Business Overview

Amazon.com, Inc. (NASDAQ: AMZN)

5 September 2026

Evidence base: Amazon SEC filings FY2021–FY2025 Forms 10-K, Forms 10-Q through Q2 2026 (filed 31 July 2026), earnings releases through Q2 2026, the 2026 DEF 14A and material 8-Ks of 27 February and 10 June 2026; supplemented by independent industry sources — competitors' own filings, the US Census Bureau, BLS, DOE/LBNL, PJM, the European Commission, the UK CMA and the US District Court docket in FTC v. Amazon.

This is a business and industry analysis. It is not a valuation and not a recommendation.

1. Executive Snapshot

2. What the Company Does

Amazon solves two unrelated problems for two unrelated customers and reports them in one income statement. For a consumer it removes the search-and-carry cost of buying a physical good: wide selection, delivered quickly, at a price checked against the market. For an enterprise it removes the need to own computing capacity — servers, storage and software services rented by the hour rather than bought and depreciated. The two businesses meet only on the balance sheet, where retail cash once funded the cloud build-out and cloud profit now underwrites both.

The AWS unit: one contracted, energised megawatt

AWS does not disclose megawatts, revenue per megawatt, or a customer count, so the unit has to be traced through what it does disclose. A customer signs a multi-year commitment; Amazon books the obligation as a performance obligation not yet recognised — approximately $496bn as of 30 June 2026, with a weighted-average remaining life of 6.4 years (Q2 2026 10-Q). Amazon then builds and energises the capacity, capitalises it into property and equipment ($446.0bn net at 30 June 2026, of which the AWS segment held $190.1bn at the end of FY2025), and depreciates it over five to six years. Revenue is recognised as the customer consumes.

The peer datapoints that exist put the price of the unit in the range of $20–25m per megawatt on Nebius's Q2 2026 contracts, and CoreWeave's disclosed 1.5GW of active power against $2.6bn of quarterly revenue implies roughly $4.6m of annualised revenue per megawatt installed (inferred). Neither is AWS. Both establish that the good being sold is power-shaped rather than software-shaped, which is why the industry's binding constraints are physical.

Cash leaves years before it returns. Amazon spent $128.3bn of net cash capital expenditure in FY2025 and $96.3bn in the first half of 2026 alone, against contracted revenue that converts over 6.4 years. Management's own framing is that “on average, it takes a little less than three years to break even on that investment” while servers have a useful life of at least five to six years (Q2 2026 call) — a claim of a two-to-three-year profitable tail per asset. That single sentence, which no external party can verify, is the load-bearing assumption behind a capital programme guided to approximately $220bn for 2026.

The retail unit: one paid unit, mostly someone else's

Sixty-one per cent of paid units worldwide are sold by third parties (Q2 2026 release), so the typical unit is not Amazon's inventory. On a third-party unit Amazon earns a referral commission, a fulfilment fee where the seller uses FBA, storage fees, and increasingly advertising. Amazon has never disclosed gross merchandise value, so the combined take must be inferred. From disclosed first-party sales, third-party services revenue and the unit mix, third-party seller services alone run at roughly 42–57% of implied third-party GMV; the FTC's 2023 complaint, drawing on Amazon's own produced documents, alleges Amazon takes “close to half of every dollar” from a typical FBA seller across selling fees, commissions, fulfilment and advertising. Two independent derivations landing in the same band is the strongest available evidence on a number Amazon does not publish.

The first-party-to-third-party shift is the most consequential change in the retail model, and it runs opposite to how it looks. A first-party sale books the full retail price as revenue and the full product cost as cost of sales; a third-party sale books only the fee. The mix shift therefore suppresses reported revenue growth while raising margin. It has now stopped: third-party units were 62% of the total in Q2 2025 and 61% in Q2 2026, the first sustained flat-to-down reading since 2021.

