Amazon.com AMZN

Reference table

1. What the business actually is — and why the vendor has it filed wrong

The vendor's profile classifies Amazon as Consumer Cyclical / Specialty Retail, with 1,576,000 employees and Andrew R. Jassy as chief executive. The classification drives a peer set of Etsy, GameStop, Casey's General Stores, Tandy Leather Factory, LightInTheBox and Revolve. That peer set is rejected in full (Section 8): the largest of them outside Alibaba is a $31B convenience-store chain, and no multiple in this dive is drawn from it.

The 10-K and 10-Q both state the actual structure: three reportable segments — North America, International, and AWS, with the chief operating decision maker being the chief executive.

Reportable segments (FY2025, from the vendor seg_geo block, which for FY2025 contains the segment table and is confirmed against the 10-K):

Product and service groups (FY2025, seg_prod — this table sums EXACTLY to reported revenue of $716.924B, so there is no double-count here):

The 10-K states "AWS sales increased 20% in 2025", which confirms the vendor's $107.556B → $128.725B (+19.7%) to the rounding. The vendor product table is corroborated by the filing and is used.

Three things follow from these tables that most coverage of Amazon misses.

First, advertising is the fastest-growing line in the company and it is not a reportable segment. At $68.635B growing 22.1%, Amazon's advertising business is larger than all but a handful of listed companies in the world, and it is buried inside North America and International. There is no disclosed advertising operating margin anywhere in the filings or the payload. Anyone modelling Amazon's margin structure without an explicit advertising assumption is guessing, and so are we — we say so rather than fabricating one.

Second, the vendor's seg_prod table is the "groups of similar products and services" disclosure, not the reportable segments. AWS appears in both, which is why it looks like a segment. North America and International do not appear in seg_prod at all. Do not mix the two tables.

Third, seg_geo is corrupted for every year except FY2025, and the 10-K explains exactly how (Section 8). The FY2024 row reads "North America Segment $438.015B, International Segment $93.832B, Germany $40.856B, Japan $27.401B, United Kingdom $37.855B" — which sums to $637.959B, precisely FY2024 revenue. But the FY2024 North America segment was approximately $387.5B, not $438.0B. The 10-K contains two separate tables: reportable segments, and "net sales attributed to countries that represent a significant portion of consolidated net sales". The vendor's FY2024 row is the country table with the United States mislabelled "North America Segment" and rest-of-world mislabelled "International Segment". FY2025's row is the genuine segment table. The two years are therefore not comparable and no year-on-year segment growth rate is computed from seg_geo.

The one thing this file cannot give us, and it matters. The brief for this dive correctly identifies segment operating income as more important than segment revenue for Amazon. Neither source supplies it. The vendor payload has no segment profit field at all. The SEC filings contain the table — the 10-Q says "Operating income by segment is as follows (in millions)" at the top of page 32 — but every numeric table has been stripped by the text extraction, in both the 10-K and both 10-Qs. What survives is the narrative, which is directionally useful (North America up on unit and advertising sales; International up on the same; AWS up on sales, partially offset by infrastructure spending) but carries no numbers. AWS operating margin is therefore not stated anywhere in this dive, because it cannot be sourced. This is stated as a gap rather than estimated around.

2. The single most important thing to understand about these numbers

Two-thirds of Amazon's reported Q2'26 profit is an unrealised mark on a private company that is also one of its largest customers, and which Amazon is itself partly funding.

The 10-Q filed 2026-07-31 states it directly:

> "Other income (expense), net was $1.1 billion and $53.4 billion during Q2 2025 and Q2 2026, and $3.9 billion and $69.1 billion for the six months ended June 30, 2025 and 2026 … The net gain of $53.4 billion in Q2 2026 and $69.1 billion for the six months ended June 30, 2026 is primarily from upward adjustments for observable changes in price relating to our nonvoting preferred stock in [a privately held artificial-intelligence developer] and the reclassification adjustment for the gains on available-for-sale debt securities from the portions of our convertible notes investments … that were converted to nonvoting preferred stock."

The arithmetic, stated precisely.

Derivation, labelled as ours: the H1'26 income tax provision was $27.8B and "included $15.9 billion of net discrete tax expense primarily attributable to the upward adjustments" (10-Q). Allocating that discrete charge pro-rata to the $53.4B of the $69.1B six-month gain that fell in Q2 gives roughly $12.3B, leaving an after-tax gain of approximately $41.1B, or $3.77 a share on 10.903B diluted shares. $5.75 − $3.77 ≈ $1.98. Cross-checking independently from operating income — $27.461B plus roughly nil net interest, taxed at the trailing 22.6% effective rate — gives $21.3B, or $1.95 a share. *The two methods agree within 2%. We use ~$1.95.

The same distortion runs backwards through the trailing year. Q1'26 contained the other $15.7B of the six-month gain and reported $2.78 of EPS against roughly $1.69 on the same operating derivation.

A 36.8x trailing multiple is a completely different investment case from a 22.3x one, and every screen in the market shows the latter.

The circularity is the part that deserves the most scrutiny. From the 10-Q's liquidity section: Amazon "invested $28.7 billion in [one private AI developer's] Series C Preferred Stock for the six months ended June 30, 2026", "also invested $10.0 billion in [a second developer's] nonvoting preferred stock" in Q2'26, and "Subsequent to June 30, 2026, we funded the remaining Commitment Amount of $21.3 billion." That is approximately $60B of Amazon equity invested into two counterparties. In the same filings, those two counterparties account for announced expansions of their compute commitments to AWS of $100.0B over eight years and more than $100.0B over ten years respectively — and the $53.4B earnings windfall is a markup on one of those very holdings.

We are not alleging impropriety; these are arm's-length transactions disclosed in full. But the analytical point stands and should be stated plainly: a material portion of Amazon's contracted backlog is owed by customers Amazon has helped capitalise, and a material portion of its reported earnings is an unrealised mark on those same customers. Both the backlog and the earnings are real. Neither is independent of the other. Nothing in this dive values Amazon on trailing GAAP earnings, and no forward multiple in Section 6 uses the FY2026 consensus row.

3. Growth and margin trajectory — genuinely accelerating

Revenue growth is going up, not down, which is unusual at this scale.

The 10-Q's own framing corroborates: "Sales increased 20% in Q2 2026, and 18% for the six months"; North America +16%, International +15%, AWS +37% in Q2 and +33% in the half. Q3'25's depressed 9.67% margin is explained by the 10-K: FY2025 operating income of $80.0B "includes charges of $2.5 billion we recorded in Q3 2025 related to the settlement of a lawsuit with the FTC and $2.7 billion, of which $1.8 billion was recorded in Q3 2025, of estimated severance costs primarily related to planned role eliminations." Adjusted FY2025 operating income is therefore approximately $85.2B, an 11.88% margin, against 10.75% in FY2024.

The useful-life question the brief flags — and the filing answers it precisely, in both directions. From the FY2025 10-K:

> "Effective January 1, 2024, we changed our estimate of the useful lives for our servers from five to six years, and effective January 1, 2025, we changed our estimate of the useful lives of a subset of our servers and networking equipment from six to five years."

> "The effect of this change in estimate for the year ended December 31, 2025 … was an increase in depreciation and amortization expense of $1.4 billion and a reduction in net income of $1.0 billion, or $0.10 per basic share and $0.10 per diluted share, which primarily impacted our AWS segment."

