The drivers, decomposed. Four drivers carry the thesis. AWS growth — settled by the record (contracted backlog, +36.8% in Q2 2026). Advertising growth — settled by the record (+22%/yr over 3.5 years; its legal risk is a debate above, not a driver dispute). Retail margin grind — settled in direction (North America segment margin up every year in the window, 7.4% TTM). Capex intensity — the one contested driver, and therefore this page's turbo trigger: everything else is corroborated; the argument is about what "normal" investment costs.

1. The one dated forecast management made about this buildout was wrong in the expensive direction. In Feb 2025 Jassy predicted constraints would "really start to relax in the second half of '25"; by Jul 2025 the horizon was "several quarters," and the capex guide has since gone $200B → ~$220B with 2027 flagged higher. Relief later means more capital, sooner, at memory-inflated prices.
2. The earnings base the price needs has never existed. Reported TTM owner earnings are −$26.9B. The +15.1% bar runs on a normalized base built from the single friendliest capex year in the series; at the trailing three years' intensity the same model demands +24.2%/yr — a decade of AWS-speed compounding at consolidated scale, with margin expansion on top.
3. The required margin is undemonstrated and the accounting runs against it. At delivered revenue growth the price needs a ~14–15% mature owner-earnings margin; Amazon's best reported figure is 2.5%. Management concedes capex runs 6–24 months ahead of billing with free cash flow "challenged" until revenue outpaces capex (Jassy, Apr 29, 2026) — while a five-year-life depreciation wave runs through the profit and loss statement (lives already shortened once, 6 → 5 years).
4. The corroboration is circular. ≥$238B — ≥48% — of the $496B backlog is commitments from two private AI labs, Anthropic and OpenAI. The same two names are also balance-sheet marks ($219.1B combined) and the stated justification for the capex program. Backlog, equity marks and capex justification rest on the same two counterparties' ability to pay — presented as three supports, they are one.
5. The legal stack now aims at the profit engine, not the perimeter. The Aug 31, 2026 ad-auction suit attacks Sponsored Products economics directly; the FTC trial (Mar 29, 2027, reported) carries structural-remedy risk; Prime already cost $2.5B.

1. Conceded in part. The walk-back is real and is why capex guidance is graded as the hole in a B+ (Management). But the failed forecast was about supply relief timing; on demand, the same record shows revenue and operating income guides beaten or met in all nine scored quarters. The error cost money through capex, not through demand that never arrived.
2. Answered by construction. The normalization is not hidden: the full intensity range is printed beside the bar (Reverse DCF), the reported base at −$26.9B, and even the friendly end requires a decade of double-digit owner-earnings compounding. The reader prices the range, not a point.
3. Conceded as the central risk; quantified, not waved. The breakeven grid shows exactly what must be true: at the demonstrated +11.8% revenue rate, only mature margins ≥14% clear the price — no normalized-start row clears at 12%. Management's mechanism (Trainium economics, ads mix, AWS mix) is named in Moat; the depreciation counterweight is in the Downside Checklist.
4. Conceded as concentration; bounded by terms. The lab commitments are contracted at ≥5-year terms (management, Q2 2026 call) and backlog life lengthened to 6.4 years; the exposure is summed and dollar-sized on this page (≥$238B ≈ 9.4% of operating EV). What cannot be bounded from outside — the labs' own funding durability — is stated as a gap.
5. Answered in part. Prime settled for $2.5B without structural change; the ad-auction >$20B figure is a complaint allegation; the trial date is press-reported, not docket-verified (Regulatory). The risk is real, dated, and checkpointed — not yet a finding against the company.

Amazon is three businesses sharing one balance sheet. (1) Commerce: North America ($453.7B TTM sales) and International ($173.6B) run at thin but rising margins (7.4% North America, 3.1% International). (2) AWS ($148.4B TTM, 19% of sales) produces 58% of consolidated operating income at a 36.8% TTM margin, and is capacity-constrained — a condition management has described across a nine-call arc (The record). (3) Advertising ($76.1B TTM, +22%/yr) — mostly Sponsored Products, increasingly Prime Video and off-Amazon surfaces — is the second profit engine and the main reason consolidated margin keeps expanding. Subscriptions (Prime) add $52.9B TTM.