The third business, which has no segment

Advertising services revenue was $19.8bn in Q2 2026, up 26%, and $68.6bn in FY2025. It is reported inside North America and International, which is why those segments' margins cannot be read as retail margins: quarterly advertising revenue is 1.8 times the two retail segments' combined operating income of $10.8bn. Amazon does not disclose segment margin excluding advertising, so the profitability of the underlying retail operation is not observable from outside.

Where Amazon has entered and exited tells the same story. It closed physical stores and took roughly $1.1bn of impairments in FY2022; bought MGM for $6.1bn net in 2022 and One Medical for $3.5bn in 2023; and walked away from iRobot in January 2024. FY2025 acquisition consideration was “immaterial.” Capital is not going into new lines of business. It is going into compute, and — through a different door — into the equity of two artificial-intelligence laboratories.

3. Industry, Competitive Position and Moat

Amazon competes in two industries with opposite structures, so the moat question has to be asked twice.

Cloud: an oligopoly sitting below the profit pool

On the most recent reported figures, the fattest margins in the cloud value chain are upstream of the hyperscaler. SK hynix earned a 76% operating margin on memory in Q2 2026, Nvidia 66% on accelerators in the quarter to July 2026, and TSMC 60% on leading-edge foundry. Hyperscaler cloud segments earned 35–41% — AWS 39.4%, Microsoft's Intelligent Cloud 40.6%, Google Cloud 35.5%. Below them, colocation operators earned 25% (Equinix) and server ODMs 7% (Wiwynn), while the best-funded neocloud, CoreWeave, ran a negative operating margin on $2.6bn of quarterly revenue and paid a quarter of that revenue in interest.

The hyperscaler's margin is struck after depreciating assets bought from the stages above it, which is why the 2026 memory shortage transferred margin upstream directly and visibly: Amazon raised 2026 capex guidance from roughly $200bn to $220bn on memory prices, and Microsoft attributed approximately $25bn of its own guidance to component pricing. Upstream scarcity rents are cyclical; the hyperscaler's position is not, because it owns the customer relationship and the contracted book.

Structurally the market is a stable oligopoly whose leader is slipping. On Synergy Research's estimates the worldwide market ran at $143.4bn in Q2 2026, growing 43%, with Amazon at 28%, Microsoft at 20% and Google at 15%; the big three have held 61–63% combined across two years while AWS moved from 29% to 28% and Google from 13% to 15%. Two things are true of AWS at once: it grew 37% in Q2 2026, its fastest in eighteen quarters, and it is losing relative share, because Google Cloud grew 82% and Oracle's OCI 93% from smaller bases.

What actually protects a hyperscaler is not the service catalogue — neoclouds grew 112% to 454% with almost none of it — but three things money cannot buy inside a planning horizon. The first is energised power: the median wait from interconnection request to commercial operation now exceeds five years (LBNL), and PJM's 2027/28 capacity auction cleared 6,623MW short of its reliability requirement with data centres accounting for roughly 5,100 of 5,250MW of forecast peak-load growth. The second is advanced packaging: TSMC's chief executive said in Q2 2026 that “our packaging capacity is so tight that now it's limiting my customers' growth.” The third is contracted demand secured before capacity exists, which is precisely what the $496bn backlog is.

Testing Amazon's account of its own position

Two of Amazon's claims survive outside evidence and two are qualified. That AWS leads the market is supported by every available series. That AI demand is capacity-constrained rather than demand-constrained is corroborated independently on all three sides of the market — by the supplier (TSMC), by the competitor (Microsoft: “we expect to remain constrained at least through 2026”), and by the grid.

That switching costs are high is only partly supported. The European Commission's preliminary position of 25 June 2026, moving to designate AWS and Azure as gatekeepers for cloud under the Digital Markets Act, describes them as benefiting from “lock-in effects and high switching costs.” But the UK CMA, having examined actual customer behaviour after the 2024 free-egress programmes, concluded the uptake data was “inconclusive in terms of whether egress fees are or are not a barrier to switching,” and Article 29 of the EU Data Act prohibits all switching charges from 12 January 2027. The lock-in that survives is contractual and architectural — committed-spend discounts, re-architecture cost and a 6.4-year backlog — not tariff-based. And the claim that custom silicon confers a price-performance advantage rests on Amazon's own benchmarks; the strongest independent evidence for custom silicon generally is Anthropic's public choice of up to one million Google TPUs “due to their price-performance,” which validates the concept rather than Trainium.

