Amazon.com (AMZN)

Investment Committee Summary

Investment Thesis

At $262 (25 August 2026) Amazon trades on 35 times forward earnings. That price accepts continued AWS growth and gives partial credit for the profit mix shifting toward AWS and Advertising — segments carrying margins several times those of the retail businesses that still supply most of the revenue — while discounting the capital bill behind them. The engine's anchors disagree with the tape: probability-weighting the scenario tree gives $247, the twelve-month target is $257, and the triangulated fair value is $229, leaving the shares trading rich to that anchor set by -13%. The gap is a capital-intensity gap. Consolidated capex is running far ahead of depreciation, so the cash-flow anchors are suppressed today while the Sustained Build state has to arrive to justify them later, and a lengthened server useful life flatters near-term operating income rather than cash. Group operating margin of 14% and stock compensation at 2.6% of revenue are the honest starting points, not the segment margins. HOLD follows from the spread, not from a view that the franchise is impaired. The most damaging risk is the AWS-deceleration path — the heaviest single bear weight — whose target lands below the 52-week low.

Narrative drafted 2026-08-16 by claude-opus-5 under supervision; reviewed by Marinus 2026-08-16.

The dashboard below is the whole argument on one page: spot ($262) against each valuation anchor, the scenario tree, technicals and the options-implied move.

Anti-Thesis (The Real Bear Case)

The steelman bear is capacity-led margin compression. Amazon is spending far more on capacity than it is depreciating; that gap is a deferred cost wave, not a saving. If AWS AI consumption lags the build — the AI Capex Bust case — growth fades while stepped-up depreciation compresses the AWS margin, and AWS carries a share of group profit far larger than its share of revenue. Simultaneously, North America retail margin, only recently rebuilt, gives ground back on cost-to-serve and competitive pricing, while international remains structurally thinner. The Anthropic relationship cuts both ways: pass-through compute is low-margin capacity revenue that flatters AWS growth optics without helping AWS margin, part of it is effectively round-tripped from Amazon's own investment, and Anthropic is not AWS-exclusive, so the custom-silicon validation can soften. Incremental returns on the build are unproven by construction. Earnings and the multiple then compress together, and the deceleration path prices below the 52-week low — an impairment of the profit engine, not a cyclical dip.

Key Debate

P/E Multiple explains 66% of Monte Carlo outcome variance — i.e. value is set by the multiple the market will pay, a rate/sentiment regime bet as much as an earnings bet.

What the Market Is Pricing In

At the current price, the market pays 21.1× consensus forward EPS, vs the house DCF terminal 20.0×, and a peer median 41.0×. The house DCF sits 23% below spot, so the market is pricing in more than the house case — roughly 2.3pp of revenue CAGR.

Variant perception: the house view is below-consensus, and the thesis is primarily growth-driven.

Scenario Analysis

The scenario tree spans a structural 'AWS Decel / Retail Mgn Hit' downside ($130) to a 'Ads + AWS Inflection' bull case ($370); the probability-weighted blend (PWEV $247) is -6% versus spot.

SBC charge: scenario targets are gross per-share prices; the PWEV is reduced by one year of stock-based-compensation dilution (1.0% of shares, on SBC ≈ 4% of revenue), trimming the gross PWEV of $250 to $247 (-1.0%). SBC is charged once, as dilution — never also deducted from FCF.

Scenario rationale — the driver path behind every target:

• AWS Decel / Retail Mgn Hit (20%, $130). AWS growth fades toward mid-teens as AI capacity outruns consumption while D&A from the build steps up, and North America retail margin gives back gains on cost-to-serve and competitive pricing. The AWS profit engine de-rates and the consolidated multiple compresses as the AI-capex thesis is questioned. The implied target sits below the 52-week low — a genuine structural impairment, not a pullback. Drivers — aws growth: ~14%; aws op margin: ~30%; na retail margin: ~5%; group multiple: compresses.
• Recession Overlay (10%, $180). A consumer/enterprise slowdown pressures retail units and advertising budgets while enterprises optimize AWS spend, capping group growth in the high-single digits. Margins hold better than revenue because regionalization and ad mix are structurally sticky, but the multiple stays capped until demand visibility returns. Drivers — aws growth: ~16%; ad growth: ~10%; na retail margin: ~6%; group multiple: capped.
• Base (35%, $260). AWS holds ~20% on steady migration plus AI consumption, Advertising compounds ~20% at ~40% margin, and North America retail margin grinds higher on fulfillment efficiency. Operating income mix shifts further toward AWS + Ads (the profit pillars), and the consolidated multiple normalizes on proven AI monetization and FCF inflection. Drivers — aws growth: ~20%; aws op margin: ~36%; ad growth: ~20%; na retail margin: ~7%.
• ME Bull (25%, $310). Retail operating margin expands well above trend as regionalization, automation and 3P/ads mix compound, and AWS reaccelerates above 22% on AI consumption. Group operating income inflects faster than revenue as the high-margin pillars carry the mix, and the multiple re-rates on durable FCF. Drivers — aws growth: >22%; aws op margin: ~38%; ad growth: ~22%; na retail margin: ~9%.
• Ads + AWS Inflection (10%, $370). Advertising sustains 20%+ at 40%+ margins (Prime Video ads + DSP) and AWS AI consumption inflects — Bedrock and Trainium capacity convert to high-utilization revenue, vindicating the capex build and lifting AWS ROIC. The two highest-margin pillars drive disproportionate operating-income upside and a full multiple re-rate. Drivers — aws growth: >25%; aws op margin: >38%; ad growth: >22%; ad op margin: >40%.

Valuation Triangulation

Four weighted anchors — an intrinsic dcf, a scenario-weighted pwev, a monte carlo median (student-t + regime switching) and a peer p/e re-rate — read them with their basis in mind. The Monte Carlo, the DCF terminal and any peer re-rate key off a market multiple, so they are not fully independent; only discounted cash flows themselves are genuinely multiple-free. The discipline is to read the spread and weight the cash-based view, not to treat four numbers as four independent votes.

The house blend DECLARES five anchor weights — DCF 35%, scenario PWEV 25%, Monte Carlo 15%, sum-of-parts 15%, peer re-rate 10%. For this name sum-of-parts is not computed, so 15% of the declared weight is redistributed across the anchors that exist — which is why the weights above differ from the declared ones. The fair value is unaffected by this disclosure; the blend has always worked this way.

Peer EV/Revenue re-rate — 0% weight: it duplicates the peer-multiple information already carried by the Peer P/E anchor while ignoring margin mix; weighting both would double-count the peer view. Shown as a cross-check.

sum-of-parts excluded from the weighted blend — diverges >55% from the Monte-Carlo / scenario core. For a high-leverage equity the per-share DCF (enterprise value less large net debt) is hypersensitive to the terminal multiple; a peer re-rate across heterogeneous margins is apples-to-oranges. Shown above for reference; the blend leans on the multiple-discipline and scenario anchors.

Monte Carlo — the outcome distribution

10,000 paths, Student-t shocks (fat tails) with a regime-switching overlay. The median lands at $211 and 30% of paths finish above spot. The variance decomposition shows the p/e multiple is the dominant swing factor (66% of variance). Value is a multiple bet: fundamentals move the answer far less than the rating does.

DCF — the cash-flow anchor

Independent of the market multiple: a 5-year path, WACC 9.5%, 20.0x terminal FCF multiple → $202. This anchor is deliberately the heaviest (41%): it is the valuation least hostage to the current multiple regime.

Peer benchmarking — relative value

Against the peer cohort, re-rating to the peer-median forward multiple (P/E 41.0x) implies $308. A premium is only justified by superior growth/margins; otherwise it is multiple risk. Weighted just 12% so market sentiment does not set the fair value.

Sum-of-parts

Valuing each piece at the multiple it deserves (North America Retail 0.8x, International Retail 0.7x, AWS 9.0x, Advertising 7.0x, Subscription / Prime 5.0x) → $70.34. 'AWS' dominates at 9.0× → $454B (64% of EV) — the segment whose multiple matters most.

Across all anchors the spread is 102% of the median — wide (genuine disagreement — the blend carries low valuation confidence).

Sensitivity

DCF/share — WACC × terminal multiple

DCF/share — revenue CAGR Δ × op-margin Δ

Tornado — DCF/share swing by driver (widest first)

Company lever — SoP/share vs AWS multiple (AI re-rating) (base 9.0x)

Peer Quality & Weighting

Quality-weighted forward P/E: 37.3× (simple median 41.0×). Direct peers count 100%, segment 50%, broad 25%.

