Amazon.com, Inc.

01 — Executive summaryOperating income of $27.5 billion. Net income of $62.6 billion. Both are correct.

Amazon.com reported net sales of $200.6 billion for the quarter ended 30 June 2026 — the first $200 billion quarter in the company's history, and a 20% increase over the prior year. Operating income was $27.5 billion, up 43%. Amazon Web Services grew 37%, its fastest rate in eighteen quarters, on a base that has already passed a $169 billion annual run-rate. Advertising grew 26%. Operating margin, at 13.7%, was the highest the company has ever reported. Management raised its 2026 capital expenditure guidance from $200 billion to $220 billion and disclosed $496 billion of contracted AWS backlog that has not yet been billed.

None of that is in dispute, and none of it is the subject of this report. The question we set out to answer is narrower: what is Amazon's earnings power, once you remove the parts of the income statement that are not the operating business — and what does the current price already assume about that number?

Our answer, compressed into a sentence: Amazon's operating businesses are performing better than at any point in the last decade, and the equity is priced as though that performance will continue without interruption through a capital cycle that is consuming every dollar the company generates. The reported earnings that make the stock look inexpensive are, increasingly, not operating earnings at all.

The single most important number in this report is not revenue or margin. It is this: over the trailing twelve months, Amazon reported net income of $135.3 billion against operating income of $93.7 billion. Net income exceeded operating income by $41.6 billion — 44% of the reported bottom line — for a company that spent $173 billion on property and equipment over the same period. In the June quarter alone the gap was $35.1 billion, because $53.4 billion of pre-tax profit arrived from revaluing an equity stake in a private company rather than from selling anything to anyone.

Six findings follow from the primary record. Each is developed in full later in this report; each is stated here with the number that supports it.

1. AWS has genuinely re-accelerated, and the margin proves it is not being bought

AWS grew 17% in the first quarter of 2025 and 37% in the second quarter of 2026. Growth of that shape, on a business of this size, is rare. What makes it more than a demand story is what happened to the margin at the same time: AWS operating margin bottomed at 32.9% in Q2 2025 and reached 39.4% in Q2 2026, an expansion of 650 basis points. A company buying growth through price or through unprofitable capacity additions cannot do that. Management attributes the expansion principally to a higher mix of in-house silicon — Trainium for training and inference, Graviton for general compute — which carries a lower unit cost than the merchant alternative. The margin line is the evidence, and it is the strongest single argument in the bull case.

2. The capital cycle has turned free cash flow negative, and it is not finished

Amazon generated $161.4 billion of operating cash flow over the twelve months to 30 June 2026 and spent $169.0 billion on property and equipment, net of disposals and incentives. Free cash flow was −$7.6 billion, against an inflow of $18.2 billion a year earlier. Capital expenditure has grown from $48 billion in 2023 to $77.7 billion in 2024 to $128.3 billion in 2025, and is guided to $220 billion in 2026 — a 4.6× increase in four years, against revenue growth of 43% over the same period.

This is a deliberate choice, and the company has been explicit about it. Andy Jassy has described AI infrastructure as a once-in-a-generation opportunity and has said the company will not have enough capacity in 2026 or 2027 to meet demand it can already see. The $496 billion backlog, which grew $132 billion in a single quarter, is the supporting evidence. But a choice is still a choice, and the consequence is that Amazon now funds growth from a combination of operating cash flow and borrowed money rather than from operating cash flow alone.

3. Reported earnings have decoupled from the operating business

Amazon holds a stake in Anthropic that it has built for $13 billion since 2023 and that it carried on its balance sheet at $190.4 billion as of 30 June 2026 — $97.9 billion of convertible notes and $92.5 billion of non-voting preferred stock. That carrying value is 34.5% of Amazon's $551.6 billion of book equity. Because of the accounting treatment, every time Anthropic raises capital at a higher valuation, Amazon books a gain above the operating line.

In the first half of 2026 those gains totalled approximately $69 billion pre-tax, of which $53.4 billion landed in the June quarter alone. Anthropic's valuation went from $61.5 billion in March 2025 to $183 billion in September 2025 to $380 billion in early 2026 to $965 billion at its Series H in May 2026. A prospective initial public offering at $2 trillion has been widely reported for as early as October 2026, which would imply a value for Amazon's holding of roughly $394 billion on an estimated 19.7% ownership — a figure Amazon has never disclosed.

4. On operating earnings, the shares are not cheap

At $248.42, Amazon trades at 20.0× trailing GAAP earnings and roughly 19× our estimate of FY2026 GAAP earnings. Those are undemanding multiples for a business growing revenue at 14% and operating income at 33%. They are also, in our view, the wrong denominator. Strip out non-operating investment gains and Amazon's trailing earnings are approximately $82 billion, not $135 billion. On that basis the shares trade at roughly 32× current-year operating earnings and 27× our FY2027 base case.

This is the analytical crux of the report. The market is not paying 20× for Amazon's retail and cloud businesses; it is paying 20× for those businesses plus a mark-to-market gain on a private company. If the mark is right, the multiple is right. If the mark is early, the multiple is 60% higher than it appears.

5. The returns question is now genuinely open

Amazon's return on invested capital, on a trailing basis, is approximately 11.9% against a weighted average cost of capital of approximately 11.2%. Those numbers are close enough that the sign of the difference is not robust to reasonable changes in methodology — which is precisely the point. Amazon is deploying capital at a rate never before seen in the company's history, into an asset class whose economic life is contested, at a return that is currently indistinguishable from its cost of capital. If that spread widens, the bull case is correct and the current price is cheap. If it narrows, the capital deployed in 2026 and 2027 will have destroyed value, and the market will not have to wait for the cash flows to know it.

We note one further asymmetry. The trailing return on invested capital is flattered by the Anthropic gains, which are pure profit with no invested capital attached. Removing them lowers the computed return by roughly two percentage points, which places it below the cost of capital. The operating business is not currently earning its cost of capital; the portfolio is.

6. Competition has stopped being a share story and become a rate story

AWS remains the largest cloud provider and is growing at 37%. It is also losing share. Synergy Research estimates that AWS's share of global cloud infrastructure spend fell roughly two percentage points year over year in the June quarter, while Google Cloud gained a similar amount; Google Cloud grew 82% and Microsoft's Azure grew 43% over comparable periods. The market itself grew 43% to $143 billion in the quarter, the fastest rate in eight years.

