AI & Machine Learning
The neutral arms-dealer of the AI build-out — AWS + Trainium + a stake in both OpenAI and Anthropic is the best-positioned compute franchise on Earth, but free cash flow has been incinerated to ~$1B and a ~$200B 2026 capex bet on 5-year-depreciated silicon is the whole thesis.
Research
The verdict
The neutral arms-dealer of the AI build-out — AWS + Trainium + a stake in both OpenAI and Anthropic is the best-positioned compute franchise on Earth, but free cash flow has been incinerated to ~$1B and a ~$200B 2026 capex bet on 5-year-depreciated silicon is the whole thesis.
Amazon is three businesses stapled to one balance sheet, reported in three segments — North America, International, and AWS. In plain terms:
Customer sets (self-listed): consumers, sellers, developers, enterprises, content creators, advertisers, employees. Scale: ~1,576,000 full + part-time employees as of 2025-12-31; ~10.73B shares outstanding (2026-01-28); market cap ~$2.65T.
Contract structure / payment terms. Retail is transactional with negative working capital (collect from customers before paying suppliers — a structural cash float). AWS is consumption-based (pay-as-you-go) with large multi-year enterprise commitments; the AI-lab deals add take-or-pay-like compute commitments — e.g. Anthropic has pledged >$100B of AWS spend over 10 years, OpenAI ~2GW of Trainium capacity. The off-balance-sheet capacity backlog is visible in $103.8B of unconditional purchase obligations + $106.3B of leases-not-yet-commenced.
The frame that matters: Amazon is no longer "an e-commerce company with a cloud side-business." It is the neutral compute + custom-silicon arms dealer of the AI build-out that also happens to own the West's largest store and its third-largest ad network. Every lens below is really about whether the AI bet pays for the capex it demands.
Amazon sits in the middle of two distinct chains — it is both a buyer of enormous scale and a seller of infrastructure.
AWS / AI compute chain (the one that matters now):
Retail chain: millions of third-party sellers (heavily China-based) + first-party vendors → Amazon fulfillment network (fulfillment centers, sortation, last-mile delivery, air cargo) → consumer. Single-source / concentration flags: China-based sellers and suppliers account for "significant portions" of 3P services + advertising revenue and of components/finished goods — explicitly called out as a tariff/trade-policy exposure. TSMC is the chokepoint for the custom-silicon moat — if Trainium can't get leading-edge wafer allocation, the cost-advantage story stalls.
Chokepoints to mark: (1) TSMC leading-edge capacity for Trainium; (2) HBM supply (shared with the entire AI industry); (3) electrical power + grid interconnect; (4) NVIDIA allocation for the GPU portion of the fleet. Amazon's vertical move into its own silicon is precisely an attempt to de-risk chokepoint #4 by owning #1.
Amazon has arguably the widest moat stack in mega-cap tech, on four legs:
Bargaining power. Over retail suppliers/3P sellers: very high (they need the demand). Over cloud customers: high for incumbents, lower for new AI-native workloads that are multi-cloud by design. Over NVIDIA: improving — Trainium is the lever that flips Amazon from price-taker to credible alternative. Over the AI labs: symbiotic, not dominant — Amazon needs anchor tenants to fill the capex as much as they need its compute, which is why it is paying (equity + financing facilities) to lock them in.
Where the moat is thinnest: raw cloud compute is commoditizing — the market is "starting to treat cloud compute as a commodity… Cloud, increasingly, does not [have a moat]". The differentiation is migrating up-stack (data, agents, silicon), which is exactly where Amazon is investing.
All segment figures `` unless noted. FY = calendar year.
Net sales by segment ($M):
| Segment | FY2024 | FY2025 | YoY | FX-neutral YoY | Mix FY25 |
|---|---|---|---|---|---|
| North America | 387,497 | 426,305 | +10% | +10% | 59% |
| International | 142,906 | 161,894 | +13% | +10% | 23% |
| AWS | 107,556 | 128,725 | +20% | +20% | 18% |
| Consolidated | 637,959 | 716,924 | +12% | +12% | 100% |
Operating income by segment ($M):
| Segment | FY2024 | FY2025 | YoY |
|---|---|---|---|
| North America | 24,967 | 29,619 | +19% |
| International | 3,792 | 4,750 | +25% |
| AWS | 39,834 | 45,606 | +14% |
| Consolidated | 68,593 | 79,975 | +17% |
The single most important segment fact: AWS is 18% of revenue but 57% of operating income ($45,606M / $79,975M). The retail business funds the lights; AWS funds the company.