Valuation-anchor screen: Sum-of-parts (excluded (>3× or <0.3× spot)). Anchor median 206.4. Extreme/excluded anchors carry no headline weight.

Historical-range cross-check: 52-week range $140–$284, centre $199 (-24% vs spot); spot sits at the 85th percentile of the range. Low-weight mean-reversion cross-check, not a fundamental anchor.

Risk / Reward & Margin of Safety

That ratio compares triangulated upside against the probability-weighted bear target, not the extreme tail; with a leg missing it is withheld rather than computed from a magnitude. Bull case (Ads + AWS Inflection): $370.

Company Overview & Business Model

Amazon.com — CONSUMER CYCLICAL · INTERNET RETAIL. Amazon.com, Inc. is an American multinational technology company which focuses on e-commerce, cloud computing, digital streaming, and artificial intelligence. It is one of the Big Five companies in the U.S. information technology industry, along with Google, Apple, Microsoft, and Facebook.

How it makes money.

Edge. Wide moat. Authored moat rationale withheld pending re-authoring.

Revenue-Segment Breakdown

The company-specific drivers behind the valuation — each segment carries its own growth, margin, multiple and capex intensity. (Tags: FACT reported · ESTIMATE from disclosures · INFERENCE judgment.)

AI revenue, decomposed — the AI lines broken out (Azure-AI / Copilot / model-API / pass-through style), so the AI contribution is auditable:

• AWS AI total (Bedrock + SageMaker + AI compute): Aggregate AWS AI/ML run-rate; the lines below decompose it and are SUBSETS — NOT additive
• Bedrock (model / API): SUBSET of AWS AI total — managed foundation-model API (Anthropic, Amazon Nova, Llama, etc.); consumption-priced
• Trainium / Inferentia (custom silicon): SUBSET — in-house accelerators; structural cost/margin advantage vs merchant GPU (lower $/training-hour, less Nvidia dependence). Capacity sold as Trn/Inf instances
• SageMaker (ML platform): SUBSET — build/train/deploy ML platform; more mature, slower-growing than Bedrock
• Anthropic-linked AWS compute pass-through: AWS revenue from Anthropic's own training/inference on Trainium under the compute commitment; capacity/cost-plus economics — low margin. Direct analog to the OpenAI/Azure pass-through; partly circular with Amazon's investment

Named Exposures

Anthropic relationship (FACT/ESTIMATE/INFERENCE)

AI capex & depreciation (ESTIMATE/INFERENCE)

Industry Context — AI Compute Stack

This name sits in the AI Compute Stack cluster as a buyer (hyperscaler) name. AWS capex (Trainium/Inferentia reduce NVDA reliance); a bust helps retail-blended FCF but caps the AWS-AI re-rate. Its scenarios are not guessed in isolation — they inherit a single, shared view of the cluster's driver cycle, so the names that depend on the same event are mutually consistent.

Value chain: MSFT (buyer (hyperscaler)) · GOOGL (buyer (hyperscaler)) · AMZN (buyer (hyperscaler)) · META (buyer (hyperscaler)) · NVDA (supplier — AI accelerators) · LRCX (supplier — wafer-fab equipment) · MU (supplier — HBM / memory)

Mapping note: name-level 'ME Bull' (25%) + 'Ads + AWS Inflection' (10%) map to cluster Supercycle (35%) — the cluster row is the SUM of the mapped scenario probabilities, not a different estimate.

On the cluster's key downside — AI Capex Bust (FY27 aggregate −30%+ (to ~$350B)) — this name implies 20% vs the cluster house view of 22% (in line with the house). The cluster's full cross-stock reconciliation governs that the names which ride the same capex cycle assign it comparable odds.

Structure: Concentration — Demand: 4 hyperscalers ≈ 60-70% of AI capex. Supply: NVDA dominates accelerators; TSMC is the single leading-edge fab; 3 HBM makers. (FACT/ESTIMATE). Barriers — CUDA software lock-in, HBM/CoWoS packaging supply, leading-edge fab access, networking (NVLink). (FACT). Pricing Power — Sits with NVDA today (~75% gross margin); erodes if custom ASICs (Google TPU, AWS Trainium, Meta MTIA) and AMD take share, or inference shifts to cheaper compute. (INFERENCE). Substitution Risk — Custom silicon, model-efficiency gains (DeepSeek-style $/token collapse), inference-vs-training mix shift, and the circular vendor-financing of neoclouds/OpenAI. (INFERENCE).