Overlaid on all of it: a ~$220B 2026 capex program (AI data centers, custom silicon), equity stakes in Anthropic and OpenAI carried at $219.1B combined (marks, not cash), and Amazon Leo, the satellite build.

Q2 operating income includes ~$1.2B of flagged one-off benefits (~$600M tariff refunds, ~$600M energy-derivative mark). Net income is dominated by marks: $53.4B of the quarter's income is pre-tax other income, primarily the Anthropic mark (Corrections). Reported net income is not a usable earnings measure right now. AWS's +36.8% sales growth was its fastest in 18 quarters.

*TTM weighted diluted shares are not printed in any single filing; ~10,872M is derived from the half-year counts. The release's period-end "shares + awards outstanding" figure (11,027M) is a different measure (label corrected in v2; no model input uses either figure).

Actual compounding (the base-rate line): revenue compounded +11.8%/yr FY21→TTM (FY22→TTM is +12.5%/yr — it flatters off the FY22 trough; cross-check only). Owner earnings have no meaningful 5-year compound rate — negative at both ends of the window, which is precisely the valuation problem the Reverse DCF prices.

Guide = range given on the prior quarter's call; delivered = as reported on that quarter's call. Deciding rows printed here; the full 12-quarter table is in the research dossier. Quarters the notes cannot score are marked, not smoothed.

Tally (all nine quarters with full guide ranges in the notes, 2023Q3→2026Q2): revenue above top 7, in range 2, below 0; reported operating income above top 8, in range 1, below 0. No quarter in the series printed below range on either line. 2024Q4 revenue and 2025Q4 on both lines are untestable from the notes (the guide ranges were not captured in those calls' notes) and are flagged, not claimed.

Capex guidance is the credibility hole, and the calls show the mechanism: FY2023 actual $48.4B → ~$75B (Q3 2024 call) → 2025 framed as a ~$105B run-rate (Feb 2025), guided to ~$125B cash capex on the Q3 2025 call ("We expect our full year cash CapEx to be approximately $125 billion in 2025, and we expect that amount will increase in 2026" — Brian Olsavsky, CFO, Oct 30, 2025) → 2026 set at $200B (Feb 5, 2026) → raised to ~$220B (Jul 30, 2026), the raise attributed to memory cost, not volume — a distinction management volunteered. (Definition flag: the calls' ~$125B is cash capex; a ~$131B figure in the Q4'25 note appears to include equipment finance leases — unreconciled, so the page uses the cash series.) The pattern across twelve calls: demand-led, disclosed promptly, always up.

The dive's central operating fact is not in any release; it is a language arc across nine calls — a soft acknowledgement (Q2 2024: "we would like to have more capacity than we already have today") followed by explicit constraint language on eight consecutive calls (Q3 2024 named chips binding; Q3 2025 monetizing, 3.8 gigawatts of power added in the trailing year; Q1 2026 shifted the evidence to components — "The cost of components, particularly memory, has skyrocketed"). Read for the thesis: demand corroborated by the same voice for two years; the one dated relief forecast was wrong in the direction that costs money. The five load-bearing moments:

Jun 30, 2026 (Q2 2026 10-Q).

Debt is face $133.0B across 11 note tranches (2014–June 2026 issues, maturities 2027–2076). Post-period: +$25.0B notes issued (2029–2066); the remaining $21.3B OpenAI funding paid (total $50.0B). No financial covenants. The filing's own aggregate for the investment portfolio is $229.7B; the ≈$231.9B component sum used in the bridge differs by a public-warrant overlap between buckets (the split is undisclosed) — ~0.1% of EV, moving no bar at 0.1-point rounding.

Lease treatment: finance leases are in the bridge (principal is a financing flow, never charged in the base); operating leases are excluded (their payments already reduce operating cash flow — adding them would double-count rent).

House stack: 10% hurdle, 10-year window, 2% terminal growth, terminal return on invested capital fades to cost of capital, solve on EV.