A market growing 43% can accommodate a leader growing 37% and losing share while still adding enormous absolute revenue. That is the situation today. The question for the next three years is whether the absolute additions continue to grow fast enough to absorb $220 billion a year of capital, at a moment when the two closest competitors are both accelerating faster than the leader. Nothing in Amazon's guidance suggests management is worried. That is not the same as evidence that it should not be.

What would change our mind

We would move to a positive rating on any of three developments: evidence over two consecutive quarters that incremental capital is earning a return above the cost of capital, most cleanly demonstrated by AWS operating margin holding above 39% while capital intensity continues to rise; a disclosure of the split between organic revenue and revenue from counterparties in which Amazon holds an equity interest; or a resolution of the Anthropic position into a listed security at a valuation that the market capitalises rather than discounts.

We would move to a negative rating on any of three others: a quarter in which AWS growth falls below 25% while capital expenditure guidance is maintained; a credit-market event that raises Amazon's cost of debt materially, given that the company has now become a regular issuer; or evidence that the $496 billion backlog contains a materially higher proportion of cancellable or conditional commitments than the headline implies.

02 — Company & business modelA logistics network that learned to sell cloud capacity

Amazon was incorporated in 1994 as an online bookseller and spent its first decade building the retail operation that made it famous and the fulfillment infrastructure that made it viable. The decision that created the modern company was made in 2003 and executed in 2006, when Amazon began renting out the internal computing infrastructure it had built for its own retail business. Amazon Web Services was not a strategy in the conventional sense; it was a by-product of solving an internal problem at scale and recognising that the solution had external demand.

Two decades later the by-product is the profit centre. AWS generated $128.7 billion of revenue in 2025 — 18% of the consolidated total — and $45.6 billion of operating income, which is 57% of the company's total. North America, the retail segment that most people picture when they hear the word Amazon, generated $426.3 billion of revenue and $29.6 billion of operating income. International generated $161.9 billion of revenue and $4.8 billion of operating income. The company is, financially, a cloud provider with a very large retail operation attached.

What the reporting structure does and does not show

Amazon reports three operating segments: North America, International, and AWS. Each contains a mixture of businesses with different economics, and the mix is disclosed only partially. North America contains first-party online stores, third-party marketplace commissions and fulfillment fees, advertising, subscription revenue, physical stores, and the healthcare operations that arrived with One Medical. International contains the same set, in different proportions and with different competitive dynamics. AWS contains compute, storage, database, machine learning, and a rapidly growing custom silicon business that is not broken out.

The company does disclose revenue by product and service line in a footnote, which is more than most large-cap retailers provide, and which allows a reader to reconstruct the revenue mix. It does not disclose operating income by product line, so the profitability of advertising — widely understood to be the highest-margin business Amazon owns — cannot be established from the filings. Any analyst who states an advertising operating margin is estimating. We say so when we do it.

How revenue actually reaches the income statement

Amazon's revenue recognition is more varied than most industrial companies, and the differences matter for anyone comparing growth rates. First-party online store revenue is recognised gross, when a product is delivered. Third-party seller services are recognised as commission and fulfillment fees, which means the same underlying transaction appears in Amazon's income statement at a fraction of the merchandise value. Advertising is recognised on delivery of impressions or clicks. AWS is recognised on consumption, with a growing share of multi-year committed contracts recognised over the commitment period.

This structure explains why Amazon's 20% revenue growth in Q2 2026 is not comparable to a 20% growth rate at a conventional retailer. The fastest-growing lines — AWS at 37% and advertising at 26% — are the ones recognised at or near gross value with the least variable cost attached. Online stores, the largest line at $269.3 billion in 2025, grew 9%. The mix is shifting toward the higher-margin, faster-growing lines, which is why consolidated operating margin has expanded even as the largest revenue category grows at a single-digit rate.

Capital intensity is the defining characteristic, and it has changed

For most of its history Amazon was capital-intensive in a way that made it a difficult company to own through a downturn. It built fulfillment centres, and the depreciation on those centres arrived before the revenue they supported. Between 2019 and 2022 that pattern compressed free cash flow and produced a multi-year stretch in which the shares went nowhere.

The capital intensity that matters now is different in kind. Fulfillment infrastructure is long-lived, geographically distributed, and depreciated over roughly ten to fifteen years. AI data centre capacity is depreciated over five to six years and, on the evidence of the last three years of accelerator releases, may be economically obsolete well before the depreciation schedule says so. Amazon's 2026 capital expenditure guidance of $220 billion is directed disproportionately at the second category. The company has not disclosed the split between fulfillment capex and technology infrastructure capex in a way that would let an outside analyst size the two, which is the single largest disclosure gap in the equity.

We flag this not as a criticism of the strategy but as a fact about the risk. A company that spends $220 billion in one year on assets with a contested useful life has made a bet on the durability of demand that is larger, in absolute dollars, than any bet any company has ever made on anything.

03 — The financial recordFour years, and the moment the arithmetic changed

Amazon's revenue grew from $574.8 billion in 2023 to $716.9 billion in 2025, a 25% increase over two years. Operating income grew from $36.9 billion to $80.0 billion over the same period, a 117% increase. That divergence — operating income growing roughly five times as fast as revenue — is the most important fact about Amazon's financial history over the last three years, and it is the product of three things happening at once: AWS scaling into its fixed cost base, advertising growing faster than anything else, and retail margins improving from a low base.

The quarterly record shows the inflection more clearly than the annual one, because it separates the period in which operating income was the story from the period in which it stopped being the story.

Q3 2025 and Q3 2026E figures are derived from full-year totals or company guidance; Q3 2026 revenue and operating income are the midpoints of guidance of $197.0–202.0 billion and $22.5–26.5 billion. Q1 2026 North America and International net sales are as reported; Q3 2025 segment figures are derived by subtraction. Other income for Q4 2025 and H2 2025 is a Farstar estimate derived from reported pre-tax and operating income. The ex-investment-gains EPS line is a Farstar calculation that removes pre-tax gains on equity securities at the reported effective tax rate; it is an approximation of operating earnings power, not a company disclosure.

Three observations from this table deserve emphasis.

Operating margin has broken out of its range. For six consecutive quarters from Q1 2025 through Q4 2025, Amazon's operating margin sat between 9.7% and 11.8%. In the first half of 2026 it printed 13.1% and 13.7%. The company has never reported a 13.7% operating margin before. The improvement is not a cost story — operating expenses have grown broadly in line with revenue — it is a mix story, and the mix is shifting in Amazon's favour.