Read that carefully. FY2024's reported operating income was flattered by an extension of server lives from five to six years. FY2025's was depressed by a partial reversal of that extension, costing $1.4B of depreciation and $0.10 of EPS, concentrated in AWS. So the FY2024-to-FY2025 AWS margin comparison is contaminated at both ends by an accounting estimate, and the direction of travel in the estimate is shorter lives, justified by "the increased pace of technology development". A third asset class was also shortened — footnote (2) to the useful-life table reads "Ten years prior to January 1, 2025" — but the class is not recoverable because the table itself was stripped in extraction. This is the single most important accounting sensitivity in the company and management has moved it twice in three years.

Management's own forward guidance points down, and is half-weighted accordingly. Q3'26 net sales of $197.0-202.0B (+9-12%, or "nearly 400 basis points higher" excluding the timing of the Prime Day promotion in both years) and operating income of $22.5-26.5B against $17.4B in Q3'25. The midpoint of $24.5B on $199.5B is a 12.3% operating margin, down from 13.69% in Q2. Note also that the vendor's Q3 consensus revenue of $201.837B sits above the guidance midpoint and near its top.

4. Balance sheet — and the enterprise-value rebuild the vendor cannot do

The vendor payload contains two balance sheets seven months apart and does not tell you. The bal_a block is dated 31 December 2025. The ratios_ttm and km_ttm blocks are computed off a mid-2026 balance sheet. Mixing them understates debt by roughly $70B. Proof, from the vendor's own arithmetic: debtToEquityRatioTTM of 0.4047 against a bookValuePerShareTTM of $51.28 on ~10.757B shares implies equity of ~$551.6B and total debt of ~$223.2B; bal_a for 31 December 2025 shows equity of $411.065B and total debt of $152.987B. Anyone reading bal_a alone concludes Amazon carries $153B of debt. The correct current figure is roughly $223B, and rising.

What the filings actually say, at 30 June 2026:

• Cash, cash equivalents and marketable securities: $123.0B (10-Q, stated at fair value; identical to 31 December 2025 despite $82.4B of debt raised in the half — it all went out again)
• Unsecured senior notes: $132.1B carrying value, fair value $123.8B, combined weighted-average remaining life 14.2 years
• Notes fair value at 31 December 2025: $61.1B — the note balance roughly doubled in six months
• Financing obligations: $415M current + $8.9B long-term = ~$9.3B, weighted-average remaining term 14.9 years, imputed rate 3.2%
• Finance leases: gross assets $55.6B, accumulated amortisation $39.8B → net ~$15.8B; only $2.1B of finance-lease additions in H1'26 — finance leases are small and shrinking in relevance
• Commercial paper: $30.0B programme, ZERO outstanding. Revolvers: zero drawn
• Subsequent to 30 June 2026: $25.0B of notes issued (the 8-K of 2026-07-09 gives $24.923B gross, $24.867B net), maturities 2029-2066, stated coupons 4.60%-6.25%, effective rates 4.67%-6.33%
• A further C$13.967B was raised on 2026-06-12 (8-K), inside the quarter

Synthos enterprise-value rebuild, pro forma for the July issue:

The vendor's enterprise value is approximately right by accident, and its net-debt figure is wrong by $45B. The implied vendor net debt is $3,129.253B − $2,984.235B = $145.019B. Against total debt of ~$223B that implies netting only ~$78B of cash and cash equivalents, omitting roughly $45B of marketable securities. The same defect is visible unambiguously in bal_a: reported netDebt of $66.177B equals total debt of $152.987B less cash of $86.810B, ignoring the entire $36.219B of short-term investments sitting on the line above it. The 10-Q settles it: cash and marketable securities is $123.0B. Corrected net debt is ~$100.0B, and corrected enterprise value is ~$3,084B. The vendor's figure survives scrutiny only because it counts the full lease liability while netting too little cash — two errors partially cancelling.

What the leverage actually means. Interest expense went from $516M in Q2'25 to $1.3B in Q2'26, and $1.1B to $2.1B across the halves. On a pro-forma note balance of ~$157B at a blended rate approaching 4.5%, the annual interest run-rate heads toward $7B, roughly double the current half-year pace. Against $93.7B of trailing operating income that is comfortably serviceable — trailing interest coverage is 28.1x and will fall, not break. The leverage is not a solvency question. It is a signal about capital allocation: a company that generates $161B of operating cash flow and still needs $82.4B of new debt in six months is telling you exactly where its priorities are.

5. Cash flow — the number that decides the dive

Derivation: the 10-Q states operating cash flow of $49.5B and $71.4B for H1'25 and H1'26, and cash capital expenditure of $55.6B and $96.3B. Applied to the FY2025 full-year figures this gives trailing operating cash flow of $161.4B and trailing capex of $172.5B. The vendor's freeCashFlowYieldTTM of −0.39% implies −$11.62B, corroborating the sign and the magnitude from an independent direction.

The three things this table says.

First, free cash flow has gone negative and the trajectory is steepening. H1'26 capex of $96.3B against $55.6B in H1'25 is +73%, on top of a +59% year in FY2025. The 10-Q says capital expenditures "primarily reflect investments in technology infrastructure (the majority of which is to support AWS business growth) and in additional capacity to support our fulfillment network, both of which we expect to increase in 2026." At the Q2 run-rate of $53.1B a quarter, FY2026 capex lands near $200B — which is roughly the figure the knowledge base has been circling since February.

Second, the depreciation consequence is arithmetic, not opinion. FY2025 D&A of $65.756B was 9.2% of revenue. Capital expenditure is running at roughly 2.2-2.3x depreciation (the vendor's capexToDepreciationTTM of 2.301 independently confirms). That gap is not a saving; it is deferred cost. At $200B a year of spend against a blended asset life somewhere between five years (servers, per the 10-K) and much longer (buildings), steady-state depreciation heads toward $150B or more. Against FY2027 consensus revenue of $942.1B, that is roughly 16% of revenue against 9.2% today — a 600-to-700 basis point operating-margin headwind arriving on a fixed schedule. Whether the revenue arrives on the same schedule is the entire question.

Third, there is no shareholder return of any kind. The 10-K is explicit: a $10.0B repurchase authorisation from March 2022, "There were no repurchases of our common stock in 2023, 2024, or 2025", $6.1B remaining. No dividend, ever. Common shares outstanding plus shares underlying stock awards went 10.8B → 10.9B → 11.0B across 2023-2025, and diluted shares rose from 10.806B (Q2'25) to 10.903B (Q2'26), roughly 0.9% annual dilution. Stock-based compensation was $19.467B in FY2025. The shareholder yield in the return bridge in Section 6b is not small — it is negative.

One genuine structural asset, correctly reported by the vendor. The cash conversion cycle is negative 63.0 days (days sales outstanding 41.5 + days inventory 36.5 − days payable 140.9). Suppliers fund Amazon's working capital. This is why a current ratio of 1.033 is not a liquidity warning and is not treated as one here.