Base construction. Reported TTM owner earnings = $161.4B operating cash flow − $169.0B net capex − $19.3B stock-based compensation = −$26.9B — printed, not buried. A negative base cannot be inverted — and shouldn't be: FY2026 capex (~$220B guided) is a deliberate supercycle, not steady state. Normalization: FY2023 was the last pre-supercycle year; its net capex intensity (8.3741% of sales, recomputed from the filings) already funded normal growth capex. Holding it on TTM sales gives normalized capex of $65.0B and normalized owner earnings of $77.1B (9.9% of sales). The judgment is disclosed — and, new in this edition, priced across the whole series:

FY2023 is the lightest capex year in the series, so the headline is the easiest bar in the set. Companion lenses at the headline normalization: +16.3%/yr on headline EV; normalized free cash flow (pre-stock-compensation, base $96,447M) +12.0%/yr; reported owner earnings (base −$26,918M) n/m — base is negative.

Terminal diagnostics (headline base): the 10.5× implied terminal EV/EBIT (earnings before interest and taxes) sits inside the mature 8–14× band — a growth bar, not a multiple fantasy.

Cells marked ✓ clear the price at the demonstrated revenue rate (+11.8%/yr, FY21→TTM).

No row clears at a mature margin of 10% or 12% — at a ≤12% mature margin, Amazon needs more revenue growth than it has demonstrated, from every starting point. Every normalized-start row clears from 14% up. The reported-start row's 14% cell prints 11.8% but is 11.803% unrounded — just above the threshold, so it is not marked clearing. The mature margin is the dominant lever (each +2 points buys ~1.6–2.4 points of required revenue growth); the starting margin, which moves the governing bar so violently above, mostly washes out over the five-year ramp — the two tables answer different questions, and both are printed so neither can be cherry-picked.

Buyback credit: none earned — credited bar = flat bar. Zero repurchases in 2023, 2024, 2025 and H1 2026 (verified verbatim in the filings); $6.1B remains of the March 2022 $10B authorization; the count is rising ~+1.3%/yr (10,613M at Q1 2025 → 10,783M at Q2 2026), so per-share compounding runs ~1.3 points/yr behind total growth. The posture has not changed in twelve calls: asked about buybacks on the Q1 2024 call, CFO Brian Olsavsky answered "we don't have anything to share with you on that today... Our first priority is to invest in — to support the growth opportunities and long-term investments within our businesses."

Sensitivity (headline base, operating EV): hurdle 8/9/10/11/12% at 2% terminal growth → +10.6 / +12.9 / +15.1 / +17.1 / +19.0%/yr. Segment lens (decomposition, not a multiple-based sum-of-the-parts): the inversion's margin burden is shared — AWS must keep compounding near +20%/yr (it did +19.3%/yr FY22→TTM) while retail margins grind from ~6% blended toward double digits; advertising (+22%/yr) is the mechanism for the second half. (Segment operating income in The record.)

• Backlog. Remaining performance obligations, primarily AWS: $244B at Dec 31, 2025 → $496B at Jun 30, 2026, weighted-average life 4.1 → 6.4 years (10-K / 10-Q). Q1's $364B explicitly excluded the >$100B Anthropic expansion. Backlog is the evidence management offers every quarter for the discipline claim — "We don't spend the capital without very clear signals that we can monetize it this way" (Jassy, Q1 2024 call, when capex was $14B a quarter; the formulation recurs at $200B and $220B).
• Counterparty commitments, summed. AWS–OpenAI $38B + $100B expansion over 8 years; AWS–Anthropic expansion >$100B over 10 years — ≥$238B combined, ≥48% of the $496B backlog, ≈9.4% of operating EV, resting on two private AI labs. Equity paid for the anchor customers: OpenAI $50.0B total; Anthropic ~$18.0B cash in, carried at $190.4B at Jun 30, 2026 (convertible notes at fair value $97.9B + preferred $92.5B); the two positions together are carried at $219.1B — marks, not cash.
• Management's return case is a timing case, in their words: "most of our AI capacity these days is being contracted for at least five-year terms" (Jassy, Q2 2026 call, Jul 30, 2026); servers/networking break even in under three years on average; data centers monetize 30+ years across 5–6 server generations. Chips: Trainium + Graviton >$25B run rate, triple-digit growth; commitments >$225B; Trainium2 "largely sold out" (Jassy, Q1 2026 call).
• The bear in management's own words (Q1 2026 call, Apr 29, 2026): "The faster AWS grows, the more short-term CapEx we will spend... typically six to 24 months before we start billing customers... in times of very high growth like now, where the CapEx growth meaningfully outpaces the revenue growth, the early years' free cash flow is challenged until these initial tranches of capacity are being monetized and revenue growth outpaces CapEx growth." The return case depends on 5–6-year server lives holding — and Amazon already shortened server life once (6→5 years, Jan 1, 2025; +$1.4B FY2025 depreciation).
• Under direct questioning on returns (Mahaney, Q4 2025 call, Feb 5, 2026): "We have... a fair bit of experience over the years in AWS of forecasting demand signals and doing it in such a way that we don't have a lot of wasted capacity... this isn't some sort of quixotic top line grab... I'm very confident we're gonna have strong return on invested capital here" (Jassy). The confidence is on the record; so is the absence of a disclosed realized return-on-invested-capital figure for the AI tranches.