Retail: a fragmented market that is concentrating where it matters

E-commerce was 17.1% of US retail in Q2 2026, so 83% of the market remains physical and fragmented — which is what makes Amazon's “intensely competitive” framing defensible on its own terms. Within the channel it actually competes in, however, concentration is rising: eMarketer estimates Amazon and Walmart together at 51.0% of US retail e-commerce in 2026 against 41.3% in 2019, the first reading above half.

Two natural experiments show which retail barriers bind. Walmart, the well-funded incumbent, is growing faster than Amazon's retail on every comparable line — US e-commerce +24%, global advertising +38%, membership fee revenue +17% — and has reproduced delivery speed essentially for free, reaching 30 minutes or less in 33 US markets using stores it already owned. What it has not reproduced is accumulated scale: its advertising business is roughly a tenth of Amazon's, and its consolidated operating margin of 5.0% sits below Amazon North America's 7.9%. Temu, the low-cost entrant, reproduced price rather than position, and the price advantage proved policy-dependent — the de minimis exemption ended for China and Hong Kong on 2 May 2025, globally on 29 August 2025, and was indefinitely suspended for all non-postal modes on 24 June 2026. Shein's US market share fell in 2025 for the first time since 2021.

The hardest single piece of evidence on Amazon's logistics position came from its own carrier. UPS disclosed that Amazon was 10.6% of its 2025 consolidated revenue and elected to cut that volume by more than half against 2024, completing the reduction by Q2 2026; its US average daily volume fell from 20.8m packages to 16.0m while revenue per piece rose 9.3%. Amazon absorbed the volume internally with no disclosed disruption. A peer-reviewed account of the US regionalisation programme (INFORMS Journal on Applied Analytics, 2026) quantifies why that was possible: in-region fulfilment from 62% to 76%, distance to customer down 15%, middle-mile touchpoints down 12%, and savings above $0.45 per unit — described as the first reduction in cost-to-serve per unit since 2018.

The industry evidence points to a clear answer on what kind of company wins in cloud: one with a cash-generative business that is not the cloud business, contracted long-dated demand, control of energised megawatts years ahead of need, and custom silicon as a margin-recapture option. Amazon is that kind of company on three counts without qualification. The fourth is now in question. Trailing free cash flow was negative $7.6bn at 30 June 2026 against positive $18.2bn a year earlier, so the retail engine no longer covers the cloud build on its own — the debt market does.

4. Growth Engine

Reported growth is essentially organic. Amazon made no material acquisition in FY2025 — consideration was “immaterial” — and the last two of consequence, MGM ($6.1bn net, 2022) and One Medical ($3.5bn, 2023), together added under 2% to a $717bn revenue base. What does need separating from the headline is currency and calendar.

Foreign exchange added $4.4bn to FY2025 net sales and $3.0bn to first-half 2026 net sales, though it was immaterial within Q2 2026 itself. Calendar matters more than it appears: Amazon's Q3 2026 guidance of 9–12% growth carries the note that “excluding the impact of Prime Day in both 2025 and 2026, third quarter 2026 year-over-year growth would be nearly 400 basis points higher,” with a further 80bp of FX drag assumed. A reader who takes the guided range as the underlying rate understates it by roughly five points.

Source: Q2 2026 10-Q. Acquisitions contributed nothing material to any line, so reported growth is organic growth; the only meaningful reconciling items are currency and the timing of Prime Day.