6. Valuation — priced in or room?

At $277.42 (market capitalisation $2.984T):

Why the FY2026 EPS row is rejected outright. Consensus FY2026 EPS of $11.70 is arithmetically impossible against the same file's quarterly consensus. H1'26 actual GAAP EPS was $8.53 ($2.78 + $5.75); Q3'26 consensus is $1.96; that leaves an implied Q4'26 of $1.21, against $1.95 actually delivered in Q4'25 — a 38% year-on-year decline nobody forecasts. The row is a blend of models refreshed and not refreshed for the Q2 investment mark. All forward valuation in this dive runs off FY2027 ($10.42, 41 analysts) and FY2028 ($13.57, 28 analysts).

A useful verification the file supports. FY2025 consensus EPS was $7.15268 against a reported GAAP diluted $7.17 — a 0.2% miss. Amazon's consensus estimates are on a GAAP diluted basis, not a non-GAAP one, which is why the contamination of the FY2026 row matters and why FY2027 onward, which analysts have not polluted with unforecastable investment marks, is the clean basis. For what it is worth, consensus undershot actual GAAP EPS by 7.6% in FY2023 and 7.3% in FY2024.

Estimate coverage is the best in this batch — 36-46 analysts on revenue, 22-41 on EPS. But look at the dispersion at the far end: FY2028 EPS ranges $10.14 to $23.40, a 2.3x spread, and FY2028 revenue ranges $1,012.2B to $1,420.6B. That dispersion is the AI-capital-expenditure uncertainty, quantified by the sell side itself. It is not noise; it is the honest admission that nobody knows.

6a. What today's price assumes (the inversion)

At $277.42 — 26.6x FY2027 consensus EPS and 3.98x trailing revenue on a corrected enterprise value of ~$3,084B — the price embeds roughly the following falsifiable claims:

• AWS revenue growth stays above roughly 25% through FY2028. (Consensus-derived: FY2027 revenue of $942.1B requires ~$225B of incremental revenue over two years from a $716.9B base; with retail growing at the 9-11% of the product tables, AWS must carry a disproportionate share.) Q2'26 delivered 37%. The falsifier is a single quarter below 30% while capex is still rising.
• Free cash flow turns positive again by FY2027 and the market does not de-rate for the interim. (Our number: trailing FCF is approximately −$11.1B and capex is guided higher into H2'26.) At the current $200B run-rate, operating cash flow must reach roughly $210B+ — a 30% increase from $161.4B — for FCF to cross zero. This is the most fragile assumption in the price.
• Depreciation does not compress operating margin below roughly 12% before FY2028. (Our arithmetic from Section 5: D&A heads toward 16% of revenue from 9.2% today.) Management's own Q3'26 guidance already implies a step down to ~12.3% from 13.69%. Company guidance, half-weighted, and it points the wrong way.
• The approximately $496B contracted backlog converts substantially at contracted economics. (Filing-derived: 10-Q, 30 June 2026, 6.4-year weighted-average life.) This is the strongest fact in the bull case. It also depends on counterparties into which Amazon has itself invested approximately $60B.
• The market keeps paying roughly 26x forward earnings for a business with negative free cash flow, no shareholder yield, and 0.9% annual dilution. (Derived from the current multiple.) Amazon's forward multiple has historically been justified by the promise of eventual cash generation. It is now being asked to survive a period in which that cash generation goes backwards.

6b. The return bridge (why the multiple moves)

Expected return over 12-24 months decomposes as: EPS growth (+30% from FY2027's $10.42 to FY2028's $13.57 on consensus) + multiple drift (COMPRESSION assumed, from 26.6x FY27E toward ~22x FY28E) + shareholder yield (NEGATIVE, approximately −0.9% from dilution, with no dividend and no buyback).

Our base case explicitly assumes multiple compression, and we should say why. The market currently pays 26.6x forward earnings for Amazon. Our $300 base is 28.8x FY2027E but 22.1x FY2028E — that is, we assume the price advances roughly in line with one year of earnings growth while the multiple the market pays at each date drifts down. The mechanism is not sentiment; it is composition. As capital intensity rises, the earnings become lower-quality: more of them are non-cash marks, more of the cash they generate is immediately reinvested, and the depreciation charge against them grows faster than revenue. A business earning $13.57 a share in FY2028 with negative or barely-positive free cash flow does not deserve the same multiple as a business earning $10.42 with a rising cash conversion rate. Peer convergence points the same way: the market is currently repricing the whole hyperscaler complex from asset-light software multiples toward capital-intensive infrastructure multiples, and today's session is what that looks like in one day.

Almost all of the expected return in our base case is earnings growth, and that is the honest framing. From $277.42 to $300 is +8.1% — roughly one quarter of the consensus FY2027-to-FY2028 EPS growth, with the rest absorbed by de-rating and dilution. The bull case at $375 requires the multiple to HOLD at ~36x FY2027E, which requires free cash flow to inflect visibly and early. That is the fragile leg and it is not in the base.

6c. Variant perception (where we differ, what would surprise)

• We differ from the street on direction, not on quality — and that is unusual enough to state plainly. The covering analysts show 84 buys, 9 holds, 1 sell and a consensus target of $322.68 (+16.3%). Our base of $300 is 7% below that. We do not think the street is wrong about the business. We think the street's target implies a multiple that does not survive the depreciation schedule in Section 5. Watchable number: the FY2027 consensus EPS line — if it starts drifting down through 2026 while the price holds, the market is re-rating upward into a downgrade cycle, which is how these things usually end.
• We differ sharply on trailing earnings quality. Every screen shows Amazon at 22.3x trailing earnings. On our derivation, excluding investment marks, the number is ~36.8x. That is not a rounding difference; it changes the name from "cheap mega-cap" to "expensively-valued mega-cap", and it is the reason a large part of the passive and quantitative bid for this stock is mispriced. Watchable event: whether the Q3 print contains another large mark in either direction — the same accounting that added $53.4B can subtract it.
• We differ from the loudest bear on the timing, not the mechanism. The most conviction-weighted claim in the knowledge base against this name — Jordi Visser, conviction 90, 2026-06-21 — calls the capital expenditure "a depreciation time bomb driving multiple compression". We agree with the mechanism and think the $496B contracted book, disclosed six weeks after that claim, is a better answer to it than the bears have yet reckoned with. That is also, notably, roughly what Visser himself concluded by 2026-08-01. Watchable number: the next disclosed remaining-performance-obligation figure. $496B rising is the bull's proof; $496B flat or falling with capex still rising is the bear's.
• We think advertising is systematically underweighted by everyone, including us. At $68.635B growing 22.1% with essentially no incremental capital, it is the highest-return line in the company and it is not a reportable segment, so no margin for it is disclosed. Watchable number: the advertising line in the next annual product table. Sustained 20%+ growth at, say, a 40-50% incremental margin would be worth several turns of multiple on its own — and we cannot verify the margin, which is precisely why nobody prices it.
• Positive surprise that would force a re-rate up: a quarter in which operating cash flow growth exceeds capex growth, i.e. visible free-cash-flow inflection; or a first use of the $6.1B repurchase authorisation, which after three years of zero would be an unmistakable signal that management believes the spending peak has passed.
• Negative surprise that would break the thesis: a further shortening of estimated useful lives. Management has already moved server lives twice in three years and the last move cost $1.4B of depreciation, "primarily" in AWS. A move from five years to four across the full server base, with $200B a year of spend behind it, would be a multi-tens-of-billions annual charge and would arrive with no warning in a 10-K footnote.