Net income has stopped tracking operating income. In every quarter from Q1 2025 through Q4 2025 the two lines were within $4 billion of each other. In Q1 2026 the gap was $6.4 billion. In Q2 2026 it was $35.1 billion. The cause is a single line item: other income, which was $2.8 billion in Q1 2025 and $53.4 billion in Q2 2026. That is not a business getting more profitable; it is a portfolio being marked up.

The GAAP-to-operating gap is now large enough to change the valuation. On trailing GAAP earnings of $12.44 per share, Amazon trades at 20.0×. On our estimate of trailing operating earnings, it trades at approximately 32×. The difference between those two numbers is the entire investment case, and it is a difference created by an accounting treatment rather than by anything that happened in a warehouse or a data centre.

We will return to each of these threads. Before turning to the segments, one framing observation is worth stating plainly, because it governs everything that follows. Amazon's equity has, over the last eighteen months, acquired the characteristics of a levered holding company. It owns a set of operating businesses that generate roughly $94 billion of operating income and consume roughly $169 billion of capital expenditure. It owns a portfolio of equity stakes, of which one is now worth more than a third of its book equity. And it has borrowed $63 billion in six months to fund the difference. Each of those three things is defensible. Together, they mean that the historical relationship between Amazon's reported earnings and its share price has broken down, and that anyone valuing the company on reported earnings is valuing something that no longer exists.

04 — Segment deep diveAWS: the only part of Amazon that sets the price

AWS revenue was $42.2 billion in Q2 2026, up 37% year over year and 12% sequentially. The annualised run-rate implied by that quarter is $169 billion, which would place AWS on its own at roughly the 24th position in the Fortune 500. Growth has accelerated in each of the last six quarters: 17%, 17%, 20%, 24%, 28%, 37%. Management described the June quarter as the fastest growth in eighteen quarters.

The re-acceleration is real and it is not a base effect. AWS revenue passed $100 billion on an annualised basis in 2023, and a business of that size does not normally find another gear. What changed is the composition of demand. For a decade AWS growth was driven by enterprises migrating existing workloads from owned data centres, a process with a natural ceiling set by how much infrastructure existed to migrate. The marginal buyer today is different: a company that did not exist, or did not consume meaningful compute, five years ago, buying capacity for model training and inference at a scale that did not have a precedent.

The backlog number, and how to read it

Amazon disclosed $496 billion of remaining performance obligations at the end of Q2 2026, against $364 billion at the end of Q1 — an increase of $132 billion in a single quarter. The company described the year-over-year growth rate as triple-digit. The figure is approximately three times the current AWS annualised revenue run-rate, which means that on the disclosed number, nearly three years of current revenue has already been contracted.

That is a genuinely impressive disclosure and it is the strongest quantitative evidence for the bull case. It should nevertheless be handled with three caveats, each of which is a matter of public record rather than speculation.

First, remaining performance obligations are not a backlog in the industrial sense. They are contracted but unbilled revenue, and the definition of "contracted" varies by customer and by product. Some portion is consumption-based commitments with minimum floors; some is take-or-pay; some is capacity reservation with renegotiation provisions. Amazon does not disclose the mix, and it does not disclose how much of the RPO balance has historically converted into revenue in each period. Without that, the number is a measure of intent at least as much as of obligation.

Second, the customers behind it are concentrated and partially self-referential. The two largest publicly disclosed AWS AI customers are Anthropic and OpenAI, both of which have made multi-year, multi-gigawatt capacity commitments. Amazon is an investor in both. Anthropic has committed to more than $100 billion of AWS spend over ten years and up to five gigawatts of capacity; in return Amazon has invested $13 billion to date and left open a further $20 billion contingent on commercial milestones. This is a legitimate commercial structure and it is disclosed. It is also a structure in which Amazon's equity investment and Amazon's revenue backlog are, to an extent that cannot be quantified from the filings, two views of the same transaction.

Third, RPO growth is an industry-wide phenomenon, not an AWS-specific one. Alphabet's Google Cloud carried $467.6 billion of RPO as of 31 March 2026, a comparable figure on a smaller revenue base. Microsoft does not disclose a directly comparable number but has referenced over $100 billion of annual Azure revenue. The fact that all three are accumulating contractual commitments at once is consistent with a genuine structural shift in enterprise computing, and also consistent with a competitive dynamic in which every provider is discounting forward to lock in capacity. Both readings are available from the same data.

Margin: the number that settles the argument

AWS operating margin was 39.4% in Q2 2026, against 32.9% a year earlier. That 650 basis point expansion, achieved while revenue grew 37% and while the segment was adding capacity at the fastest rate in its history, is the most important financial fact in the AWS franchise. It is difficult to buy growth at that scale without paying for it in margin; Amazon did not.

The stated driver is mix. Amazon's custom silicon — Trainium for AI training and inference, Graviton for general-purpose compute, Nitro for virtualisation — carries a lower unit cost than the merchant alternatives, and the share of AWS capacity running on it has risen materially. Management has said that the chips business and the AI services business each crossed a $25 billion annualised run-rate in Q2 2026, both growing at triple-digit rates. Graviton5, released to general availability in the quarter, is described as delivering up to 30–40% better price-performance than comparable instances and is used by 98% of the top 1,000 EC2 customers.

The economic logic is straightforward. A merchant accelerator is bought at a price that includes the vendor's gross margin, which for the market leader is approximately 75%. A custom accelerator is bought at a price that includes TSMC's margin and the design cost amortised over the volume Amazon can underwrite. If Amazon's internal demand plus its committed customer demand is large enough to amortise the design, the saving flows to AWS's margin. The margin data suggests it is.

The constraint is software. Trainium delivers its cost advantage through Amazon's proprietary Neuron software development kit rather than through the CUDA ecosystem that most AI engineers already know. Porting an existing training pipeline to Neuron is real work, and the size of that work is the reason the merchant alternative has not been displaced more quickly. Amazon's answer is Bedrock, which now serves OpenAI's GPT-5.6, Anthropic's Claude Opus 5, Google DeepMind's Gemma 4 and others behind a single API. If the model layer is commoditised behind an interface, the silicon underneath matters less to the customer — which is precisely the outcome Amazon needs.

Geographic and structural considerations

Amazon does not disclose AWS revenue by geography or by industry vertical. What can be observed from the outside is that the growth is concentrated in North America, that the government business is expanding materially — Amazon announced a $50 billion commitment to expand AI infrastructure for US federal agencies, adding 1.3 gigawatts of capacity, alongside expanded government access to Claude through Bedrock — and that the Middle East, Japan, and parts of Europe are contributing sovereign AI demand that did not exist three years ago.