The drivers, ranked

• AWS capacity conversion — structural, management-driven. AWS grew 37% in Q2 2026 to $42.2bn and contributed $16.6bn of operating income, 61% of the company's total. Growth is set by how fast contracted capacity is energised rather than by selling effort: the backlog went from $195bn at 30 June 2025 to $364bn at 31 March 2026 to $496bn at 30 June 2026, while its weighted-average life lengthened from 4.0 to 6.4 years. Two disclosed expansions explain much of the step — an OpenAI commitment increased by $100bn over eight years in Q1 2026, and an Anthropic collaboration increased by more than $100bn over ten years in Q2 2026.
• Advertising — structural. $19.8bn in Q2 2026, up 26%, growing faster than Alphabet's advertising business (+14%) from roughly a quarter of its base. Incremental margin is close to complete because the inventory is a page Amazon was rendering anyway and the attribution data is a by-product of the order system it already operates. Growth is now coming from new surfaces — Prime Video advertising, live sport, off-site demand-side inventory — rather than more slots on the search page, which is itself the clearest signal that on-site ad load is near its practical ceiling.
• Third-party seller services — structural, but the mix lever has stopped. Up 16% in Q2 2026. The rate can rise even when the mix does not, because take-rate expansion across the sector is coming from advertising rather than headline commissions. But third-party units have held at 61–62% of the total since 2024 and ticked down over the last year, so the shift that drove a decade of margin improvement is substantially complete.
• Retail unit volume — structural with a cyclical overlay. Paid units grew 17% in Q2 2026. The underlying channel shift continues at roughly 0.55 percentage points of US retail share a year, about 55–60% of its pre-COVID pace, and the COVID spike fully round-tripped before the trend resumed.
• Price, mix and currency — cyclical and temporary. Currency, Prime Day timing and a $640m tariff refund under the International Emergency Economic Powers Act, booked to North America cost of sales in Q2 2026, all flatter the current period. Amazon stated the refund “represents the significant majority of refunds we expect to receive.”

One lever has not been pulled. Subscription services grew 12% in Q2 2026 with the US Prime price unchanged at $139 since February 2022. Amazon has raised realised revenue per member instead through the Prime Video advertising tier, introduced in January 2024 and repriced upward by $2 a month in March 2026. The headline price is an untested lever whose elasticity is unknown, because Amazon has disclosed no member count since 2021 and has never disclosed a renewal rate.

5. Margin, Cash and Capital Allocation

Consolidated operating margin was 11.2% in FY2025 and 13.7% in Q2 2026, but that number blends three businesses with different physics. AWS converts revenue to profit through utilisation: its costs are depreciation, power and people, all largely fixed once capacity is energised, so incremental revenue on installed capacity falls through at a high rate and idle capacity is expensive. That is why AWS margin swings so far — 27.1% in FY2023, 37.0% in FY2024, 35.4% in FY2025, 39.4% in Q2 2026.

The retail segments convert through density instead. Fulfilment and shipping costs scale with units and cube rather than with dollars, so profitability depends on packages per route and distance per package, not on revenue. Worldwide shipping costs rose 19% in Q2 2026 against 17% unit growth — cost per unit up roughly 1.7% — and management attributed the gap to fuel and line-haul rates, stating that excluding those, shipping costs grew more slowly than units. The underlying network is still deflating per unit; input prices are currently offsetting it. The comparison that matters is the third-party cost floor: UPS took 9.3% and FedEx 10% on revenue per piece over the same year.

The financial spine

Sources: FY2023 and FY2025 Forms 10-K. FY2024 and FY2025 free cash flow are as reconciled by Amazon; FY2022 and FY2023 are computed on the same definition because the FY2025 10-K reconciles only two years. Comparability warning: these four years are not on a constant depreciation basis, and FY2025 operating income absorbs $2.5bn for the FTC settlement and $2.7bn of severance.

The depreciation point deserves stating plainly, because it is the single largest threat to reading this table as a trend. Amazon lengthened server lives from four to five years effective January 2022, telling investors to expect roughly $3.1bn of benefit to 2022 operating income; lengthened them again from five to six years effective January 2024, with the same $3.1bn estimate for that year; then shortened a subset from six back to five years effective January 2025 “due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning,” costing $1.4bn of additional depreciation and $1.0bn of net income in FY2025, primarily in AWS. Heavy equipment moved from ten years to ten-to-thirteen years effective January 2025 with no dollar effect quantified. A margin line drawn from FY2022 to FY2025 crosses three different depreciation policies.