Trend / cause:
All `` unless noted.
| Metric | Q1 2025 | Q1 2026 | YoY |
|---|---|---|---|
| Total net sales | $155,667M | $181,519M | +16.6% |
| — Net product sales | 63,970 | 71,304 | +11% |
| — Net service sales | 91,697 | 110,215 | +20% |
| Operating income | 18,405 | 23,852 | +30% |
| Other income (expense), net | 2,749 | 15,647 | — |
| Income before taxes | 21,679 | 39,834 | — |
| Provision for income taxes | (4,553) | (9,560) | — |
| Net income | 17,127 | 30,255 | +77% |
| Diluted EPS | $1.59 | $2.78 | +75% |
| Diluted shares (M) | 10,793 | 10,874 | +0.8% |
Beat/miss vs. consensus: Reported diluted EPS $2.78. Note the quality caveat below — the headline number is heavily flattered by a non-cash mark.
What drove it. AWS (+28%) and advertising were the operating drivers; operating income +30% on +16.6% revenue is genuine operating leverage. But ~$15.6B of the pre-tax income is a non-cash "other income" gain — an upward markup of Amazon's nonvoting Anthropic preferred + reclassified Anthropic convertible gains. Strip it: pre-tax operating-quality income ≈ $24.2B; after a normalized ~24% rate, operating-quality net income ≈ $18-19B and diluted EPS ≈ $1.70-1.75. The clean operating result roughly matched a strong quarter; the GAAP headline doubled because of an investment markup. This is the central earnings-quality flag.
Margin moves. Consolidated operating margin 11.8% → 13.1% (Q1'25→Q1'26). Tech-and-infra opex +29% to $29.6B (AI build-out + the 5-year depreciation). Cost of sales fell as a % of sales (retail efficiency + ad mix).
Guidance / outlook. Management reiterated continued AI investment and warned macro/tariff/AI factors will keep impacting results into Q2 2026. Separately, Amazon has guided ~$200B of capex for 2026 (≈ +53% vs. 2025).
Balance-sheet flags (the real story):
Market reaction. Stock ~$240 in mid-June 2026, trailing P/E ~29x, forward ~27x — i.e. the market is paying a premium through the FCF trough, betting the capex converts to AWS revenue. The reaction to the deals was positive (stock +5% on the Globalstar/AI news days).
No transcripts/ on disk (transcripts=0) — this lens is ``.
What management is focused on (Q1 2026 call, Jassy): AI as a generational opportunity; Trainium 3 shipping and "already a multi-billion-dollar business," nearly fully subscribed; AWS AI run-rate >$15B; Project Kuiper → Amazon Leo entering the market mid-2026; defending the ~$200B capex as demand-driven, not speculative.
Tone shift over the last ~4 calls:
Read: management conviction is at a multi-year high — which is bullish if the demand is real and a classic top-of-cycle tell if it isn't. The honest tension of this name lives in that sentence.
| Company | Ticker | Mkt cap | Fwd P/E | EV/EBITDA | Cloud growth (latest qtr) | Note |
|---|---|---|---|---|---|---|
| Amazon | AMZN | ~$2.65T | ~27x | n/a | AWS +28% | #1 cloud, 30% share |
| Microsoft | MSFT | n/a | ~23.6x; sources range 23–33x | n/a | Azure +40% | #2 cloud, 25% share |
| Alphabet | GOOGL | n/a | ~19–21x fwd | ~21.6x | GCP +63% | #3 cloud, 13% share, cheapest |
| Oracle | ORCL | n/a | ~19.4x NTM | n/a | OCI fast-growing off small base | RPO-driven |
| Dividend yield | — | AMZN 0% | MSFT ~0.7% | GOOGL ~0.4% | — | AMZN pays no dividend |
| 5-yr avg ROE | — | n/a for any peer | — | — | — | (compiled data absent) |
Read. On forward P/E, AMZN (~27x) is the most expensive of the four cloud mega-caps — GOOGL (~20x) and ORCL (~19x) are cheaper, MSFT (~24x) similar. The premium is justified only by the breadth of the three-engine story (AWS reaccel + ads + the silicon/AI-lab optionality) and the depressed-by-capex FCF that should normalize. The uncomfortable comp fact: Amazon has the biggest cloud share but the slowest cloud growth (AWS +28% vs. Azure +40% vs. GCP +63%) — bulls pay up for the largest base, bears note the share-of-incremental is shifting to Azure/GCP. (Caveat: cross-vendor growth rates are not like-for-like — different fiscal calendars and segment definitions; the 10-Q's AWS +28% is the authoritative AMZN figure.)
