PASS WITH ARGUMENT and the argument has to be the contracted current RPO rather than the multiple.

4.2 Terminal-margin sensitivity (exit 23.05x, WACC 11.6%)

Two readings, and both belong in the record. On v2's terminal-only instrument the screen's 21.9% was the only setting that produced a FAIL, so the verdict looked manufactured by the terminal margin — exactly the pattern the audit found across 21 names. On the corrected instrument the terminal margin is no longer the swing parameter: every setting from 21.9% to 30.0% fails on the trailing actual, and even 32.0% clears by under 1pp. The parameter that decides Oracle is not the terminal margin and not the exit multiple — it is whether the interim cash burn is funded inside the model. That is a finding about this name specifically, and it only became visible once the instrument was fixed.

4.3 WACC sensitivity (terminal margin 30.0%, exit 23.05x)

4.4 On the ex-operating-lease EV, for readers who prefer that convention

At EV $448,764m, terminal margin 30.0%, exit 23.05x: required CAGR 16.80% FCF-inclusive (10.77% terminal-only), margin +0.55pp on the FY2026 actual and −6.31pp on the 3-year CAGR. The convention choice moves the answer by ~1.4pp of required CAGR — less than the 5.96pp the interim-FCF instrument moves it. Stated so the reader can see the size of both rather than inherit either. Note the direction: on the ex-operating-lease convention the name scrapes a positive margin on the trailing actual, so the lease convention alone can flip the sign of the valuation margin. That is why §2 argues the convention rather than asserting it.

---

5. RPO — the entire thesis, with the exact disclosed figures

All nine figures below are EDGAR XBRL RevenueRemainingPerformanceObligation, i.e. the tagged number in the filing, not a transcript paraphrase.

Current RPO (the next-twelve-month portion)

Oracle does not tag a current-RPO figure; it discloses a duration schedule in Note 1 prose. FY2026 10-K, verbatim: "of which we expect to recognize approximately 12% as revenues over the next twelve months, 34% over the subsequent month 13 to month 36, 34% over the subsequent month 37 to month 60 and the remainder thereafter."

$76,560m of current RPO is 85.1% of the $90bn FY2027 revenue guide and 113.7% of all FY2026 revenue. Weighted-average duration on midpoints ≈ 3.4 years; 88% of the book sits beyond twelve months.

How much is attributable to named AI compute contracts

Oracle names no counterparty and discloses no single-customer share of RPO, anywhere. What it discloses is a statement about a different metric: "No single customer accounted for 10% or more of our total revenues in fiscal 2026, 2025 or 2024." On a $638bn book against $67bn of revenue those are not the same question.

It is nevertheless bounded by Oracle's own words. Q1 FY2026 release: "We signed four multi-billion-dollar contracts with three different customers in Q1. This resulted in RPO contract backlog increasing 359% to $455 billion." RPO rose $317,500m in that one quarter — 49.8% of today's entire book — from three counterparties. That averages ~$105.8bn per counterparty, 1.84x Oracle's entire FY2025 revenue, each.

Q4 FY2026 (Hilary Maxson, prepared remarks): "Most of the RPO increase in both Q3 and Q4 were large scale AI contracts."

Attributable to named AI compute contracts: ≥49.8% by Oracle's own arithmetic; the true figure is higher and is UNDISCLOSED. Recorded as INDETERMINATE, not estimated.

The longer series — this is a break, not a short-history artifact

EDGAR tags RevenueRemainingPerformanceObligation back to 2018-08-31, and the full 32-quarter series matters because it establishes that the eight-quarter +551% is a genuine structural break rather than a small base flattering a growth rate:

RPO grew at a 6.4% CAGR across FY2019–FY2022 and 4.6x across FY2022–FY2025. Then it grew 4.6x in a single year. Seven years of stable, slow-compounding disclosure is the strongest available evidence that the FY2026 figure is not a definitional change or a disclosure-policy artifact — the tag, the filer and the presentation are unchanged; only the number moved. Per MEMO_BRIEF's rule, this is unprecedented-and-verified, and it is adopted.