Cash conversion has inverted

Operating cash flow rose 20% in FY2025 to $139.5bn and 33% in the twelve months to June 2026 to $161.4bn. Free cash flow on Amazon's own definition fell from $38.2bn in FY2024 to $11.2bn in FY2025 and to negative $7.6bn on a trailing basis at 30 June 2026. The entire gap is capital expenditure: $128.3bn net in FY2025, $169.0bn trailing, and guided to approximately $220bn for calendar 2026.

Reported operating cash flow is also being helped by payables. Accounts payable rose to $121.9bn at end-2025 and $147.4bn at June 2026, and $27.0bn of the 2025 balance was property and equipment acquired but not yet paid for, against $16.8bn a year earlier. Days payable computed on the headline balance moved from 106 to 125 days; excluding unpaid capital expenditure, from 87 to 97 days. The extension is real but roughly half the headline suggests.

Where the cash went

Ranked, the answer is unambiguous and has become more so. Across FY2019–FY2023 the order was capital expenditure ($209.6bn net), finance-lease principal ($43.8bn), acquisitions and investments ($20.9bn), debt repayment ($9.2bn), buybacks ($6.0bn, in 2022 only) and dividends (none, ever). Since then the concentration has increased: FY2025 net capex alone was $128.3bn; there were no repurchases in 2023, 2024, 2025 or the first half of 2026, with $6.1bn still unused under a 2022 authorisation; and no dividend has been declared. The diluted share count consequently rises about 1% a year on stock compensation, from 10,492m in FY2023 to 10,903m in Q2 2026.

The funding source changed decisively in 2026. Long-term debt went from $65.6bn at 31 December 2025 to $128.9bn at 30 June 2026 across US dollar, euro, Swiss franc and Canadian dollar issues, with a further $25.0bn of notes issued in July 2026 and a $17.5bn undrawn delayed-draw term loan signed with Citibank on 8 June 2026 whose commitments expire on 30 September 2026. Interest expense more than doubled year on year to $1.3bn in the quarter. Amazon states that it is “not subject to any financial covenants under the Notes” and that it expects “to undertake additional financing activities in 2026.”

Two commitments that post-date the reported figures

On 27 February 2026 an Amazon subsidiary agreed to purchase $35.0bn of OpenAI Series C preferred stock, guaranteed by Amazon, drawable at Amazon's discretion but mandatory on OpenAI meeting specified milestones or completing a public listing, with an outside date of 31 December 2028. The 8-K states this is “separate from and in addition to” a $15.0bn purchase obligated on 31 March 2026, and that affiliates simultaneously entered an AWS cloud-services arrangement and a joint collaboration agreement, neither of which carries a disclosed value. Amazon had invested $28.7bn of the $50bn by 30 June 2026 and the remaining $21.3bn subsequently. Separately it invested $10.0bn in Anthropic preferred stock in Q2 2026 and extended a facility of up to $20.0bn that becomes drawable as compute-capacity milestones are met, of which $15.0bn remained available. Amazon is now supplier, equity holder and lender to the same two counterparties.

The compensation structure is consistent with this capital behaviour. Every named executive earns a $365,000 base salary, there is no annual bonus, no performance-conditioned equity and no severance; pay is periodic restricted stock vesting over five or more years on a back-end-weighted schedule, and no equity was granted to any named executive in 2025. The proxy states directly that “the Company does not use any financial performance measures to link executive compensation to company performance.” Jeff Bezos holds 8.8% of a single-class share structure and has never taken stock compensation. Whatever else this arrangement produces, it does not produce pressure to defend a quarterly margin.