Synthos fair values

All three anchors are multiples of FY2027 consensus EPS of $10.42 (41 analysts), cross-checked against FY2028's $13.57 (28 analysts).

• Bear ~$229 — 22.0x FY2027E. Cross-check: 16.9x FY2028E; 3.35x trailing revenue. The scenario: FY2026 capex closes above $210B, free cash flow stays negative through 2027, AWS growth decelerates below 30% while the depreciation from three years of spend lands, and the market completes the re-rating of hyperscalers from software multiples toward infrastructure multiples. Note that $229 is still 2.7% above the 200-day average of $235.66 minus a normal correction, and roughly where the stock traded in the first quarter of this year. −17.4%.
• Base ~$300 — 28.8x FY2027E, 22.1x FY2028E. Cross-check: 4.3x trailing revenue; roughly 38x our derived trailing ex-mark EPS of $7.54, falling toward 22x as the earnings base doubles. The scenario: the backlog converts broadly on schedule, AWS holds above 30%, advertising keeps compounding, capex plateaus near $200B, free cash flow crosses back above zero during FY2027, and the multiple compresses in an orderly way rather than a disorderly one. +8.1%.
• Bull ~$375 — 36.0x FY2027E, 27.6x FY2028E. The scenario: the $496B book converts faster than modelled and grows; the in-house accelerator programme materially lowers Amazon's own cost per unit of compute, which is the one development that breaks the depreciation objection at its root; robotics begins visibly removing cost from the ~$90B fulfilment base; and free cash flow inflects positive in 2027 rather than 2028, taking the multiple with it. +35.2%.

Base is 8.1% above spot; asymmetry roughly 2:1 to the upside (17% down, 35% up). That is a respectable asymmetry attached to a thin base case, at a 52-week high, on a day the stock gapped down. It is a Watch, not a Buy — and Section 11 pre-registers exactly what turns it into one.

7. Knowledge base — 462 raw hits, 11 used, 451 discarded

Raw hits: 462. Used: 11. Discarded: 451. This is by an enormous margin the noisiest search in this batch, and the triage is itself a finding.

How the 462 break down. 361 matched on entity (the claim names Amazon, AMZN or AWS in its entity list) and 101 matched only on text — Amazon appearing in prose about something else entirely. All 101 text matches are discarded outright as incidental colour. Of the 361 entity matches, 260 predate 2026 and 101 are dated 2026. The 2026 entity lane runs 62 bullish, 30 bearish, 9 neutral.

Four further discard categories, with reasons:

1. Management voice — 20 of the 101 2026 entity claims, and 38 since January 2025, come from a channel of the chief executive's own remarks. Under house policy these are half-weighted at best. None is quoted as evidence here. Where the dive needs a company view it takes it from the 10-Q's guidance section, labelled as guidance.

2. Promoter-class supplier voice — 9 of the 101 2026 claims come from the chief executive of Amazon's largest equipment supplier, describing how much compute Amazon will buy. That is a voice with a direct commercial interest in the answer, discussing his own order book, and in most of those claims Amazon is a customer-mention rather than the subject. All discarded, notwithstanding high stated conviction (75-90).

3. Mega-cap list membership — Amazon appearing as one name in a basket claim about something else. Passive-flow mechanics, digital advertising as a "creative monopoly", copper supply, Metcalfe's-law network effects, COVID-era e-commerce, remote work and wage inflation, machine-vision suppliers, video licensing. All discarded as incidental.

4. Single-speaker duplication across channel labels — and this one is important. 28 of the 101 2026 entity claims are attributable to one speaker, Jordi Visser, appearing under four different channel labels (jordi_visser, jordi_visser_ai, jordi_visser_m, and as a guest on anthony_pompliano). One video generated four near-identical claims across four labels; the "S&P 500 isn't safe past 2030" claim appears four separate times between 2026-06-06 and 2026-06-17. Counting those as breadth would inflate a single opinion into a consensus. They are collapsed to roughly twelve distinct positions and attributed to the speaker, with the channel noted separately. This is exactly the failure mode the attribution policy exists to prevent, and it is live in this file.

The eleven claims used. All quoted verbatim, attributed to the speaker rather than the channel, dated.

The bear lane — capital expenditure as cost:

> 2026-06-21 · bearish · conviction 90 · Jordi Visser (independent) · entities: MSFT, AMZN, META, GOOGL

> "Core repeated trade: long his AI thematic portfolio, short the hyperscalers — 'your capex is my opportunity'; massive capex spend is a depreciation time bomb driving multiple compression."

> 2026-07-30 · bearish · conviction 72 · eurodollar_university (speaker not named in record) · entities: ORCL, AMZN, GOOGL, META, NVDA, MSFT

> "The AI bubble is a credit bubble, not an equity one — CDS spreads and weak bond demand show the debt cycle turning as capex is financed at the wrong price, risking terrible returns."

> 2026-08-01 · bearish · conviction 55 · eurodollar_university · entities: ORCL, NVDA, META, AMZN, GOOGL

> "AI-buildout debt is drawing growing caution — CDS on Oracle (215bps), SpaceX, Nvidia, Meta, Amazon, Alphabet moved higher in July; Amazon's $25B deal saw its order book shrink from $62B to $41B."

> 2026-06-18 · bearish · conviction 75 · Jeremy Grantham · entities: TSLA, AMZN, GOOGL

> "The magic seven were seven easy monopolies dominating separate niches; now they're all fighting in one AI market — capital-heavy this time, not capital-light — a bloodcurdling dogfight where only one can be first mover."

The bull lane — the backlog answers the capex:

> 2026-08-01 · bullish · conviction 72 · Jordi Visser (independent; channel: anthony_pompliano) · entities: AMZN, MSFT, GOOGL, ORCL

> "Amazon, Microsoft, Google and Oracle backlog of contracted orders nears/exceeds $2 trillion; they raised CapEx with orders to match—I'm not worried about them despite widening CDS."

> 2026-06-03 · bullish · conviction 75 · Bill Ackman (channel: all_in) · entities: MSFT, META, AMZN

> "Owns Microsoft, Meta, Amazon; calls these high-quality 'old-fashioned' companies undervalued and left behind as capital chases chips, semis and energy — analogous to Berkshire in the 2000 bubble."

> 2026-05-15 · bullish · conviction 82 · Gavin Baker (independent) · entity: AMZN · TEST: "Trainium 3 ramps in second half of the year and Trainium becomes the standout accelerator story of 2026"

> "Trainium is by far the most underestimated AI accelerator; Trainium 3 ramping in 2H will make Trainium to 2026 what TPUs were to 2025, aided by aggressive design and a working switched scale-up network."

> 2026-04-18 · bullish · conviction 75 · "Daniel" (channel: we_study_billionaires) · entity: AMZN · TEST: "robotics cuts the majority of Amazon's ~$90B fulfillment costs within about 5 years"

> "Added Amazon below $200; AWS grows into $200B capex, robotics could cut majority of ~$90B fulfillment costs, marketplace margins inflecting — earnings power beyond the last decade."

> 2026-01-09 · bullish · conviction 75 · Jacob Shapiro (independent) · entity: AMZN

> "Amazon is one of the largest positions held for years and not reduced — the one Mag-seven name trading at still-sane valuations; likes the business, including its robotics push."