Sovereign and government demand is structurally attractive because it is less correlated with commercial capital cycles. It is also politically determined, lumpy, and frequently subject to procurement timelines that slip. We do not give it credit in our base case beyond what is already contracted.

What we would need to see to be more constructive on AWS specifically

Three things. First, disclosure of the custom silicon revenue line as a reported segment metric rather than a run-rate statement on an earnings call — the $25 billion figure is management commentary, not an audited number, and the difference matters for anyone modelling the margin trajectory. Second, evidence that the non-Anthropic, non-OpenAI share of AWS growth is rising; the current disclosure does not permit that calculation. Third, a quarter in which AWS margin holds above 39% while capital intensity continues to increase, which would demonstrate that the margin expansion is structural rather than a function of the current point in the depreciation cycle.

05 — Segment deep diveNorth America: a low-margin business that is quietly getting better

North America is Amazon's largest segment by revenue and its second largest by profit. FY2025 net sales were $426.3 billion — 59% of the consolidated total — and operating income was $29.6 billion, a 6.9% margin. In Q2 2026 the segment generated $116.2 billion of revenue, up 16%, and $9.1 billion of operating income, an 7.8% margin.

A 7.8% operating margin on a retail business would be unremarkable for a pure-play retailer. It is remarkable for Amazon, because the segment contains a mixture of activities with radically different economics, and because the margin has more than doubled since 2023 — from 4.2% to 7.8% in three years, an improvement of 360 basis points.

Three forces explain that improvement, and it is worth separating them because they have different durability.

Advertising mix. Advertising services revenue was $68.6 billion in FY2025 and $19.8 billion in Q2 2026 alone, growing 26%. Advertising is recognised inside the North America and International segments but is not separately reported at the operating income line. The business is understood to carry a substantially higher margin than retail, because the incremental cost of serving an additional sponsored placement on a page view that has already been generated is close to zero. As advertising grows from 9.6% of consolidated revenue toward a larger share, reported segment margins rise mechanically, without any improvement in the retail operation.

Fulfillment density. Amazon's logistics network has been built out over fifteen years, and the marginal cost of delivering an additional package into a region that already has a same-day facility is materially lower than the average. Regionalised fulfillment, the shift from a national hub-and-spoke model to a set of regional networks, has shortened the average distance a package travels and reduced the number of touches per order. This is a genuine cost improvement and it is durable.

Third-party mix. Third-party seller services — commissions, fulfillment fees, and seller-facing services — generated $172.2 billion in FY2025, up 10%, and is now a larger revenue category than AWS. The economics of this business are structurally better than first-party retail because Amazon does not carry inventory risk, and because the take rate on a third-party transaction is applied to merchandise value that never appears in Amazon's income statement.

The retail revenue line is not the retail profit line

Online stores revenue was $269.3 billion in FY2025, growing 9%. That is the slowest-growing major category in the company, and it is the one that receives the most external attention. The attention is misplaced. A 9% growth rate on a $269 billion base adds $24 billion of annual revenue, which is more absolute revenue than most retailers generate in total. And the growth rate is a function of a maturing category rather than of competitive weakness: US e-commerce grew at a similar rate over the same period.

The question that matters is whether the retail operation is being managed for margin or for growth, and the segment margin data suggests the former. Amazon has slowed its fulfillment network expansion, closed or subleased excess capacity built during the pandemic, and reduced headcount in corporate functions. Those are margin-accretive decisions that cost revenue growth in the short run.

Prime, subscriptions, and the durability question

Subscription services revenue was $49.6 billion in FY2025, up 12%. This line is predominantly Prime membership fees plus associated digital content. It is the most durable revenue Amazon has: it is prepaid, it renews automatically, and it is the mechanism through which the retail, media, and advertising businesses reinforce one another. Amazon does not disclose Prime membership numbers, which means the metric cannot be tracked directly.

What can be tracked is engagement. Management disclosed in Q2 2026 that Alexa for Shopping active users had nearly doubled and interactions had increased more than fivefold year over year, that Amazon Now — the rapid-delivery service — was available in nine countries and more than 250 cities, and that Amazon Pharmacy new customers had more than doubled in six months with same-day prescription deliveries up nearly fivefold. These are all disclosures designed to demonstrate that the retail flywheel is still turning. None of them is audited and none has a revenue figure attached.

06 — Segment deep diveInternational: twenty years of investment, and a 2.9% margin

The International segment generated $161.9 billion of revenue in FY2025, up 13%, and $4.8 billion of operating income — a 2.9% margin. In Q2 2026 the segment reported $42.2 billion of revenue, up 15%, and $1.7 billion of operating income, a 4.0% margin.

This is the least-discussed segment and, on our reading, the one where the capital allocation question is sharpest. Amazon has been operating internationally since 1998, when it acquired Bookpages and Telebook to enter the UK and German markets. Twenty-eight years later the segment earns a margin roughly 40% of the North America figure on a revenue base that is 38% of North America's. The cumulative operating losses sustained by this segment over its history are not separately disclosed but are large.

The trajectory is improving. International margin went from −2.1% in FY2023 to 2.7% in FY2024 to 2.9% in FY2025 to 4.0% in Q2 2026. That is a genuine inflection, and the mechanism is the same as in North America: advertising mix, fulfillment density, and a shift toward third-party transactions. The difference is that the base is lower and the fixed-cost infrastructure is less mature, so the same forces produce a smaller absolute result.

Two structural questions sit underneath. The first is whether the international markets in which Amazon competes are winnable at scale. In several large markets Amazon faces a dominant domestic incumbent with a lower cost base, stronger logistics density, and, in some cases, regulatory protection: Mercado Libre in Latin America, Flipkart and now quick-commerce operators in India, Coupang in Korea, and a set of regional leaders across Southeast Asia. Amazon has chosen to compete in most of these markets rather than exit, and it has periodically retreated from others.

The second is capital allocation. International consumes capital — fulfillment infrastructure, technology, marketing — at a rate that has historically exceeded its cash generation. Every incremental dollar deployed there is a dollar not deployed into AWS, where the marginal return on invested capital is demonstrably higher. Management has never framed the choice that way publicly, and there is a reasonable argument that international scale is strategically necessary to support global enterprise customers. But the arithmetic is what it is: a 4% operating margin business competing against $220 billion of domestic capital requirements.

07 — AdvertisingThe business Amazon does not have to explain, and therefore does not

Advertising services revenue was $68.6 billion in FY2025, up 22%, and $19.8 billion in Q2 2026 alone, up 26%. On an annualised basis the business is running at roughly $80 billion, which would make it larger than AWS was in 2021 and roughly the size of a top-five global advertising platform.