6. Cyclicality, Constraints and What to Monitor

Both engines are at or near their own cycle highs, which is the necessary context for every margin in this document. AWS revenue growth troughed at 12% in Q2 and Q3 2023 and reached 37% in Q2 2026, its fastest in eighteen quarters. AWS operating margin troughed at 27.1% in FY2023 and printed 39.4% in Q2 2026, a record. North America's segment margin was negative 0.9% in FY2022 and 7.9% in Q2 2026, also a record. A record margin at a trough and a record margin at a peak are opposite facts, and these are the second kind.

The cloud precedent is specific and worth carrying. Between Q1 2022 and Q2 2023 AWS growth fell from 37% to 12% — two-thirds of the growth rate in five quarters — while sector producer prices, on the BLS index for data processing and hosting, actually rose. It was volume and configuration rather than price: AWS operating income fell only 6% in the worst quarter and was up 29% a quarter later, because Amazon pulled the capex brake, with trailing purchases of property and equipment down 15% year on year by September 2023.

That lever is materially weaker in 2026. Total contractual commitments were $650.0bn at 30 June 2026, including $137.2bn of leases not yet commenced and $130.1bn of unconditional purchase obligations, and power contracts run for years — Amazon holds energy contracts covering roughly 200 million megawatt-hours with a weighted-average remaining duration near sixteen years. A demand pause arriving while contracted depreciation is still ramping therefore compresses margin on both blades at once: revenue growth falls into a cost base that cannot be withdrawn at the same speed. That configuration is exactly what the 2022–23 episode does not cover, because that capital expenditure was discretionary and short-dated.

The retail downside runs through mix rather than volume. Consumers rotate from discretionary to consumable goods rather than stopping: Target's comparable sales still grew 3.3% in Q1 2022 while its gross margin fell 430 basis points. Consumables are cheaper, heavier and bulkier per dollar and monetise less advertising, so units rise while merchandise value does not — the worst combination for a network whose costs scale with units and cube. North America went from +2.6% to −0.9% segment margin in a single year on decelerating, not declining, volume. Advertising decelerates but has not historically declined: Amazon's advertising growth troughed at 18% in Q2 2022 while Meta's advertising revenue fell 1% for that full year.

Exposure is concentrated in two places. Geographically, the United States was $489.7bn of $716.9bn of FY2025 net sales, or 68%. By segment, AWS was 18% of FY2025 revenue but 57% of segment operating income — 21% and 61% respectively in Q2 2026. Amazon's earnings are far more exposed to enterprise IT and AI capital budgets than its revenue mix implies, and increasingly so each quarter.

Durable against borrowed

Leading indicators, and where each is published

7. Risks, Unknowns and Questions for Deeper Work

Cyclicality is covered in Section 6. What follows is what cyclicality does not capture.