1. Runway + successor determination — answered. Backlog tenor lengthened (4.1 → 6.4 years) while doubling; Trainium commitments >$225B and reserved 2027 capacity say the successor products are being determined inside the same franchise.
2. Innovation productivity — thin-evidence. Custom silicon is the claimed edge (Trainium2 "about 30% better price performance," management-asserted; >$25B Trainium + Graviton run rate). No independent benchmark exists in the record — the claim is management's own, labeled as such.
3. Go-to-market quality — answered. Prime subscriptions ($52.9B TTM) and third-party seller services give the commerce engines a distribution flywheel no rival replicates; AWS sells into an installed enterprise base it already bills.
4. Margin level + what management is doing to defend it — answered, with the defense named. Every margin display on this page carries the same defense: ads mix (+22%/yr on the highest-margin revenue), AWS mix, and Trainium economics (management claims several hundred basis points of margin advantage at scale). The counterweight: a five-year-life depreciation wave on ~$220B/yr of new plant. Consolidated operating margin has expanded four straight years while the defense was tested (2.4% → 12.1%) — the strongest single moat datum on the page.
5. People — answered. Jassy (CEO since Jul 2021), Olsavsky (CFO since Jun 2015), Garman (AWS CEO since Jun 2024): tenure, coherence, and a guidance record (The record) that functions as the people-evidence.
6. The industry-peculiar edge — answered. Logistics density, reserved power positions (3.8 gigawatts added in the trailing year), and silicon designed for its own racks: a stack of physical commitments competitors must duplicate at 2026 prices.
7. Horizon + dilution economics — answered, and it is the contested driver. Data centers monetize 30+ years; servers depreciate over 5; the count drifts +1.3%/yr with zero buybacks (Reverse DCF).
8. Candor → integrity — the closing gate; see Management. One walked-back forecast, disclosed promptly, against nine scored quarters of kept guides.

Re-verification (Oct 16, 2026, refreshed the same day against live EDGAR). (a) Form 4 population — all 154 Form 4/4A filings, Oct 2024–Oct 2026, 19 owners: zero open-market (code-P) buys, zero pledges/forwards; live refresh confirms no filing of any form after Oct 6, 2026. (b) Latest proxy (2026 DEF 14A, Apr 9, 2026): Jassy 2025 reported pay $2,069,861; no CEO grant since the 2021 promotion award; no named-executive-officer grants in 2025; single share class; Bezos 950,434,581 shares (8.8%). (c) Buybacks verbatim: zero repurchases 2023–2025 and H1 2026; $6.1B authorization remaining; count +1.60% over five quarters (≈+1.28%/yr). (d) No call since the grade date (last: Q2 2026, Jul 30, 2026).