Balance Sheet & Liquidity

Balance-sheet data as of 2025-12-31 (Alpha Vantage).

Capital Allocation

Free-Cash-Flow Quality

Accounting quality: SBC 4% of revenue; cash conversion (OCF/NI) 180% — cash-backed.

Competitive Moat

Moat sources:

• FACT: AWS switching costs (data gravity, committed-spend contracts, re-architecture cost) on the largest cloud installed base
• FACT: fulfilment/logistics network scale and Prime membership flywheel that raise the cost of matching delivery speed
• FACT: third-party marketplace network effects (selection to traffic to sellers) plus a high-margin advertising business monetising that traffic
• INFERENCE: the moat is thinnest in first-party retail, where margins are structurally low and competition (Walmart, Temu) is intense

Earnings-Call Disconfirmation & Sentiment

Derived signals from the MCH market-data store (Alpha Vantage transcripts + news). Quantitative tone only — a disconfirmation flag, not a substitute for reading the call.

Management vs analyst tone (2026Q2): management +0.52 vs analyst floor +0.00 → delta +0.52 (n=11 mgmt / 6 Q&A; 73rd pctile across the S&P book, z +0.7).

Flag: TYPICAL — management-vs-analyst tone within the normal cross-sectional range.

News (last 365d, 2663 articles): avg ticker sentiment +0.18 (bullish 12% / bearish 1%)

Consensus & Market Expectations

_Consensus figures: Alpha Vantage sell-side aggregates. Where the house view sits materially above or below the street, the divergence is itself a datum — see the thesis.

Catalyst Calendar

• 2026-10-13 (~50d) — Prime Big Deal Days / holiday retail read-through (authored)
• 2026-12-01 (~99d) — AWS re:Invent 2026 - custom-silicon (Trainium) and AI-service roadmap (authored)
• 2027-02-15 (~175d) — FTC/DOJ antitrust case procedural milestone (authored)

Forecast Track Record

• EPS surprise: beat 88% of the last 8 quarters; average surprise +24.0%.
• Prior-forecast backtest (23 snapshots, 2026-04-24→2026-08-20): directional hit-rate 100%; mean predicted +2.0% vs realised +3.5%. Disconfirming track record is reported, not suppressed.

Catalyst Timeline

5 catalysts in the next 90 days (of 15 tracked). Importance 1–3; confidence 0–1.

_Sources: extended.catalysts, data/catalysts/.json, data/catalysts/macro.json, AV DIVIDENDS, opex 3rd-Friday calc.

Regulatory & Legal Risk

Probabilities and sensitivities are analyst estimates, not market-implied.

Scenario Macro & Key Risks

Decision Rules (Machine-Checked)

Stance: Hold — 0 bullish / 0 bearish / 0 caution rules triggered of 6 evaluable (0 lacked data).

Machine-checked rules over disclosed inputs. The authored falsification triggers elsewhere in this report are analyst judgment and are NOT evaluated here. The stance is a portfolio-management signal and does not modify the published research rating.

Reasons the Thesis Could Fail (Falsifiable)

Pre-registered signals that would break the thesis — each polices a specific scenario boundary and is checked at every earnings update:

• AWS constant-currency revenue growth (YoY) < 17% for two consecutive quarters (2 consecutive prints). (rationale withheld pending re-authoring — frozen figure or verdict)
• North America segment operating margin < 6% for two consecutive quarters (2 consecutive prints). (rationale withheld pending re-authoring — frozen figure or verdict)
• Full-year capex plan revision > 15% cut versus the prior stated full-year plan (single event). A deep cut relieves near-term free cash flow but concedes doubt about AI demand; the cluster reads it as the bust signal and the AWS multiple de-rates with it.
• Advertising revenue growth (YoY) < 15% for two consecutive quarters (2 consecutive prints). (rationale withheld pending re-authoring — frozen figure or verdict)
• Server useful-life assumption / datacentre asset impairment reversal-or-writedown any shortening of the ~6-year server life or an impairment charge (single event). The life extension flatters current D&A and operating income; a reversal or writedown concedes that capacity was built ahead of demand and pulls the deferred cost wave forward.
• Anthropic training-compute allocation to AWS / Trainium material shift away announced migration of primary training workloads to a non-AWS platform (single event). Anthropic is the anchor Trainium customer; losing primary-training status removes both the custom-silicon validation and the pass-through revenue that supports the AWS AI narrative.