5.1 WHERE ORACLE SITS BETWEEN MU AND KLA — the memo's central finding

The queue asks ORCL to be placed between two poles measured the same night. The honest answer is that it is directionally the anti-KLA, structurally weaker than MU on every hardness test that can be checked, and a third case neither pole spans.

Direction: ORCL is the opposite of KLA, decisively. KLA's backlog was being consumed — a pull-forward converting into revenue and shrinking. Oracle's is being built faster than it converts, and the first year of its published conversion path landed on the number ($18.1bn against an $18bn OCI target). Nothing about Oracle resembles the KLA pattern.

Hardness: ORCL is weaker than MU on 5 of 5 checkable tests. MU's backlog is hard in the two ways that survive a counterparty changing its mind — a contractual take-or-pay obligation, and cash already posted. Oracle discloses neither. Its $638bn is a performance obligation, which is a real accounting construct, but it is unbilled to 97.6% and uncollateralised to at least 97.6%.

And here is the axis neither pole measures, which is where Oracle actually differs:

MU and KLA both fund their backlog out of positive free cash flow. Oracle funds its at −$23.7bn of FCF and a guided $40bn of fresh debt and equity. The discriminating question is not hard-vs-soft backlog — it is WHO CARRIES THE CAPITAL RISK OF CONVERSION.

On MU, the customers carry it: they posted ~$18bn of cash deposits and signed take-or-pay. On KLA, the question is moot: no capital cycle stands between backlog and revenue. On Oracle, shareholders and bondholders carry it: $167bn of obligations, capex above 100% of revenue, and a $20bn ATM issuing equity at the 23rd percentile of its own multiple.

So the placement is not a point on a line between MU and KLA. Oracle is hard on direction, soft on collateral, and uniquely exposed on financing — and the financing axis, which the MU/KLA comparison does not contain, is the one that determines the outcome. That is why the corrected reverse DCF (§4.0) moves the verdict: it is the only test in this document that prices the financing axis at all.

---

6. THE FINANCING SIDE — is the RPO converting to cash? No, and the gap is widening

EBITDA is computed as EBIT + |D&A from the cash-flow statement| = $20,606m + $9,293m = $29,899m. AV's ebitda field says $32,138m — overstated by $2,239m (7.5%) because it is built on AV's wrong ebit and on income-statement D&A of $7,944m, which understates the cash-flow figure of $9,293m by $1,349m (14.5%). av_vs_edgar.py flagged this independently at 2025-08-31: IS D&A $420m vs CF D&A $1,771m — 0.24x (defect 6).

The conversion timeline management has committed to — in its own words

Q4 FY2026 call, Hilary Maxson, prepared remarks, verbatim:

"Our net cash outlay for capital expenditures for the full year was $48 billion, taking into account prepayments and timing impact of around $8 billion… We will continue those investments in our fiscal year 2027, with an expected net cash outlay for capital expenditures of around $70 billion. This includes customer prepayments and timing impacts expected at around $20 billion to $25 billion, so our reported CapEx will be higher by this amount."

"To support our capital investments program, we expect to raise around $40 billion in debt and equity in our fiscal year 2027, and that includes our already announced $20 billion at-the-market equity issuance. We do not anticipate raising additional debt funding in calendar year 2026."

"Our fiscal year 2027 guidance: you can start to see the strong translation of our RPO into revenues, with expected growth in our total revenues of +34% in constant currency… Our fiscal year 2027 gross margin will step down."

>>> CASH-CONVERSION STATUS: NOT CONVERTING, AND MANAGEMENT HAS COMMITTED TO NO DATE.

Read the arithmetic rather than the adjective. Net cash capex is guided from $48bn to ~$70bn, +45.8%. Reported capex is guided from $55.7bn to ~$90–95bn, +62–71%. Against revenue growing 34% to $90bn. Capex will exceed revenue in FY2027. Oracle will raise $40bn of fresh debt and equity to fund it. There is no stated year in which free cash flow turns positive, no stated peak-capex year, and no FCF target. What management commits to is revenue conversion (+34%) — a different variable.