> 2026-07-31 · bullish · conviction 55 · Dan Howley (journalist; channel: darius_dale) · entity: AMZN

> "AWS growth story reaching a 'show me the money' moment — accelerating sequentially, $25B AI run rate plus $25B chips run rate; selling its own chips to third parties could be a good play for Amazon."

The claim that describes today, written a week before it happened:

> 2026-07-28 · neutral · conviction 45 · Josh Brown (independent) · entity: AMZN

> "Amazon has been running in place, going nowhere; capex is the swing factor — holding the $200B guide reads as relief, hiking toward $220B is a negative even if right for the business."

What this lane establishes.

First, the lane is genuinely split and the split is on exactly the right question. Not on whether Amazon is a good business — nobody in 462 claims argues it is not — but on whether $200B a year of capital expenditure is an asset or a cost. That is a rare and valuable configuration: the disagreement is precise, dated, and testable.

Second, the most prolific bear partially recanted, and the timing matters. Jordi Visser at conviction 90 on 2026-06-21 called it a "depreciation time bomb". Six weeks later, on 2026-08-01, at conviction 72: "I'm not worried about them despite widening CDS", explicitly citing contracted backlog. When the loudest voice on one side moves toward the other within six weeks, the correct inference is that the question is unresolved, not that it has been answered. That reversal is the single largest input into the Moderate conviction rating and the Watch verdict.

Third, two claims are directly testable against the filings and both check out. The 2026-08-01 claim referring to "Amazon's $25B deal" is confirmed by the 8-K of 2026-07-09: $24.923B gross, $24.867B net, coupons 4.60% to 6.25%. The order-book detail ($62B shrinking to $41B) is not in any filing and cannot be verified, but the deal itself is real and the pricing at the long end — 6.25% on 2066 paper, 6.33% effective — is consistent with a demanding market. Separately, the "$2 trillion of contracted orders" claim across four hyperscalers is corroborated at the Amazon level by the 10-Q's ~$496B of remaining performance obligations at a 6.4-year weighted-average life. The knowledge base's central bull fact is confirmed by the filing.

8. Data integrity — what we rejected from the vendor file and why

Ten material defects. Three are cases where a filing directly contradicts the vendor.

1. seg_geo FY2024 and every prior year is the COUNTRY table mislabelled as the segment table — FILING CONTRADICTS VENDOR; we use the filing. The vendor reports FY2024 "North America Segment $438.015B" and "International Segment $93.832B" alongside Germany, Japan and the United Kingdom. Those five sum to $637.959B — exactly FY2024 revenue — but the FY2024 North America segment was approximately $387.5B. The 10-K contains two distinct tables: reportable segments, and "net sales attributed to countries that represent a significant portion of consolidated net sales". The FY2024 row is the second table with the United States mislabelled "North America Segment" and rest-of-world mislabelled "International Segment". Only the FY2025 row is a genuine segment table. No seg_geo year-on-year growth rate is computed anywhere in this dive.

2. The payload contains TWO balance sheets seven months apart and does not say so — understating current debt by roughly $70B. bal_a is dated 31 December 2025 (equity $411.065B, total debt $152.987B). ratios_ttm is computed off a mid-2026 balance sheet: debtToEquityRatioTTM 0.4047 against bookValuePerShareTTM $51.28 implies equity ~$551.6B and total debt ~$223.2B. The 10-Q settles it — unsecured notes alone were $132.1B at 30 June 2026 against a $61.1B fair value at 31 December 2025. We use the 10-Q. Never combine bal_a leverage with ratios_ttm valuation for this name.

3. netDebt omits marketable securities in BOTH blocks — a $36.2B error in bal_a and roughly $45B in km_ttm. bal_a.netDebt of $66.177B equals total debt $152.987B less cash $86.810B, ignoring the $36.219B of short-term investments on the line directly above it. Corrected 31 December 2025 net debt: $29.958B. km_ttm.enterpriseValueTTM of $3,129.253B implies net debt of $145.019B against ~$223B of total debt, i.e. netting only ~$78B of cash — the 10-Q states cash and marketable securities of $123.0B. Corrected net debt ~$100.0B; corrected enterprise value ~$3,084B. The vendor's headline enterprise value is only 1.5% too high because it counts the full lease liability while netting too little cash — two errors partially cancelling, which is worse than one, because it means the figure cannot be trusted even when it looks right.

4. capitalLeaseObligations of $87.339B is not capital leases — FILING CONTRADICTS VENDOR. The 10-Q states gross finance-lease assets of $55.6B with accumulated amortisation of $39.8B (net ~$15.8B) and only $2.1B of finance-lease additions in H1'26. Amazon's finance leases are small and shrinking. The $87.3B figure is the aggregate lease liability, overwhelmingly operating leases. Including it in totalDebt is a defensible analytical choice; labelling it "capital lease obligations" is not, and it is the classic double-count risk this file was checked for. The precise operating/finance split at 30 June 2026 is not recoverable — the 10-Q's lease-liability table was stripped in text extraction — and is stated as a gap.

5. Trailing GAAP earnings and every multiple derived from them are economically meaningless — REJECTED as a lens. priceToEarningsDilutedRatioTTM of 22.32x is arithmetically correct on trailing GAAP diluted EPS of $12.43, of which roughly $4.89 is unrealised marks on a private holding (Section 2). This dive uses derived trailing EPS excluding investment marks of ~$7.54, giving ~36.8x, and values primarily on FY2027 and FY2028 consensus EPS. Also rejected on the same grounds: priceToEarningsGrowthRatioTTM (0.249), forwardPriceToEarningsGrowthRatioTTM (−2.014, negative purely because the contaminated FY2026 estimate exceeds FY2027), earningsYieldTTM (4.53%), and returnOnEquityTTM (30.5% — the ex-mark figure is approximately 20%).

6. ratios_ttm contains two mutually exclusive "EBIT" figures and an EBITDA that is 48% too high — REJECTED. ebitMarginTTM of 0.2305 implies $178.8B; operatingProfitMarginTTM of 0.1208 implies $93.7B, which ties exactly to the sum of the four reported quarterly operating income figures. The $85B difference is the investment marks. ebitdaMarginTTM of 0.3275 implies $254.0B of trailing EBITDA; the honest figure is operating income of $93.7B plus D&A of roughly $78B, or approximately $172B. evToEBITDATTM of 12.32x and enterpriseValueMultipleTTM of 12.32x are therefore rejected; the corrected figure is approximately 17-18x (corrected EV ~$3,084B ÷ ~$172B), with the D&A component labelled as our estimate because the H1'26 depreciation line is not extractable. The vendor's own ebtPerEbitTTM of 1.8666 documents the problem from inside the file.

7. est.ebitdaAvg is corrupt — the FY2023 row is roughly 3x too high. REJECTED, per the standing data contract. FY2023 ebitdaAvg is $239.922B against FY2024's $103.746B; actual FY2023 operating income was $36.852B and D&A $48.663B, so true EBITDA was near $85B. The entire ebitdaAvg/ebitAvg series is discarded and all forward valuation runs on epsAvg, which independently verifies well: FY2025 epsAvg of $7.15268 against a reported $7.17.