For the grade: the 12-quarter guidance machine (The record); bad news disclosed with numbers attached across the series (server-life shortening and its −$400M net FY2025 operating-income impact, Q4'24 call; the $4.3B charges separated from the $21.7B ex-charges print, Q3'25; the ~$1B Leo drag in the Q1'26 guide, Q4'25; the ~$1.2B one-off benefits inside Q2'26 operating income, Q2'26); the Anthropic trade ($1.25B in, July 2023 → $45.8B year-end-2025 mark → the OpenAI deal at a $350B anchor); Jassy net-accumulates across the Form 4 window (vest acquisitions 402,640 shares vs sales 223,408).

Against: capex is the largest guided number and it keeps missing upward ($105B run-rate framing → ~$125B cash delivered 2025; $200B → ~$220B for 2026); the one dated operating forecast (capacity relief H2'25) was walked back a quarter later; buybacks are exactly zero for 3.5 years while the count drifts up; MGM ($8.45B) and One Medical (~$3.9B) returns never broken out; Leo's losses undisclosed.

Context: Bezos is a programmatic seller under disclosed 10b5-1 plans (2025 sales 25.0M shares, ~$5.65B, under a plan adopted Mar 4, 2025; a third plan adopted Nov 14, 2025 was already executing Aug 2026) plus recurring gifts. Context only; not insider signal. No dividend ever.

Integrity capstone (Fisher's closing gate): PASS with the named exception — prompt, numbered disclosure of bad news across twelve calls; the exception is the capex guide itself, where the numbers arrived promptly but always larger. Candor about the present, optimism about the trajectory: the B in candor, and why the grade is B+ and not higher.

• FTC monopolization (W.D. Wash. 2:23-cv-01495). Liability-only bench trial March 29, 2027 — reported (MLex / Law360 reporting the scheduling order; the order text was not retrieved from PACER). Structural relief is among the remedies sought; no settlement talks reported. Amazon's motion to push the trial to June 2027 was denied Oct 9, 2026 — also reported, order text not retrieved.
• Prime settlement (Sep 25, 2025; Tier 1 — FTC releases). $1.0B penalty + $1.5B redress; >$845M paid by Sep 2026; the Sep 2026 order expansion (max per consumer $51→$200, automatic payments from Oct 1, 2026) keeps administration running through ~April 2027.
• New — ad auctions (Aug 31, 2026; Tier 1 — FTC release + complaint 2:26-cv-03097). FTC + 22 state attorneys general allege undisclosed "soft reserve" prices in Sponsored Products auctions since ~2018–19; the complaint alleges >$20B extracted (allegation, not a finding) — ≈26% of TTM advertising revenue ($76.1B), aimed at the highest-margin engine. First filing appearance: the Q3 10-Q. Texas sued separately the same day.
• California price case. Trial set Jan 19, 2027 (a reported agreed continuance leaves the date unconfirmed); the court denied Amazon's summary-judgment motion on its Buy Box crossclaim in April 2026. DC settled its Prime-delivery ZIP-code suit for $8.25M on Oct 1, 2026 (immaterial).
• EU DMA. The Commission's preliminary view (Jun 25, 2026) is that AWS should be designated a gatekeeper; a final decision is reported as possible in Nov 2026.
• Filings. The 10-K / 10-Q legal notes fold the FTC matters into an omnibus paragraph (no standalone accrual); quantified items: Luxembourg GDPR (General Data Protection Regulation) fine (€746M, on appeal), Italian competition fine (€1.13B → €752M, appealed Dec 2025).