The economics of this business are different in kind from anything else Amazon owns. Advertising revenue is generated on traffic that already exists and has already been paid for by the retail operation. The marginal cost of an additional sponsored product placement is a fraction of the marginal revenue, and the business therefore converts revenue to operating income at a rate that no other Amazon segment approaches. Analysts commonly assume an operating margin in the 40–60% range; Amazon has never disclosed the figure, and we treat any specific number as an estimate.

What can be established from the filings is the direction and the magnitude. If advertising carries an operating margin of, say, 50%, then the $68.6 billion of FY2025 advertising revenue contributed approximately $34 billion of operating income — more than AWS contributed, on one-ninth of the revenue. Under that assumption, advertising alone would represent roughly 43% of Amazon's consolidated operating income. Every dollar of advertising growth is therefore disproportionately accretive to consolidated margin, which is the mechanism behind the segment margin improvements documented earlier in this report.

Three growth vectors, and one that is not yet proven

Sponsored products. The core business, and still the largest. Amazon's search results are a purchase-intent surface with no equivalent in digital advertising: a user typing a product query is closer to a transaction than a user scrolling a social feed, and advertisers pay accordingly. The growth here is a function of ad load, which Amazon has increased gradually, and of advertiser count, which continues to expand as smaller merchants adopt the platform tools.

Prime Video advertising. Amazon introduced an ad-supported tier to Prime Video and has been scaling it. This is a genuinely new inventory pool, and it addresses a category — brand advertising, as opposed to performance advertising — in which Amazon has historically been weak. The strategic value is that it monetises content spend that was previously a pure cost, and that it gives brand advertisers a reason to move budget to Amazon from linear television.

Retail media beyond Amazon. Amazon has begun placing advertising on third-party properties and in physical stores. This is early and not separately disclosed.

The unproven vector is artificial intelligence in the ad stack. Amazon has invested heavily in generative tools for creative generation, audience targeting, and campaign optimisation. If those tools raise advertiser return on ad spend, they support pricing; if they mainly reduce the labour cost of running campaigns, they accrue to advertisers rather than to Amazon. Management has not disclosed the effect either way.

The regulatory exposure nobody discusses

Amazon's advertising business sits inside a regulatory structure that was designed for a different industry. The Federal Trade Commission's antitrust case against Amazon, scheduled for trial in March 2027, centres in part on the relationship between the marketplace and the advertising business — specifically whether Amazon's use of seller data and its control of the search placement algorithm constitute an unfair method of competition. A finding against the company in that case would not necessarily break the advertising business, but it could constrain the mechanisms by which it is monetised.

The European Union's Digital Markets Act imposes separate obligations on gatekeepers, including restrictions on self-preferencing and requirements around data use. Amazon has been designated a gatekeeper and has been adjusting its European practices accordingly. The direction of travel in both jurisdictions is toward more constraint on exactly the practices that make retail media economics so attractive.

08 — The Anthropic position$13 billion in, $190 billion on the books, and no ownership percentage disclosed

This section is the analytical core of the report. Amazon's headline results are not in question. What is in question is how much of those results represent the operating business, and how much represents the appreciation of a single financial asset that Amazon has never told investors how much of it owns.

The history, reconstructed from disclosure and press reporting

Amazon's relationship with Anthropic began in September 2023 with a $1.25 billion investment as part of a broader $4 billion commitment, completed in March 2024. AWS became the primary platform for training and serving Anthropic's Claude models. A second $4 billion investment followed in November 2024, at a valuation reported by private markets at approximately $40 billion. In April 2026 Amazon invested a further $5 billion, with an additional $20 billion left open subject to commercial milestones. In return, Anthropic committed to more than $100 billion of AWS spend over ten years and up to five gigawatts of Trainium capacity.

Total capital deployed by Amazon: approximately $13 billion. Anthropic's valuation over the same period: $61.5 billion in March 2025, $183 billion in September 2025, $380 billion at a Series G in early 2026, and $965 billion at a $65 billion Series H that closed on 28 May 2026. Anthropic's annualised revenue run-rate was disclosed at $47 billion in May 2026 and $65 billion in July 2026.

What is on the balance sheet

As of 30 June 2026 Amazon carried its Anthropic position at $190.4 billion, split between $97.9 billion of convertible notes and $92.5 billion of non-voting preferred stock. Against Amazon's $551.6 billion of book equity, that single position represents 34.5% of the total. It is the largest single asset Amazon owns after property and equipment, and it is larger than the net book value of the entire North America segment.

The structure of the instrument matters and is frequently misunderstood. Amazon's holdings are convertible notes that convert to preferred stock as Anthropic raises capital, and preferred stock carried at fair value. The consequence is that every time Anthropic completes a round at a higher valuation, Amazon books a gain — not when it sells, and not when Anthropic lists, but at the moment of the round. This is why the gains appear in "other income, net" and why they are so lumpy.

In Q1 2026, Amazon recorded a $12.3 billion upward adjustment on Anthropic preferred stock. In Q2 2026, pre-tax other income was $53.4 billion, which the company described as "primarily" attributable to its Anthropic investments. Across the first half, that is approximately $69 billion of pre-tax income generated by the revaluation of a private company, against $51.4 billion of operating income in the same period.

The IPO scenario, priced honestly

Anthropic filed a confidential S-1 during the summer of 2026, and press reporting has pointed to a listing as early as October 2026 at a valuation of approximately $2 trillion. At that valuation, and on the 19.7% ownership estimate, Amazon's stake would be worth roughly $394 billion — approximately $36.50 per Amazon share, or 14.7% of the current market capitalisation.

It is important to be precise about what that event would and would not do. It would not create $394 billion of new accounting gain in one step, because Amazon has already marked the position to $190.4 billion. It would reprice the gap between $190.4 billion and $394 billion — approximately $204 billion, or roughly $19 per share. And it would convert a private, illiquid, periodically revalued asset into a liquid, daily-marked one, which removes a discount that the market currently applies and introduces a new source of earnings volatility.

Our base case credits the position at its carrying value plus a modest step-up. Our bull case credits it at approximately $30 per share. Our bear case credits it at carrying value and assumes no further appreciation, on the reasoning that a private valuation marked at 31 times a $65 billion revenue run-rate is not a conservative entry point.

The half of the trade nobody prices

The equity upside is the loud part of the Amazon–Anthropic relationship. The quiet part is the commercial commitment. Anthropic has pledged more than $100 billion of AWS spend over ten years and up to five gigawatts of Trainium capacity. That commitment is a substantial contributor to the $496 billion AWS backlog discussed earlier, and it is the reason Amazon can underwrite Trainium design costs against a customer it can see.