Fact / Inference / Speculation

• FACT: Spot $262; 52-week range $140–$284; engine rating HOLD; house target $257 (-2%). (source: Alpha Vantage 2026-08-24, 25 August 2026)
• INFERENCE: Triangulated FV $229 (-13% vs spot · triangulated FV); the rating tracks the Monte-Carlo + scenario-PWEV core; the cash-flow anchor sits below the multiple-discipline core.
• SPECULATION: At current prices the embedded bet is that the market keeps paying the current multiple through the capex cycle — a regime call the engine cannot verify from fundamentals alone.

Conviction Score

Research Conviction measures the strength and quality of the research setup — input coverage, financial quality, valuation support, process agreement. It is NOT the probability that the recommendation succeeds; the forecast track record section is the evidence on that.

67.0/100 (confidence band 59.1–75.0), 83rd percentile of 858 covered names (as of 2026-08-24). Weighted composite under config ros-1.19.0 — every component and its inputs below.

Score history: 68.3 → 68.3 → 68.4 → 68.5 → 68.5 → 67.5 → 66.9 → 66.9.

Probability-Weighted Return Profile

Horizon: 1 year — at this horizon CAGR equals total return by definition. Expected values are the probability-weighted sums over the full scenario set below.

The gross expected return is the probability-weighted scenario return (it reconciles to the gross PWEV); the diluted figure applies the SBC share-count charge and matches the published PWEV-based target.

Expected Alpha

Expected return minus the return this name is REQUIRED to deliver for its risk (1-year horizon). Constants are pre-registered (preregistered-static (amendment #2, 2026-07-29); not fitted to MCH outcomes).

A negative expected alpha does not change the rating — it says the expected return does not clear the risk-adjusted hurdle at today's price. Rating mechanics are unchanged by this section.

Probability Cross-Checks

Not authoritative. The scenario probabilities in this report are AUTHORED — a judgement about how the world might go, not a measurement. Nothing below modifies them, the target, the rating or any position size. These checks ask only whether anything outside our own model agrees with us; where it does not, that is information for the reader, not a correction we have quietly applied.

Authored set: 5 scenarios, probabilities summing to 1.0, mean target $249.66.

Factor Exposures

Cross-sectional percentiles over 858 covered names (style scores sector-demeaned; thematic = return-beta to the theme's proxy ETF). 50 = estate median.

Market interaction: correlation vs SPY +0.67, vs QQQ +0.67 (trailing ~1y daily returns).

Options Intelligence

Preferred structure: Long Stock. The selector reads the equity view (direction) and the volatility surface — nothing here re-prices the chain.

• no directional edge and options are cheap — options add little; hold the stock
• Direction neutral from the overlay conviction/rating (read-only input).
• IV/RV at the 2nd percentile of the cross-section → low vol bucket. This is the measure that selects the structure above: it ranks how rich this name's implied vol is against its own realised vol, relative to other names.
• Reported alongside and not used to select: this name's own ATM IV sits at the 62nd percentile of its own month-end history (decile 7). The two measures disagree here — vol is high for this name by its own history while its options are not unusually rich against its realised vol. Where they diverge, the cross-sectional measure is the one acting.
• IV term structure is in contango (longer-dated richer, slope +7.8pp) — favour longer-dated ownership (LEAPS) or calendars that are long the cheaper front.
• No live-chain Long Stock was priced for this name — shown as the indicated strategy; size against a freshly pulled chain.

IV term structure (contango, slope +7.8pp): 32-DTE 30% · 88-DTE 36% · 389-DTE 38%

No live-chain Long Stock was priced for this name — shown as the indicated approach; size against a fresh chain.

Alternatives: . IV rank shown via the name's own monthly IV history (advisory); structure selected on the cross-sectional IV/RV percentile. Model output for research and education only — not individualised investment advice. Position-sizing and stance figures are mechanical outputs of the disclosed rules and inputs; they do not consider any person's objectives, financial situation, or tax status.