• Valuation — FLAG. The governing bar rests on a normalization, and the headline uses the friendliest year in the series; the range (+15.1% → +24.2%) is printed wherever the bar appears, which mitigates but does not remove the flag. Terminal value is 60.8% of operating EV. (Provenance: the boom-normalized base — capitalizing a chosen year's economics as permanent.)
• Leverage — FLAG. The buildout is debt-financed at the margin: $132.2B carrying debt plus $25.0B of notes issued after Jun 30, 2026, funding a ~$220B/yr program whose assets depreciate over 5-year server lives — lives already shortened once (6 → 5 years, +$1.4B FY2025 depreciation); a second shortening is a named, precedented risk. Mitigants: maturities termed out to 2076, no covenants, $123.0B liquid.
• Management & ownership — FLAG. The largest guided number (capex) has missed upward two years running; buybacks are zero for 3.5 years while the count drifts +1.3%/yr — the unused $6.1B authorization is signaling, not policy. PASS items in the same category: the revenue/operating-income guide record and the numbered bad-news disclosure pattern. Bezos's programmatic 10b5-1 selling is disclosed context, not a flag.
• Moat — FLAG. The capacity constraint that evidences the moat can invert: if lab demand falters, reserved capacity and 5-year depreciation become oversupply arithmetic. And the moat's profit engine (advertising) is under direct legal attack (Regulatory). PASS item: backlog tenor lengthened to 6.4 years while doubling — demand contracted further out than at any prior point in the series.
• Personal bias — FLAG. Three named traps: (1) contracted ≠ riskless — the two labs are simultaneously the backlog, the capex justification, and $219.1B of balance-sheet marks; treating them as three corroborations is the bias the concentration sum exists to break. (2) Base-rate neglect on margins — the required ~14–15% mature owner-earnings margin has never been printed (best: 2.5%); the grid, not the narrative, is the discipline. (3) Guide halo — nine quarters of revenue/operating-income beats can smuggle credibility into the one guide (capex) never kept; the grades keep them separate.

No leverage or moat flag is left unaddressed: both surface in the verdict box and carry dated checkpoints below.

1. Net sales vs $197.0–202.0B — fail: below $197.0B, the first below-range revenue quarter in the scoreboard.
2. Operating income vs $22.5–26.5B — and how much of any beat is one-offs (the checkpoint is whether management's flagging continues).
3. Capex: Q3 run-rate vs the ~$220B frame — fail: any raise of the frame; and whether the reason is again cost (memory) or volume.
4. Capacity language: does "not enough capacity" survive into the 2027 guide season, and is any relief horizon dated (the Q2'25 walk-back is the grading precedent)? Binary.
5. Backlog: remaining performance obligations vs $496B, weighted life vs 6.4 years — a sequential decline fails the capacity debate's threshold.
6. AWS growth/margin vs +36.8% / 39.4% in Q2 — and any Trainium3 / 2027 capacity-reservation color.
7. Q3 10-Q legal note: first filing appearance of the ad-auction suit; any accrual (binary: quantified or not).
8. Buybacks/share count: any repurchase at all (zero for 14 straight quarters of filings — binary); count vs 10,783M.
9. Investment marks: Anthropic/OpenAI carrying values; any change in the Anthropic facility's disclosure.

• Dispositive-motion rulings in the FTC case after Aug 11, 2026 (Dkt. #699) were not retrieved from PACER; the Mar 29, 2027 trial date and Oct 9, 2026 continuance denial are press-reported (MLex / Law360), labeled "reported," not verified.
• The Q3 2025 and Q3 2024 call notes did not capture the following quarters' guide ranges — those scoreboard cells are marked untestable, not filled.
• The calls' FY2025 capex figures (~$125B cash vs a ~$131B comparison figure) are unreconciled in the notes; the filing-based series in The record is authoritative.
• The Anthropic financing facility's pricing is undisclosed; OpenAI's $28.7B Series C consideration mix (cash vs AWS credits) is undisclosed; the Anthropic marks are Level 3 / recent-round marks, not cash. The labs' ability to fund ≥$238B of commitments is not examinable from Amazon's filings — counterparty quality is the page's largest unexaminable input.
• Leo segment economics are undisclosed (only guided cost headwinds); MGM/One Medical returns undisclosed; segment-level capex is never broken out — the Reverse DCF normalization is therefore consolidated, by necessity.
• No investor day / capital markets day in the last 24 months (none found) — the investor-day check is satisfied by absence; strategy claims rest on calls and filings only.