Mostly ``; the pattern is what matters.
What the market actually reacts to for AMZN: (1) AWS growth rate — the single most price-sensitive variable; an AWS beat/miss moves the stock more than the giant retail line. (2) The FCF / capex trajectory — 2022 proved the market will de-rate hard if capex looks undisciplined. (3) Margin direction at the consolidated level. Retail revenue surprises barely move it. This name trades on AWS growth × capex discipline.
All exec data ``.
Capital-allocation history. Reinvest-and-compound is the religion — no dividend, minimal buybacks, near-100% of cash plowed into fulfillment + AWS + now AI silicon and data centers. Historically this created enormous value (AWS itself). The current test: $200B/yr into rapidly-depreciating AI hardware is the most aggressive capital deployment in the company's history. M&A is disciplined (One Medical $3.5B in 2023; immaterial in 2024-25), then suddenly large again with Globalstar (~$11.6B, April 2026). The equity stakes in Anthropic + OpenAI ($48.1B carrying value, Mar 2026) are a new capital-allocation category — strategic minority positions used to lock in compute demand.
Red flags (management): none structural. The earnings-quality optics of booking $15B+ of Anthropic markups through "other income" while FCF is ~$1B is a presentation concern, not a governance one. Bezos's 15M-share plan is the only insider item of note, and it's routine.
Acting as a forensic equity analyst. Every figure ``.
Auditor: Ernst & Young LLP, continuously since 1996; FY2025 critical audit matter = uncertain tax positions; goodwill impairment test (April 1, 2025) found no impairment, fair value "substantially exceeded" carrying value.
Regulatory findings (required). Read regulatory/regulatory-findings.md (fetched 2026-06-18, sources SEC EDGAR EFTS LR + AAER): 0 SEC Litigation Releases and 0 AAERs naming Amazon in 2021-2026. No accounting-fraud enforcement history. Non-SEC enforcement (web): the headline item is the FTC Prime "dark patterns" settlement — $2.5B (Sep 25, 2025), $1B penalty + $1.5B restitution to ~35M customers, settled 3 days into trial, no admission of liability. This is now closed and was booked as a $2.5B Q3 2025 charge. Separately, a $1.1B Italy tax/lawsuit resolution (Q4 2025). Item 3 / Note 7 (10-K + 10-Q): ordinary-course IP, tax, antitrust, labor, privacy litigation; the disclosed patent matter (Rensselaer/Alexa, $140-267M claimed) was won by Amazon (patent invalidated, affirmed by the Federal Circuit Feb 2026). No material unresolved monetary legal overhang disclosed. Verdict: clean accounting/enforcement record; the regulatory exposure is competition/consumer-protection (FTC/EU/DOJ), not fraud — and the biggest single item (FTC Prime) is resolved.
Built bottom-up from the latest actuals + guidance. All output with arithmetic; consensus reference points are. Note the GAAP-vs-operating distinction: because "other income" (Anthropic/OpenAI marks) is non-cash and unpredictable, I project operating-driven EPS and treat investment marks as a separate, un-forecastable swing factor.
Consensus anchors ``: FY2026 EPS ~$7.72-7.75 (one outlier $6.72), FY2027 ~$9.34; FY2026 revenue ~$795.6B; avg PT ~$313 (S&P Global, 67 analysts, "Strong Buy").