• The backlog and the balance sheet key off the same two counterparties. AWS's contracted obligations grew roughly $301bn in twelve months, and the two publicly identified expansions — OpenAI at $100bn over eight years and Anthropic at more than $100bn over ten — account for the majority of it. Amazon is simultaneously those customers' equity holder, with $122.3bn of private-company investments at 30 June 2026, and Anthropic's lender of up to $20bn drawable as compute milestones are met. If either laboratory's funding fails, Amazon loses contracted revenue, writes down equity and is left holding purpose-built capacity, in the same quarter. No counterparty concentration inside the backlog is disclosed.
• Depreciation converges on capital expenditure, and the mechanism is arithmetic rather than competitive. Trailing capex of $169.0bn already exceeds AWS's entire FY2025 revenue of $128.7bn, and with five-to-six-year lives the annual depreciation charge tends toward the annual capex figure once the build is in service. AWS depreciation rose 61% in FY2025 to $21.5bn. If revenue growth decelerates while the asset base keeps ramping, margin compresses with no pricing action by anyone. Amazon has already shown the sign of it: AWS margin fell from 37.0% to 35.4% in FY2025 while revenue grew 20%.
• Net income has stopped describing the business. Other income was $69.1bn in the first half of 2026, of which $62.8bn was unrealised upward revaluation of privately held Anthropic stock priced off other investors' funding rounds, followed by $15.9bn of discrete tax expense. Reported first-half net income of $92.9bn sits on operating income of $51.3bn. These are Level 3 fair values and Amazon states that “market sensitivities are not practicable” for them. Any analysis anchored on reported earnings per share is now measuring something other than the operating business.
• The FTC case attacks the mechanism rather than the conduct. Trial is set for 29 March 2027 in the Western District of Washington, with Sherman Act Section 2 counts and the Project Nessie claim having survived dismissal in September 2024. The two practices at issue — Buy Box eligibility and the conditioning of Prime on Fulfilment by Amazon — are the link that converts marketplace share into fulfilment volume, and therefore into route density, and therefore into cost per unit. A remedy touching them would work backwards through the cost structure rather than simply capping a fee.
• Financing risk here is structural, not covenant-driven. Debt doubled in six months and there are no financial covenants, so there is no tripwire — and equally no external constraint. The exposure is refinancing $132bn of notes at a weighted-average remaining life of 14.2 years into an unknown rate environment, against assets whose economics rest on a two-to-three-year break-even that no external party can verify.
• Seller economics are the quiet variable. Third-party unit mix has stopped rising for the first time since 2021 while the inferred take rate sits near half of seller revenue. If the mix reverses, Amazon books gross merchandise revenue and full product cost in place of a fee — raising reported revenue and lowering margin, precisely inverting the last decade's trend. No independent data on Amazon seller profitability exists.

What the sources could not answer

• Gross merchandise value, total or third-party. Amazon has never disclosed it, so the take rate must be inferred; that inference drives any view of marketplace economics.
• Prime membership count and renewal rate. The last company figure was 200 million, in the 2020 shareholder letter published in April 2021. Amazon now says only “double-digit growth” with no base.
• Any AI-specific revenue, the training-versus-inference split, or their relative margins. No hyperscaler discloses these, and the durability of the backlog turns on them.
• Segment operating margin excluding advertising — so the profitability of the underlying retail operation is unobservable.
• Customer concentration. The FY2025 10-K contains no customer-concentration disclosure at all: no customer is named at 10% of revenue, and there is no statement that none exists.
• The maintenance-versus-growth split of capital expenditure, and the dollar effect of the January 2025 heavy-equipment life extension.
• Whether AWS's “over $25bn” chips run-rate and its “over $25bn” AI run-rate, both cited on the Q2 2026 call, describe the same revenue pool.
• Full-year 2026 capital expenditure does not appear in any filing — the approximately $220bn figure exists only in management's call commentary. Nor do transcripts, shareholder letters or investor-day materials exist in the source folder, so management's own framing here rests on releases and MD&A.

Resolving the first three would change the analysis most. Without a member count the strength of the consumer franchise cannot be sized; without an AI revenue and margin disclosure the $496bn backlog cannot be valued as either an annuity or a project book; and without gross merchandise value the marketplace's pricing power can only be triangulated from a litigant's allegation and an inference.

8. Investor Takeaways

• What this business really is: a capital-intensive compute utility earning 39% margins, funded by a normal-margin retailer that happens to own a near-full-margin advertising business. AWS is 21% of revenue and 61% of segment profit.
• The core economic engine is a contracted, energised megawatt of compute. Amazon signs the demand first — $496bn at a 6.4-year weighted life — then builds, energises and depreciates the asset over five to six years.
• The main growth lever is the speed at which contracted capacity is energised, not demand generation; advertising is the second lever, and it now grows through new surfaces rather than higher ad load.
• What could break the story is a demand pause arriving while committed depreciation is still ramping. In 2022–23 AWS lost two-thirds of its growth rate in five quarters and defended margin by cutting capex; with $137.2bn of leases not yet commenced, that lever is weaker now.
• Monitor the AWS backlog and its weighted-average life, AWS margin against the depreciation ramp, free cash flow against a roughly $220bn capex year, and the March 2027 FTC trial.