1. Anthropic carrying value (Buildout economics). v2 stated Anthropic was "carried at ~$220B combined." Wrong: Anthropic alone is carried at $190.4B. The $219.1B figure is Anthropic plus OpenAI preferred ($28.7B) and is now stated as the combined figure wherever it appears. (~$29.6B overstatement, corrected to the dossier's own arithmetic.)
2. Q2 2026 Anthropic-related income (The business). v2 stated a "$52.5B pre-tax Anthropic mark." The filing figure is $53.4B of pre-tax other income, primarily from the Anthropic investments (release; 10-Q other income, net = $53,415M; the Anthropic preferred upward adjustment alone is $50,486M). The $52.5B figure matches no line in the filing and is withdrawn.
3. Capacity-constraint call count. Reconciled to a nine-call arc — a soft acknowledgement (Q2 2024) followed by explicit constraint language on eight consecutive calls (Q3 2024–Q2 2026), replacing v2's unqualified "eight consecutive calls" everywhere.
4. Holdings data removed. v2's header carried a position-context line. Under the Oct 16, 2026 holdings ruling, reports disclose no holdings; the line is removed and no holdings data appears on this page.
5. Disclaimer added. This page previously carried no disclaimer footer; the locked house footer now appears below.
6. Presentational (v4). Verdict box, Key Debates, driver decomposition, breakeven grid and Downside Checklist are new; SWOT is abolished as a section — its factors are migrated, not dropped. The normalization range and grid are new analysis on unchanged v2 machinery, not corrections.
7. Verdict box — downside flags surfaced (Oct 16, 2026). The verdict box previously surfaced neither the Downside Checklist's leverage flag nor the full moat flag — a breach of this page's own veto-gate rule (V4-14), under which no checklist flag may stay buried below the verdict. A "Downside flags (veto gate)" line naming both flags, in the checklist's wording, was added to the verdict box on Oct 16, 2026: Leverage — FLAG (the buildout is debt-financed at the margin — $132.2B carrying debt plus $25.0B of notes issued after Jun 30, 2026, funding a ~$220B/yr program whose assets depreciate over 5-year server lives, lives already shortened once, 6 → 5 years) and Moat — FLAG (the capacity constraint that evidences the moat can invert — if lab demand falters, reserved capacity and 5-year depreciation become oversupply arithmetic — and the moat's profit engine, advertising, is under direct legal attack). No figure, bar, grade, checkpoint, or threshold changed.

SWOT migration, attested item by item: strengths (backlog, AWS scale, advertising, silicon, liquidity) → Verdict quality line, The business, and Fisher items 1–3 and 6; weaknesses (negative reported cash flow, capex-guide misses, zero buybacks, undisclosed Leo economics, thin International margin) → the bar row, the Capex regime debate, Reverse DCF buyback credit, Gaps, and The record's segment table; opportunities (reserved-capacity monetization, Trainium, ads surfaces, Leo, International convergence) → the buildout economics and Fisher items 1 and 4; threats (FTC trial, ad-auction suit, EU designation, component inflation, lab concentration) → Key debates, Regulatory, and the Downside Checklist, each sized and dated. No factor was dropped.

1. SEC EDGAR: FY2025 10-K, FY2023 10-K (FY21–FY23 comparatives), Q1/Q2 2026 10-Qs, Q2 2026 release (Ex-99.1, Jul 30, 2026), Q4 2025 release (Feb 5, 2026), 2026 proxy DEF 14A (Apr 9, 2026, accession 0001104659-26-041026), full Form 4 population Oct 2024–Oct 2026 (154 filings, with a live-EDGAR refresh on the re-verification date).
2. Earnings-call transcripts — all 12 quarters, 2023Q3–2026Q2, read in full from the desk library (Motley Fool; MarketBeat full text for 2025Q1/Q2) — quotes dated in text, paraphrases never presented as quotes.
3. Regulatory record: FTC releases Sep 25, 2025 / Aug 31, 2026 / Sep 17, 2026; California and District of Columbia attorney general releases; European Commission DMA preliminary view (Jun 25, 2026); MLex/Law360 for the FTC trial schedule (labeled reported).
4. Dates & price: Amazon Q3-date announcement (Oct 15, 2026). Price: Yahoo Finance chart API, pulled Fri Oct 16, 2026, 11:47 AM ET.
5. Full working notes (research dossier): transcript synthesis, DCF workbooks (v2/v3), price snapshot, regulatory status, Anthropic mark reconciliation, and the reverse-DCF model and its output.