Two things are true at once and the memo says both. The RPO is real, filed, tagged, and its first year converted on the number: Oracle's published five-year OCI path ($18bn FY26 → $32bn → $73bn → $114bn → $144bn FY30) delivered $18.1bn in FY2026. And the cash cost of that conversion is accelerating faster than the revenue it produces. A company booking $638bn of RPO while free cash flow runs −$23.7bn and rising is a financing story wearing a backlog story's clothes. The $75bn of customer prepaid-and-supplied hardware genuinely reduces Oracle's capital requirement — but total contract liabilities are only $15,395m, so the cash-prepaid share cannot exceed that and is probably far less; the majority is customer-supplied hardware, which cuts capex, not credit exposure. 97.6% of the $638bn RPO is unbilled.

---

7. Own-multiple history — point-in-time, six years

Daily series, 1,506 sessions from 2020-07; TTM EBIT/revenue/net income as known at each date (AV normalized quarterlies lagged 45 days to approximate the filing date); net debt and share count point-in-time from 81 quarterly balance sheets. This retires the superseded version's disclosed distortion.

Oracle is not at the 1st–17th percentile the superseded version reported. On a correctly constructed series it sits at the 23rd–35th percentile of its own six years — cheap-ish, not distressed. The prior figure was an artifact of holding today's $135bn of leverage constant across a period when Oracle carried a fraction of it.

---

8. The 12-month target

Step 1 — near-term estimates as the base. FY2027 guidance: revenue +34% cc (~$90.0bn), non-GAAP EPS $8.05. The trailing multiple history is built on GAAP EPS, so GAAP is what must be used. FY2026 GAAP EPS $5.86 (diluted, $17,087m / 2,914.0m) less ~$0.93 of Ampere/Bloom investment gain = clean FY2026 GAAP EPS $4.93. Applying Oracle's own "18% adjusted" growth gives $5.82; a 25% case gives $6.16. Base FY2027 clean GAAP EPS: $5.95.

Note a new equity-method item to watch: from Q3 FY2026 Oracle holds a 15% stake in TikTok US, accounted for under the equity method (Doug Caring, prepared remarks). That will inject a non-operating, non-cash line into GAAP EPS of exactly the kind that requires cleaning — see the MSFT memo for how badly this can distort a headline.

Step 2 — named product-cycle events inside 12 months (each dated in ORCL_Catalyst_Calendar.md): Q1–Q3 FY2027 prints, each disclosing RPO and OCI revenue against the $32bn FY2027 OCI path; execution of the $20bn ATM; the first quarter in which the FY2027 depreciation step-up lands in reported gross margin.

Step 3 — the multiple, on Oracle's OWN range, percentile stated. Trailing P/E 20.10x = 23.2nd percentile of the six-year point-in-time distribution; 0.8th of the three-year.

Regime caveat, disclosed rather than assumed away. The six-year window spans pre-AI Oracle (a low-growth licence business at 16–22x) and AI-capex Oracle (30–55x). Today's business is neither. The three-year window is entirely the re-rating and its unwind and is not usable as a central tendency. I therefore anchor on the six-year p25 (20.72x) with today's 20.10x as the floor, and state plainly that the multiple anchor is the weakest element of this valuation.

12-month target: $123, +4.4% to spot.

Lower than the superseded $134 (+11.9%) for two reasons, both corrections: the point-in-time p25 is 20.72x not 22.6x, and today's P/E sits at the 23rd percentile not the 12th, so there is less reversion room than v1 credited. No external professional target is on file for ORCL in this book.

---

9. Downside case with a named cause (MEASURED — logged and scored; does not reject the name)

Named cause: non-performance by one of the three counterparties that signed ~half the RPO book in Q1 FY2026 — a mechanism written into Oracle's own risk factors and quantified in §5.

Quantified: clean GAAP EPS held flat at $4.93 (the mechanism stalls rather than reverses) on Oracle's own six-year P/E p10 of 18.45x → $91, −22.8%. On the six-year minimum of 16.84x → $83, −29.6%.