There is a legitimate question about whether the two halves are independent. If Amazon's equity investment is, in substance, a mechanism for financing a customer's purchases, then some portion of the AWS revenue growth and some portion of the reported investment gain are two entries describing one economic event. Amazon does not disclose what proportion of revenue comes from entities in which it holds an equity interest, and it does not disclose what proportion of receivables are owed by such entities. Until it does, the question cannot be settled from public information.

What a markdown would look like

The asymmetry that receives least attention is the downside. Amazon's reported earnings now contain a large, periodically revalued asset, and the revaluation runs in both directions. A 30% decline in Anthropic's valuation from the Series H level would produce a pre-tax charge of roughly $57 billion — larger than any single quarter of Amazon's operating income has ever been. The charge would be non-cash, would not affect operating cash flow, and would be excluded from most adjusted earnings measures. It would nevertheless be reported, and it would be the headline.

We are not forecasting such a decline. We are noting that Amazon's reported earnings per share has become a variable with a distribution that is wider than the underlying operating business, and that the market has not yet had to price that. The first quarter in which the mark goes the other way will be informative.

09 — Quality of earningsThree numbers that do not appear in the press release

Amazon's reported results are audited, prepared under US GAAP, and not in question. What follows is not an allegation of accounting impropriety. It is an assessment of whether the headline numbers describe the operating business, or the operating business plus a set of financial transactions that would not exist if the operating business were smaller.

Finding one: net income now exceeds operating income

Over the twelve months to 30 June 2026, Amazon reported operating income of $93.7 billion and net income of $135.3 billion. The gap of $41.6 billion is 30.7% of reported net income and 44% of reported operating income. In the June quarter alone, operating income was $27.5 billion and net income was $62.6 billion — a ratio of 2.3 to 1.

This is an unusual configuration for an operating company. It arises because Amazon's non-operating income line contains realised and unrealised gains on equity securities that are marked through the income statement, and because the company's tax rate on those gains is the same as its tax rate on operating profit. A company whose net margin exceeds its operating margin is, by definition, reporting earnings that come substantially from outside operations.

Finding two: cash conversion has halved

Operating cash flow as a percentage of net income was 279% in FY2023, 196% in FY2024, 180% in FY2025, and 119% over the trailing twelve months to June 2026. Free cash flow as a percentage of net income over the same periods: 121%, 65%, 14%, and −6%.

Part of the deterioration in the free cash flow ratio is definitional: the denominator has grown faster than the numerator because net income is being inflated by non-cash gains. Removing the Anthropic revaluations, trailing net income is approximately $82 billion rather than $135 billion, and free cash flow of −$7.6 billion is then a negative 9% of adjusted earnings rather than negative 6%. That is worse, not better.

The operating cash flow ratio is the more informative figure, because it is less affected by the non-cash gains — operating cash flow starts from net income and subtracts the non-cash items, so the equity gains largely wash out. A decline from 279% to 119% means that working capital and other accruals are absorbing an increasing share of reported profit. For Amazon specifically, the largest single driver is inventory and receivables associated with the AWS buildout, plus the growth in operating lease liabilities as data centre capacity is leased rather than owned.

Finding three: capital intensity is consuming the cash flow statement

Amazon's capital expenditure has grown from $48.1 billion in 2023 to $169.0 billion over the trailing twelve months — a 3.5× increase in three years. Over the same period operating cash flow grew from $84.9 billion to $161.4 billion, a 1.9× increase. Capital expenditure has therefore grown roughly twice as fast as the cash flow available to fund it, and the difference has been met from the balance sheet.

This is the arithmetic that turns a capital allocation question into a financing question. A company that funds capital expenditure from operating cash flow is making a strategic decision. A company that funds it from operating cash flow plus $104 billion of bond issuance in twelve months is making a strategic decision and a leverage decision, and the second one has a counterparty.

10 — Capital intensity$220 billion, and the question of what it buys

Amazon has guided to approximately $220 billion of capital expenditure in 2026, raised from $200 billion at the Q2 earnings call. The increase of approximately $20 billion was attributed by management to higher memory prices, specifically to the diversion of high-bandwidth memory manufacturing capacity away from conventional DRAM toward AI data centre use. Memory makers are reported to earn three to five times more revenue per wafer on AI-grade product, and the resulting supply shift raised contract prices sharply through 2026.

To put the number in context: $220 billion is roughly 27% of Amazon's expected 2026 revenue, and approximately 2.1 times its expected 2026 operating income. It exceeds the GDP of roughly 130 sovereign states. It is the largest single-year capital programme ever undertaken by a non-government entity, and Amazon has said it expects to keep adding data centre capacity at significant scale for the next two to three years.

What the capital buys, and what it does not

Management's framing is that the spending is demand-led rather than speculative. The supporting evidence is real: $496 billion of RPO, most 2027 capacity already reserved, a "significant portion" of 2028 capacity committed, and Jassy's statement that Amazon will not have enough capacity to meet demand in 2026 or 2027 and that demand already visible for 2028 is "striking." The company is also on pace to double power capacity by the end of 2027 relative to 2025.

Three qualifications apply.

The capital is not fungible. Data centre capacity built for a specific workload and a specific customer is not easily redeployed. A facility optimised for large-scale model training has different power, cooling, and networking characteristics from one optimised for inference, and both differ from general-purpose cloud. If the mix of demand shifts — as it is currently shifting, with inference growing faster than training — some of the capital deployed in 2026 may be less well-matched to 2028 demand than the underwriting assumed.

The depreciation clock starts immediately. AI accelerators and the facilities built to house them are depreciated over five to six years at the major hyperscalers. The useful economic life of a training accelerator is contested: NVIDIA has moved to an annual architecture cadence, which means a part bought in 2026 will be two generations behind by 2028. The counter- argument, which has force, is that an older accelerator remains perfectly economic for inference on smaller models, for fine-tuning, and for the long tail of enterprise workloads. Rental rates for older generations have held up better than the depreciation sceptics predicted. The truth is somewhere between the two positions, and Amazon's disclosure does not let an analyst locate it.

The capital is being deployed against a market that all three major providers are simultaneously expanding. Combined 2026 AI infrastructure spending across the four largest hyperscalers is approximately $725 billion, up from roughly $410 billion in 2025. When supply in a market grows 77% in a year, the risk is not that demand fails to appear; it is that the price of capacity falls faster than the cost of building it.