Revenue build:
Operating-margin build: FY2025 consolidated op margin 11.2% ($80.0B/$716.9B). Operating leverage from AWS + ads, partly offset by AI depreciation:
To operating EPS (net interest roughly neutral as rising debt service offsets the cash pile's income; normalized tax ~24%; shares ~10.9B flat-to-+1%/yr; excludes investment marks):
Why my operating EPS sits a touch below the street's headline: consensus GAAP EPS (~$7.75 FY26) likely bakes in some positive investment marks that I deliberately exclude as un-forecastable. On clean operating numbers I'm ~$6.90; including a "normal" Anthropic/OpenAI markup, GAAP could exceed $8. The gap is the earnings-quality issue — bulls should size on operating EPS, not the marked-up GAAP line.
Scenarios (FY2027 operating EPS):
Per the --watchlist rules, no Brier forecast logged (forecast.ts create skipped in breadth mode). If promoted to a thesis, the scoreable base call would be: "AMZN FY2026 operating (ex-investment-mark) diluted EPS ≥ $6.75."
Acting as an institutional-grade equity analyst — adversarial.
Bull case. Amazon is the best-positioned single company in the AI build-out, and you get the build-out optionality on top of two already-great businesses. (1) AWS is #1 and re-accelerating (+28% Q1 2026) with the AI run-rate >$15B compounding. (2) The silicon moat is real and widening — Trainium 3 at NVIDIA-parity TCO-50%-lower, "nearly fully subscribed," lets Amazon capture margin NVIDIA otherwise takes, and rent it to the labs. (3) Amazon is the neutral arms dealer — it now banks compute demand from both OpenAI (2GW) and Anthropic ($100B+/10yr, 5GW), i.e. it wins regardless of which lab wins. (4) Advertising ($68B, +22%) is a high-margin, under-appreciated third engine. (5) FCF is depressed by choice, not weakness — when capex intensity peaks, $140B+ of OCF drops toward the bottom line. (6) Optionality from Amazon Leo / Globalstar (satellite + Apple D2D) and robotics/autonomy is unpriced.
Bear case — risks that could permanently impair or de-rate. (1) The capex/depreciation trap. ~$200B 2026 capex on hardware that may be economically dead in ~3 years but is depreciated over 5; industry-wide AI depreciation could exceed combined profits. If demand disappoints, this is 2022's over-build at 4x the scale — and 2022 halved the stock. (2) AWS is losing share of incremental growth — Azure +40%, GCP +63% vs. AWS +28%; the #1 is the slowest-growing, and cloud compute is commoditizing up-stack. (3) The AI-lab stakes are a circular-financing risk — Amazon invests equity into its own customers (Anthropic/OpenAI), who then spend it on AWS; the $48.1B mark-to-model asset and the $15B+/qtr non-cash gains flatter GAAP and could reverse violently. Pre-mortem (18 months out, thesis broke): AWS growth rolled back to mid-teens as enterprises optimized AI spend and the labs' revenue didn't scale into the capacity; the 5-year depreciation was cut to 3-4 years; Anthropic/OpenAI marks were written down; FCF stayed near zero through 2027; and a ~27x forward multiple on a no-FCF, decelerating-cloud mega-cap compressed to ~18x — a 30%+ de-rating even with fine revenue. (4) Regulatory — FTC/EU/DOJ competition pressure is structural even with the Prime settlement closed.
Is the multiple too high? At ~27x forward, AMZN is the priciest cloud mega-cap, carried entirely by the AI-optionality narrative and a temporarily-suppressed FCF that the market is looking through. It is not cheap; it is a momentum-and-optionality multiple. Justified only if AWS sustains ≥22-25% and capex converts to FCF by 2027.
Contrarian view (what the market is refusing to see): the bull consensus treats the OpenAI + Anthropic deals as pure positives. The contrarian read is that Amazon is now partly financing its own demand — paying equity to lock in compute commitments — which is brilliant if AI inferencing scales and a hall-of-mirrors if it doesn't. The honest debate isn't "is AWS good" (it is); it's "is ~$200B/yr of 5-year-life capex, partly demand-financed with equity in mark-to-model private labs, a moat or a margin time-bomb?"
You are a skeptical short-seller dismantling the bull case.
Europe's only credible frontier lab — a sovereignty-and-infrastructure bet, not a model-quality bet; the moat is regulatory/political, the risk is that "good enough, in Europe, on-prem" gets out-priced by Chinese open weights before the €20B mark is earned.