Is it a going-concern case? No, and that is argued. Interest coverage ~4.5x on FY26 interest of $4,599m against a 3.0x covenant; $31.9bn of liquid assets; the $75bn prepaid/customer-supplied hardware means a meaningful share of the GPU capital is not Oracle's to fund; and management has publicly committed to "preserving our investment-grade credit rating." The risk is a large permanent impairment of equity value, not solvency. Probability: 25% (judgement, logged for Brier scoring).

Second, mechanical downside: the $20bn ATM plus ~$20bn more of FY2027 issuance is ~6% dilution into a 23rd-percentile multiple. Issuing equity below intrinsic value is value-destructive by construction, and Oracle has announced it will.

Third, new to this version: the ROIC-below-WACC finding in §4.1 is itself a downside mechanism that requires no counterparty failure. If Oracle simply earns its cost of capital and no more on $167bn of invested capital, the warranted multiple is ~10.5x, not 23x. That is a −55% multiple outcome from arithmetic alone.

---

10. What would change the analysis

---

11. Defects found in this re-underwrite

Verified clean: revenue ties AV↔EDGAR across all tested quarters (av_vs_edgar.py: disagree 0/4); DebtLongtermAndShorttermCombinedAmount ties to the component sum exactly; no stock split since 2000-10-13, so no per-share basis risk; EntityPublicFloat $346.8bn at 2026-05-31 does not exceed same-date market cap.

---

12. v3 verification pass — what was checked, corrected, and left unverifiable

Every figure below was checked against EDGAR XBRL (companyfacts / companyconcept, CIK 0001341439) — a source Alpha Vantage did not produce, which per MEMO_BRIEF is the only check with power, since a vendor that is wrong consistently defeats every internal consistency test.

Tied to the dollar

All ten balance-sheet lines in the net-debt rebuild; revenue and operating income FY2023–FY2026; every opex-bridge line (R&D $10,272m, S&M $8,331m, G&A $1,618m, intangible amortisation $1,671m, restructuring $1,779m); OCF $31,977m and capex $55,663m, hence FCF −$23,686m; all nine RPO periods plus 23 further periods back to 2018-08-31; diluted shares 2,914m; spot $117.85 (Alpaca latest trade and the point-in-time consensus snapshot agree). The multiple-history percentiles were reproduced by re-running the point-in-time script, and the required CAGR of 12.224% was reproduced from reverse_dcf.py at v2's stated inputs before the FCF correction was applied.

Corrections to v2

The two traps in the brief, checked rather than assumed

• The WDC trap (longTermDebt = 0 was genuinely correct for WDC). On ORCL, AV's longTermDebt $122,342m and shortTermDebt $7,199m both tie exactly to EDGAR (LongTermNotesAndLoans, NotesPayableCurrent), and their sum ties to the independent DebtLongtermAndShorttermCombinedAmount tag at $129,541m. Reconciled from three tags, assumed from none. The zero-in-newest-quarter defect does not fire on this name.
• The MU trap ($8.5bn of refundable customer deposits counted as cash with no offsetting liability). On ORCL, contract liabilities of $15,395m are properly presented as a liability and are not netted into cash; the $31,289m is plain cash. The only adjustment required ran the other way — adding the $605m of current AFS securities that AV's cashAndShortTermInvestments omits. No MU-style overstatement is present.

Left unverifiable — disclosed rather than glossed

EDGAR has not tagged us-gaap:GrossProfit for ORCL since FY2018. The 65.20% gross margin, and therefore the 58.0% terminal gross-margin anchor that the whole opex bridge hangs from, is single-source (Alpha Vantage). Indirect corroboration: revenue − EDGAR total opex − EDGAR operating income implies $23,080m of cost of revenue against AV's $23,441m, a $361m gap identical to the bridge residual, so the two are internally consistent to 0.54pp. That is corroboration, not verification. Recorded as SINGLE-SOURCE. It is the weakest input in §3, and it is a defect of the data layer rather than of this memo (logged as defect 20 below).

Additional defects found in the v3 pass