Return on invested capital: the number that decides it

Amazon's trailing return on invested capital is approximately 11.9% on a third-party computation, against a weighted average cost of capital of approximately 11.2%. Computed on our own methodology — net operating profit after tax divided by total debt plus equity less cash — the figure is closer to 10.7%, which is below the cost of capital.

We want to be careful about what that means, because it is easy to over-read. A company in the middle of a large capital programme will always show a depressed return on invested capital, because the denominator grows immediately while the numerator grows only as the assets are commissioned and monetised. Amazon's capital expenditure has grown 3.5× in three years; the revenue it supports has not yet had time to arrive. A J-curve is expected, and the honest question is not whether the current return is low but whether the terminal return justifies the programme.

Two observations bear on that question. The first is that the current return is flattered by the Anthropic gains, which are pure profit with no invested capital attached; removing them lowers the computed return by roughly two percentage points. The operating business is not currently earning its cost of capital. The portfolio is. The second is that AWS margin expansion of 650 basis points over four quarters is direct evidence that at least part of the capital programme is earning an above-hurdle return today — because if it were not, the margin would be falling, not rising.

11 — Balance sheet & financingFrom a net cash fortress to a regular issuer of debt

Amazon's balance sheet has historically been one of the strongest in technology. The company has never paid a dividend, has run minimal leverage for most of its history, and has funded its capital programme from operating cash flow. That description no longer fits.

Long-term debt stood at $65.6 billion at 31 December 2025. It was $119.1 billion at 31 March 2026 and $128.9 billion at 30 June 2026. The July 2026 bond offering — an eight-part deal of at least $25 billion — settled after the balance-sheet date and lifts the total to approximately $154 billion. Over twelve months Amazon raised approximately $104 billion of bonds: $15 billion in a US offering in November 2025, roughly $54 billion across US and European markets in the first half of 2026, $10 billion in Canada in June 2026, and $25 billion in July 2026. The company has told underwriters it will not issue more debt in 2026.

Cash, total debt, net debt, equity and the Anthropic carrying value at 30 June 2026 are drawn from Amazon's Form 10-Q and from third-party aggregation of the reported balance sheet. Total debt includes operating and finance lease obligations, which is why it exceeds long-term debt by a wide margin; Amazon reports lease liabilities separately and does not present a "total debt" figure. The 31 December 2025 columns for cash, total debt, equity and net debt are Farstar estimates derived from the change in long-term debt and reported equity movements. Shareholders' equity grew partly because of the Anthropic revaluation, which flows through the income statement into retained earnings.

Why the borrowing is defensible, and what it changes

Amazon did not need to borrow to fund operations. It generated $161.4 billion of operating cash flow over the trailing twelve months and had $123 billion of cash and equivalents on the balance sheet. A company in that position that raises $104 billion of bonds is making a deliberate capital-structure decision, and the rationale is straightforward: the cost of debt is materially below the expected return on the capital being deployed, so financing the programme with debt rather than equity preserves the upside for existing shareholders.

That logic holds if the expected return is above the cost of debt, which we think is likely, and if the debt can be refinanced on acceptable terms through a period in which the assets have not yet generated their returns, which is less certain. Amazon's interest coverage is approximately 28×, its debt-to-EBITDA is approximately 1.35×, and its credit profile remains among the strongest in the corporate market. The Altman Z-score is 4.17, comfortably in the safe zone. On any conventional measure, the leverage is modest.

What has changed is the direction of travel and the strategic implication. Amazon has moved from funding growth internally to funding it in the capital markets, and it has done so at the same time as the other major hyperscalers. Alphabet raised $80 billion from stock sales to fund its buildout. NVIDIA issued $24.9 billion of debt in a single quarter. Meta, Oracle and Microsoft have all been active. The AI infrastructure cycle is now substantially debt-financed across the sector, which means the transmission mechanism for a demand shock is no longer a decision to stop buying but a credit market that becomes unwilling to fund the next gigawatt.

Capital return: nothing, and that is a deliberate choice

Amazon pays no dividend and, uniquely among the largest technology companies, has conducted no meaningful share repurchase. The share count has increased 0.88% over the past year, entirely from stock-based compensation. Every dollar of operating cash flow not consumed by capital expenditure is retained.

This is consistent with the capital allocation philosophy that has governed Amazon since its founding: reinvest everything as long as the marginal return exceeds the cost of capital. It is the correct policy for a company in that position. It also means that the shareholder receives no return at all until either the capital programme matures or the company changes its policy — and that the entire investment case rests on the terminal value of assets that have not yet been built.

12 — Custom siliconTrainium is the most consequential thing Amazon has built since S3

Amazon's custom silicon programme is the least understood part of the company and, on our reading, the one with the largest effect on the long-run economics of the equity. It comprises three product families: Graviton server processors for general-purpose compute, Trainium AI accelerators for training and inference, and Nitro virtualisation hardware. Management disclosed in Q2 2026 that the chips business had crossed a $25 billion annualised revenue run-rate, growing at a triple-digit rate, alongside an AI services business that crossed the same threshold on the same growth rate.

The strategic logic is not primarily cost. It is that AWS buys roughly 75% of its accelerator supply from a single vendor whose gross margin is approximately 75%, and that vendor is also competing for the same capacity allocation. Every dollar of accelerator spend that moves from merchant silicon to Trainium is a dollar on which Amazon captures the margin rather than paying it, and a dollar on which Amazon controls the supply chain.

Trainium3 and the specification question

Trainium3 is manufactured on TSMC's 3-nanometre process. A 144-chip UltraServer is rated at 362 MXFP8 petaflops of aggregate compute, which places it broadly at parity with NVIDIA's Blackwell NVL72 at rack scale. Third-party analysis has estimated a total cost of ownership advantage of approximately 50% relative to the merchant alternative, though that figure depends heavily on the utilisation rate assumed and on whether software engineering time is included. The interconnect is NeuronLink-v4 over a NeuronSwitch-v1 all-to-all switched fabric, which offloads inter-chip communication from the compute cores — the same architectural decision NVIDIA made with NVLink.

The comparison that matters is not peak throughput. It is model-FLOPs utilisation — the share of theoretical peak that a real training workload actually achieves. Independent benchmarking has consistently found that merchant alternatives to NVIDIA achieve a lower share of their theoretical peak in multi-node training, and there is no published MLPerf submission placing Trainium at parity on that metric. Amazon's position is that the cost per token, not the utilisation rate, is what customers buy, and that a 50% capital cost advantage survives a 15–20% utilisation disadvantage. That arithmetic works, but only just, and it works better for inference than for training.