8. The FY2026 consensus EPS row is internally inconsistent with the same file's quarterly consensus — REJECTED. FY2026 epsAvg of $11.70 (38 analysts) against H1'26 actual GAAP EPS of $8.53 and Q3'26 consensus of $1.96 implies a Q4'26 of $1.21, versus $1.95 actually delivered in Q4'25. It is a blend of refreshed and unrefreshed models around the Q2 investment mark. No forward multiple in this dive uses it.

9. The peers block is unusable — REJECTED in full. Etsy ($8.2B), GameStop ($8.6B), Casey's General Stores ($31.5B), RH ($3.6B), Revolve ($1.9B), Liquidity Services ($1.2B), ATRenew ($1.5B), LightInTheBox ($30M) and Tandy Leather Factory ($20M) are not comparables for a $2.98T company deriving its profit growth from cloud infrastructure and advertising. Alibaba ($309B) is the only member with any claim to relevance and is a different business in a different regulatory regime. No peer multiple appears in this dive. The economically relevant cross-references for Amazon are its counterparties, not its peers — the semiconductor and equipment names in this batch — and they are used that way in Section 9.

10. researchAndDevelopementToRevenueTTM of 15.61% is not research and development — FILING CONTRADICTS VENDOR. The vendor maps Amazon's "Technology and infrastructure" line ($108.521B in FY2025) to R&D. The 10-Q defines that line as including "servers, networking equipment, and data center related depreciation and amortization, rent, utilities, and other expenses necessary to support AWS and other Amazon businesses." It is largely cloud cost of goods sold, not research. A 15.6% R&D-to-revenue ratio for Amazon is not comparable to a 15.6% ratio at a software company and is not used.

Also rejected as meaningless rather than wrong: grahamNumberTTM ($120.47), grahamNetNetTTM (−$31.23) and netCurrentAssetValueTTM (−$294.8B) — Graham-style balance-sheet screens have no application to a company of this structure. priceToFreeCashFlowRatioTTM of −256.71x is arithmetically correct on negative free cash flow but conveys nothing; free cash flow is reported in dollars throughout this dive instead.

A technical-block disagreement, flagged. quote.yearHigh is $281.07 and quote.yearLow is $196.00; the tech block gives $284.02 and $198.79. The quote's $281.07 is identical to today's intraday high and the tech block's $284.02 is identical to yesterday's close, so the quote field appears to be a partial-session artefact. This dive uses the tech block ($284.02 / $198.79) throughout, because the derived fields (drawdown from peak, percentage from high) are computed on the same series.

Correctly reported and used: seg_prod sums exactly to reported revenue in every recent year — no double-count — and its AWS growth rate is confirmed by the 10-K's "AWS sales increased 20% in 2025". capexToDepreciationTTM of 2.301 is consistent with the cash-flow statement. freeCashFlowYieldTTM of −0.39% correctly signals negative free cash flow and corroborates the filing-derived −$11.1B. cashConversionCycleTTM of −63.0 days is real and structural. The zero-buyback and zero-dividend fields are confirmed verbatim by the 10-K.

Non-equity tripwire — checked and passed. AMZN is common stock: price $277.42 (not par-like), beta 1.461 (not bond-like), zero dividend, 65.4M shares traded today, and a 52-week range of $198.79-$284.02 (a 43% band). Not a preferred, not a baby bond, not an ADR (isAdr false). The check was performed and passed.

9. Technicals — at the high, and gapping the wrong way

• Price $277.42, −2.32% today, on volume of 65.4M. Opened at $276.06 against a prior close of $284.02 — a gap down of 2.8% at the open — traded $275.82-$281.07, and closed near the middle of that range without filling the gap.
• $284.02 was the 52-week high, set at yesterday's close. The stock is 2.32% below its 52-week high and 39.6% above its 52-week low of $198.79. max_dd_from_peak is −2.32%, meaning today's decline is the entire drawdown from the peak.
• Above both moving averages, by a wide margin: +12.2% over a 50-day average of $247.19 and +17.7% over a 200-day average of $235.66, both rising. RSI 61.5, MACD +4.81. Structurally this is an uptrend in good health. Tactically it is extended.
• Relative performance is the tell. 12-month +29.2% against SPY +24.3% and QQQ +30.8% — ahead of the broad market, marginally behind the growth index. 6-month +14.2% against SPY +11.1% and QQQ +15.7% — the same pattern. 3-month +2.0% against SPY +7.6% and QQQ +7.7% — a 5.6-point deficit. Amazon has been underperforming for a quarter, which is precisely what Josh Brown described on 2026-07-28 as "running in place, going nowhere".

Today's move and what it does to the entry

This is the most informative single session in the batch, and for Amazon it points the opposite way to everything else.

Every name that rose today sells into the capital expenditure Amazon is making. The one name that fell is the company making it. Nine names up between 4% and 13%; the payer down 2.3%. That is not sector rotation and it is not noise. It is the market pricing hyperscaler capital expenditure as a cost on the buyer's income statement and as revenue on the seller's — which is, accounting-wise, exactly what it is.

The knowledge base named this trade explicitly and repeatedly before today. Jordi Visser, conviction 90, 2026-06-21: "long his AI thematic portfolio, short the hyperscalers — 'your capex is my opportunity'." Today that trade paid off in a single session, in the same direction and roughly the same magnitude the claim implied. For a research process built on expert-versus-consensus divergence, that is a claim graded correct in real time, and it should be weighted accordingly.

So what does it do to the entry?

It makes it worse, and the honest answer is to wait. The three arguments:

1. This is a gap DOWN from a 52-week high, which is the mirror image of the problem the rest of this batch has. For the semiconductor names the question is whether to buy into a one-day gap up. For Amazon it is whether a gap down from the highs is a dip or the first day of a de-rating. You cannot know on day one, and the cost of finding out is one quarter of patience with an 86-day gap to the next print.

2. The base case does not pay you to be early. $300 against $277.42 is +8.1%. That is not enough to justify buying an extended chart into a guided sequential margin decline with no earnings catalyst for 86 days and no dividend or buyback paying you to wait. Amazon is the only name in this batch offering literally zero carry while you hold it.

3. The structural levels are a long way down and they are where the risk/reward becomes compelling. The 50-day average at $247.19 is 10.9% below spot; the 200-day at $235.66 is 15.1% below. A retest of the 50-day would put the base case at +21%, which is a different proposition entirely.

We are not asking the reader to buy into today's move in either direction. The recommendation is to watch, with pre-registered triggers in Section 11 — one of which is simply a price. If you already own Amazon, nothing here says sell. The business is excellent, the trend is intact, and the bull case at $375 is real. This is a statement about the entry, not about the company.

10. Insiders — routine, small, all disposals

The reading, stated carefully. Every open-market transaction in this file is a sale, and there are no purchases of any kind at any price. But the sizing argues against reading much into it. The chief executive's disposals on 2026-05-21 total 20,000 shares at $261.95-$265.61, against a post-transaction holding of 2,205,766 — approximately 0.9% of his position, executed in a single day across a laddered price range immediately following a restricted-stock-unit exercise. That is the signature of a scheduled sell-to-cover or a pre-arranged trading plan, not a discretionary exit. The Worldwide Stores chief executive sold 1,000 shares against a holding of 484,527 — 0.2%.