Position Sizing Framework

Model output for research and education only — not individualised investment advice. Position-sizing and stance figures are mechanical outputs of the disclosed rules and inputs; they do not consider any person's objectives, financial situation, or tax status.

Options Overlay

A defined-risk way to express the HOLD equity view. Chain as of 2026-08-24 — end-of-day marks — indicative, not executable quotes.

Market signals — ATM IV 29.8% (subdued regime) · expected move ±7.0% (2026-09-25) · put/call OI 0.71 · ATM Δ 0.57 / Θ -0.16 / ν 0.30. Direction: NEUTRAL (implied return -12.6% to triangulated fair value $229.16).

Covered Call (if held) (Income / neutral) — Short 280 C · 2026-09-25 · premium $3.27 · yield 1.2% · priced from the listed chain (EOD marks)

Converts a near-fair holding into income by agreeing to sell at a higher strike — worth weighing when upside looks limited near fair value and being called away is acceptable. Illustrative — no outcome is implied or guaranteed.

Put Spread (income) (Income / would-own) — Short 240 P / Long 225 P · 2026-10-02 · net $1.59 · net entry $238.41 · yield 0.7% · RoR 12.0% · max loss $13.40 · priced from the listed chain (EOD marks)

Gets paid to wait for a lower entry, with the tail capped: the sold put collects premium while the cheaper long wing below it caps the maximum loss at the spread width — a defined-risk alternative to a naked cash-secured put.

Protective Collar (if held) (Hedge) — Long 235 P / Short 290 C · 2027-02-19 · net $5.32 · floor -10.0% · cap +11.0% · priced from the listed chain (EOD marks)

For a holder who wants to stay invested but cap risk: the sold call funds most of the protective put, fencing the position into a band — a way to ride out a de-rating without selling.

Indicative options structures — a defined-risk expression of the equity view, model-and-market illustrated from the last-close chain. NOT personalised advice or an executable quote; premiums, IV and greeks move intraday. Options carry the risk of total loss of premium. Not a registered financial adviser.

Rating Bridge

Rating = HOLD because:

• Probability-weighted scenario value implies -6% vs spot
• Monte Carlo median implies -19% vs spot
• DCF fair value implies -23% vs spot — but this is terminal-value sensitive (exit-multiple $202 vs Gordon $156, 22% apart), so it carries less weight
• Bear case (AWS Decel / Retail Mgn Hit) downside is -50% vs spot
• Net: the valuation anchor itself sits 12.6% below spot, so there is no reward leg to weigh against the bear case and the reward-to-risk ratio is withheld rather than computed off a negative upside. The rating is not asymmetric enough for a Buy and not impaired enough for a Sell — hence Hold.

Model Appendix

DCF — line items

FCF is bridged: NOPAT + D&A − Capex − ΔNWC (capex intensity 12% of revenue, weighted from the segments) — not a single conversion fudge.

WACC 9.5% · Σ PV(FCF) $313B + PV(terminal) $1804B = EV $2117B; + net cash $40.0B → equity $2157B ÷ diluted shares $10.70B = $202/share (exit-multiple terminal).

• Gordon terminal at 2.5% → $156/share — a genuinely non-multiple, cash-based cross-check; the exit-multiple and Gordon values bracket the terminal-value risk.
• Incremental ROIC on the forecast capex ≈ 9% vs WACC 9.5% → below WACC — the incremental build is value-dilutive.

Peer set

Implied prices at the peer medians: peer-median fwd P/E → $308; EV/Rev → $323.

Weighted fair-value math

Assumption Register

Sensitivity-ranked drivers (widest fair-value swing first): Capex intensity ±15% (98.0); Op margin ±3pp (88.0); Revenue CAGR ±3pp (61.0); Terminal × ±15% (51.0); WACC ±1pp (17.0).

Inputs, Sources & Confidence

Every load-bearing input, labelled by type and confidence. (reported fact · company guidance · consensus estimate · market data · house estimate · inference.)

Research Provenance

Load-Bearing Assumptions

DCF: WACC 9.5%, terminal multiple 20×, FY+5 revenue $1,263B. Triangulation leans 41% on DCF, 29% on PWEV, 18% on the Monte Carlo median, 12% on peer-implied value.