What is worth noting is the price context. Those sales were executed at $262-$266 in May. The stock is now $277.42, so nothing here looks prescient. The genuine signal is the absence of the opposite: no insider has bought a share in the open market in the entire disclosed window, at any price, including during the run from $198.79 to $284.02. Combined with three consecutive years of zero corporate buybacks and $6.1B of unused authorisation, the complete picture is that neither the company nor its officers have chosen to buy this stock with cash at any point in three years. That is not an accusation. It is a fact that sits uncomfortably alongside a street configuration of 84 buy ratings.

11. Verdict, kill-criteria and flip conditions

Watch.

Amazon is one of the two or three best businesses in the world and this is not a call on the business. Revenue growth has accelerated for four consecutive quarters to 19.6%. AWS grew 37%. Operating margin expanded 226 basis points. Advertising compounds at 22% on a $68.6B base with no capital behind it. The company discloses approximately $496B of contracted customer obligations with a 6.4-year weighted-average life. Estimate coverage is the deepest of any listed company. Everything about the franchise argues for owning it.

What argues against owning it at $277.42, today.

The reported earnings are not what they appear. Two-thirds of the largest quarterly profit in corporate history is a non-cash mark on an illiquid private holding that is simultaneously a major contracted customer and partly funded by Amazon itself. On our derivation, the quarter beat consensus by 7%, not 216%, and the trailing multiple is ~36.8x rather than the 22.3x every screen displays.

The cash flow is negative and the trajectory is steepening. Trailing free cash flow is approximately −$11.1B. First-half capital expenditure rose 73% year on year to $96.3B and the company says it expects it to increase further. Capex runs at 2.3x depreciation, and every dollar of that gap is a future charge on a fixed schedule.

The funding has changed character. Unsecured notes roughly doubled in six months to $132.1B, with a further $25.0B issued after quarter-end at coupons up to 6.25% and effective rates to 6.33% on paper maturing in 2066. Return on invested capital is 8.48%. The spread between what the new capital costs and what the existing capital earns is thin.

There is no yield to wait on. No dividend. No buyback in three years, with $6.1B authorised and unused. Diluted shares rising 0.9% a year. The shareholder-yield term in the return bridge is negative.

And we have no edge. Our base fair value of $300 is below the street's $322.68, against a configuration of 84 buys, 9 holds and 1 sell. When our number is below consensus and our reasoning is a known, discussed, widely-modelled risk, the honest verdict is not Buy. Base +8.1%, bear −17.4%, bull +35.2%, at a 52-week high, on a gap-down day, with 86 days to the next catalyst. That is a Watch.

Pre-registered KILL criteria — what would move this from Watch to Avoid:

• A further shortening of estimated useful lives. Management has moved server lives twice in three years; the last move cost $1.4B of depreciation, "primarily" in AWS. A move from five years to four across the full base, against $200B a year of spend, would be a multi-tens-of-billions annual charge appearing without warning in a footnote.
• AWS growth below 30% in any quarter while capital expenditure is still rising. 37% in Q2'26 is the number to hold. The adverse combination is decelerating revenue against accelerating spend, which is the precise shape of a capacity overbuild.
• FY2026 capital expenditure closing above $220B — the level Josh Brown named on 2026-07-28 as a negative "even if right for the business" — without a corresponding rise in the disclosed remaining-performance-obligation figure above $496B.
• The remaining performance obligation flat or falling at the next disclosure. This is the bull case's single load-bearing fact and it is disclosed quarterly.
• Operating margin below 12% for two consecutive quarters. Q3'26 is already guided to roughly 12.3% at the midpoint.
• Evidence of stress in the debt issuance — a materially wider new-issue spread, a pulled deal, or reliance on the commercial paper programme, which currently carries zero drawn against a $30.0B capacity.

Pre-registered FLIP conditions — what turns this into Buy — Tactical:

• A single quarter in which operating cash flow growth exceeds capital expenditure growth, i.e. visible free-cash-flow inflection. This is the primary trigger and it is checkable in one line of the cash-flow statement.
• A price at or below approximately $247 (the 50-day average) on sector beta rather than company-specific news, which puts the base case at roughly +21% and the bear case at only −7%.
• First use of the $6.1B repurchase authorisation. After three years of zero, any repurchase is an unmistakable statement that management believes the spending peak has passed.
• The next disclosed remaining performance obligation materially above $496B with the weighted-average life holding near 6.4 years — the backlog outgrowing the capex is the cleanest possible refutation of the depreciation objection.
• Verifiable evidence that the in-house accelerator programme is materially lowering Amazon's own cost per unit of compute — Gavin Baker's conviction-82 claim carries an explicit test ("Trainium 3 ramps in second half of the year"), and the second half of 2026 is now.

Where AMZN fits in the Synthos Framework Portfolio. Not yet. A 0% position with a monitoring flag, to be sized at 2-3% in the mega-cap compounder sleeve on any of the flip conditions above. For readers who already hold Amazon, this dive argues to hold and not to add; for readers who do not, it argues to wait for either the cash-flow inflection or the price. Note the batch overlap explicitly: an investor holding AMD, Micron, Lam, KLA, Applied, Intel, Marvell or ASML already owns the receiving end of Amazon's capital expenditure. Those positions and an Amazon position are not diversification — they are opposite sides of the same question, and today demonstrated that they can move violently against each other in a single session. Sizing should be done across the pair, not name by name. Logged as a tracked Synthos call (Watch) as of 2026-08-04 at $277.42.

Single biggest risk: the depreciation schedule. FY2025 depreciation and amortisation was $65.8B, or 9.2% of revenue. Capital expenditure is running near $200B a year against a five-year life on servers per the 10-K. Steady-state depreciation from that spending level heads toward $150B or more — roughly 16% of FY2027 consensus revenue. That charge arrives on a contractual, arithmetic schedule regardless of customer behaviour. The $496B contracted backlog is the answer to it, and it may well be sufficient. But the depreciation is certain and the conversion is not, and the market began pricing that asymmetry today.

Most fragile assumption in the price: that free cash flow turns positive again by FY2027. Everything else in the valuation follows from it. If operating cash flow does not reach roughly $210B — a 30% increase from $161.4B — while capital expenditure holds near $200B, Amazon spends a third consecutive year consuming rather than generating cash, funded by debt at 4.60-6.25%, with no dividend or buyback, and a rising share count. A $3 trillion company that consumes cash for three years does not keep a 26x forward multiple, however good its business is.

Provenance & disclosures

• Traceability: 462 raw knowledge-base hits across the terms AMZN, Amazon, AWS and "Andy Jassy" over 51,928 distilled claims. 11 used, 451 discarded (kb_breadth 10 distinct speakers, net conviction split-contested). Discards by reason: 101 match: text hits where Amazon appears in prose about something else, all rejected outright; 260 entity matches predating 2026, retained as background but not quoted; 20 claims from a channel of the chief executive's own remarks, half-weighted under house policy and quoted nowhere; 9 claims from the chief executive of Amazon's largest equipment supplier, discarded as promoter-class voice with a direct commercial interest and Amazon as a customer-mention; a large residue of mega-cap basket claims (passive flows, digital advertising, copper, network effects, COVID e-commerce, remote work, machine vision, video licensing) discarded as incidental. Critically, 28 of the 101 dated-2026 entity claims are attributable to a single speaker, Jordi Visser, appearing under four separate channel labels, with one video generating four near-identical claims — these are collapsed to roughly twelve distinct positions and attributed to the speaker, not the channel, to prevent one opinion being counted as breadth. All quotes are verbatim from the stored claim text.
• Data as-of: income statement through 2026-06-30 (Q2'26, reported 2026-07-30, five days before this dive, 10-Q filed 2026-07-31) · vendor balance sheet and cash flow 2025-12-31 (FY2025, 10-K filed 2026-02-06), superseded throughout by the 10-Q's 30 June 2026 figures where they differ · estimates 2026-08-04 · prices 2026-08-04 (quote.timestamp 1785873601 = 2026-08-04T20:00:01Z, the 16:00 ET close; $277.42, −2.32%; 50-DMA $247.19; 200-DMA $235.66; RSI 61.5) · knowledge-base claims 2026-08-04. Filings used: 10-K 2026-02-06, 10-Q 2026-04-30, 10-Q 2026-07-31, 8-K 2026-06-12, 8-K 2026-07-09, 8-K 2026-07-30.
• Where a filing contradicted the vendor, the filing won, and we said so: (i) the vendor's seg_geo FY2024 row reports "North America Segment $438.015B"; the 10-K shows two separate tables and that figure is the United States country attribution, not the segment — we use the segment table only, and only for FY2025. (ii) The vendor's bal_a shows total debt of $152.987B at 31 December 2025; the 10-Q shows $132.1B of unsecured notes alone at 30 June 2026 against a $61.1B fair value six months earlier, plus $25.0B issued after quarter-end — we use the 10-Q and treat bal_a as stale. (iii) The vendor reports capitalLeaseObligations of $87.339B; the 10-Q shows gross finance-lease assets of $55.6B with $39.8B of accumulated amortisation and only $2.1B of H1'26 additions — the $87.3B is the aggregate lease liability, overwhelmingly operating, and we say so. (iv) The vendor maps $108.521B of "Technology and infrastructure" to research and development; the 10-Q defines that line as including server, networking, data-centre depreciation, rent and utilities — it is largely cloud cost of sales and the 15.61% R&D ratio is not used.
• Data rejections (detailed in Section 8): netDebt in both bal_a ($66.177B, omitting $36.219B of short-term investments; corrected $29.958B) and km_ttm (implied $145.019B, omitting roughly $45B of marketable securities; corrected ~$100.0B, corrected enterprise value ~$3,084B); all trailing GAAP earnings multiples including priceToEarningsDilutedRatioTTM (22.32x, corrected to ~36.8x excluding investment marks), priceToEarningsGrowthRatioTTM, forwardPriceToEarningsGrowthRatioTTM (−2.014), earningsYieldTTM and returnOnEquityTTM (30.5%, ~20% ex-mark); ebitMarginTTM (0.2305, contradicted by operatingProfitMarginTTM of 0.1208 in the same block); ebitdaMarginTTM, evToEBITDATTM and enterpriseValueMultipleTTM (12.32x, corrected to approximately 17-18x); the entire est.ebitdaAvg/ebitAvg series (FY2023 row roughly 3x too high); the FY2026 epsAvg row of $11.70 (implies a Q4'26 of $1.21 against $1.95 delivered a year earlier); the peers block in full; researchAndDevelopementToRevenueTTM; grahamNumberTTM, grahamNetNetTTM, netCurrentAssetValueTTM and priceToFreeCashFlowRatioTTM as meaningless rather than wrong; and quote.yearHigh/yearLow in favour of the tech block's $284.02/$198.79.
• Basis note — the most important one in this dive: Amazon's Q2'26 net income of $62.647B includes $53.4B of pre-tax "Other income (expense), net", which the 10-Q attributes primarily to upward marks on nonvoting preferred stock in a privately held artificial-intelligence developer and to reclassification of gains on converted convertible notes. The associated H1'26 discrete tax expense was $15.9B. Derived diluted EPS excluding the mark is approximately $1.95 against a consensus of $1.82; trailing EPS excluding marks is approximately $7.54 against a reported $12.43. Both derivations are ours, labelled as such, and cross-checked by two independent methods agreeing within 2%. Amazon's consensus estimates are on a GAAP diluted basis — verified by FY2025 epsAvg of $7.15268 against a reported $7.17 — which is why the FY2026 row is contaminated by realised marks and the FY2027 row is not.
• Gaps stated rather than estimated around: segment operating income is not available from either source. The vendor payload has no segment profit field; the filings contain the table but every numeric table was stripped by the text extraction in all four filings. No AWS operating margin appears anywhere in this dive. Also unavailable: the operating-versus-finance lease liability split at 30 June 2026; the H1'26 depreciation figure (the D&A estimate of ~$78B used in the corrected EV/EBITDA is ours); quarterly segment and advertising revenue (annual only); any advertising margin; and earnings-call transcripts, which are not available on the current data plan, so management Q&A is absent throughout.
• Estimate coverage: the deepest in this batch — 36-46 analysts on revenue and 22-41 on EPS. FY2027 ($10.42, 41 analysts) is the primary forward anchor; FY2028 ($13.57, 28) is the cross-check. Note the dispersion: FY2028 EPS estimates range $10.14 to $23.40 and FY2028 revenue $1,012.2B to $1,420.6B — a 2.3x and 1.4x spread respectively, which is the sell side's own quantification of the capital-expenditure uncertainty.
• Peer note: the vendor peer set (Etsy, GameStop, Casey's General Stores, RH, Revolve, Liquidity Services, ATRenew, LightInTheBox, Tandy Leather) is rejected in full and no peer multiple is drawn. The economically relevant cross-references are Amazon's capital-equipment counterparties — AMD, Micron, Lam Research, KLA, Applied Materials, Intel, Marvell, ASML and Broadcom — whose 2026-08-04 closing moves are used in Section 9 exactly as supplied in the batch reference table, with no other figure for those names invented.
• Fair-value caveat: the $229 / $300 / $375 anchors are multiples of FY2027 consensus EPS of $10.42 (22.0x / 28.8x / 36.0x), cross-checked against FY2028 consensus EPS of $13.57 (16.9x / 22.1x / 27.6x). Stated scenario arithmetic, not a discounted cash flow. The base case assumes multiple compression, and the reason is stated in Section 6b rather than assumed away. The bull case requires the multiple to hold near 36x FY2027E, which requires a visible free-cash-flow inflection, and is flagged as the fragile leg.
• Timing: Q2'26 was reported 2026-07-30, five days before this dive — a clean post-print window with the freshest quarter fully in the data and the 10-Q filed. The next print is 2026-10-29, 86 days away; no earnings banner is warranted. The company has already guided Q3'26 revenue to $197.0-202.0B and operating income to $22.5-26.5B against $27.461B in Q2 — a guided sequential decline, half-weighted as company guidance but noted as an explicit verdict input.
• Half-weight management framing applied throughout: Q3 guidance, the "we expect capital expenditure to increase in 2026" language, and all knowledge-base claims sourced from the chief executive's own channel are discounted and labelled. None of the eleven quoted claims comes from company management.
• Not investment advice. Independent research, educational and informational only, never personalised. No recommendation to buy, sell or hold any security is made to any person.
• Version: 2026-08-04-full.