The software moat, viewed from the other side

NVIDIA's advantage is not the chip. It is eighteen years of CUDA, a library ecosystem with no equivalent, and a developer population that has never had a reason to learn anything else. That advantage is real, and it is the reason a cheaper accelerator with comparable specifications has not displaced the incumbent.

Amazon's response has three parts. First, price: the Trainium cost advantage is large enough that a customer with a well-characterised workload and an engineering team can justify the port. Second, the model layer: Bedrock abstracts the underlying hardware behind a single API, so a customer consuming a model rather than training one never touches Neuron or CUDA. Third, managed services: Amazon will run the workload for you, which removes the porting problem entirely.

The competitive risk to NVIDIA is therefore not that Trainium is better. It is that the fraction of AI compute spending that is indifferent to the underlying hardware is growing — because more of it is inference on established models rather than training of new ones, and because the interface layer is consolidating. Amazon is well positioned for that world. So is Google, whose TPUs run the majority of Gemini inference. The company least well positioned for it is the one that owns the developer ecosystem but not the model layer or the cloud. That is not an argument against NVIDIA today; it is an argument about where the next five years of spending will be contested.

What is not disclosed

Amazon does not report custom silicon revenue as a line item, does not disclose the proportion of AWS capacity running on its own silicon, and does not disclose the gross margin of the silicon business. The $25 billion run-rate figure is management commentary on an earnings call, not an audited number, and it is not additive to the AWS segment total — it represents the revenue attributable to instances powered by those chips, not a separate revenue stream. A reader should treat it as a directional indicator rather than a modelling input.

13 — Competition ICloud: the leader is growing fast and losing ground

The global market for cloud infrastructure services reached $143 billion in the second quarter of 2026, up 43% year over year and the fastest growth rate in eight years. Over the last eleven quarters the market has doubled in size. Generative-AI-specific cloud services are growing at approximately 165% year over year, and — more importantly for the incumbents — AI is also accelerating growth across the broader range of cloud services, because every AI workload arrives with an attached data, networking, and storage requirement.

Within that market, the three largest providers held 67% of spend in Q2 2026. AWS remained the largest, but its share fell roughly two percentage points year over year while Google Cloud gained a similar amount. On growth rates: AWS 37%, Microsoft Azure 43% (fiscal fourth quarter), Google Cloud 82%. Google Cloud now has a revenue run-rate approaching $100 billion and is capturing more share than at any point in its history.

The competitive question for Amazon is not whether it will lose the market — it will not, in any horizon we model. It is whether the loss of share is a temporary consequence of Google's TPU advantage and Microsoft's enterprise distribution, or the beginning of a durable shift. Three factors bear on that.

Google's vertical integration is genuine. Google designs its own accelerators, runs the majority of Gemini inference on them, owns the model layer, and owns a distribution channel in Search and Android. That is a more complete stack than either AWS or Azure possesses, and it is the reason Google Cloud can grow 82% while still reporting improving margins. The constraint on Google historically has been enterprise relationships, and that constraint is loosening.

Microsoft's position is built on software distribution rather than silicon. Azure's advantage is the installed base of enterprise agreements that make Microsoft the default procurement decision. That is durable, and it is why Azure has grown in the low-to-mid 40s for several consecutive quarters despite a less differentiated silicon story.

Amazon's position is built on scale, silicon, and the model layer. It has the largest absolute revenue base, the most mature custom silicon programme, and — through Bedrock — the widest selection of third-party models. It has the weakest enterprise software distribution of the three, which is why it has been investing in the government and sovereign segments where procurement is relationship-driven and where it has been winning.

The capital expenditure chart above is the context in which all of this happens. Four companies will spend approximately $725 billion on AI infrastructure in 2026, up 77% from approximately $410 billion in 2025. Amazon's share of that total falls slightly, from approximately 32% to 30%, even as its absolute spend rises 68%. When supply in a market grows 77% in a year and the leader's share of the investment is flat to down, the leader's relative position is being diluted by capital rather than defended by it.

This is not a reason to be negative on Amazon. It is a reason to be precise about what the company's growth is a function of. AWS's 37% growth is being delivered against a market growing 43%. The absolute dollars are enormous and the contracted backlog is real. But the company is no longer the only credible supplier of frontier-scale AI capacity, and the pricing power that accompanies that position is weaker today than it was three years ago. The 39.4% AWS operating margin is the evidence that pricing power has not yet broken. It is the number to watch.

14 — Competition IICommerce: a mature market with a durable leader and a rising second

Amazon's retail operation is not the growth story it once was, and the company does not present it as one. Online stores revenue grew 9% in FY2025. Third-party seller services grew 10%. Physical stores grew 6%. Those are the growth rates of a mature business competing in a mature market, and they are the reason the consolidated revenue growth rate of 12–14% depends on AWS and advertising.

The competitive picture is nonetheless favourable to Amazon in the medium term, for reasons that have little to do with price.

Logistics density is the moat. Amazon's delivery network, built over two decades at a cost that is now largely sunk, allows same-day and next-day delivery at a unit cost that no competitor can match without replicating the investment. Walmart is the only retailer with a comparable physical footprint and it has been closing the gap, but the gap remains. The emergence of rapid-delivery specialists — the various quick-commerce operators across Europe, India and the United States — is a genuine competitive development in categories where speed is the only purchase criterion, but it is a niche rather than a threat to the whole catalogue.

Selection is a two-sided network. Amazon's marketplace has more sellers than any competitor, which produces more selection, which attracts more buyers, which attracts more sellers. Third-party sellers account for the majority of units sold on the platform. That structure is self-reinforcing and difficult to displace.

The low-price Chinese platforms are a real but bounded threat. Temu, Shein and their competitors have taken meaningful share in low-price, long-delivery categories. Their economics depend on the de minimis exemption in US customs law, which has been narrowed, and on direct-from-factory supply chains that cannot replicate Amazon's speed. They compete with Amazon on price and selection rather than on service, and Amazon has responded with its own low-price storefront. We treat this as a margin headwind in specific categories rather than a structural threat to the platform.

Walmart is the more serious competitor over a ten-year horizon. Walmart's e-commerce business has been growing faster than Amazon's for several years, from a much smaller base, and it has the physical infrastructure to build a delivery network at lower incremental cost than Amazon needed. It has also built an advertising business that, while smaller, grows faster. On the current trajectory Walmart becomes a credible second at scale in the United States within the decade. It does not become the leader.

The physical store question

Amazon's physical stores segment generated $22.6 billion of revenue in FY2025, up 6%, and has historically been loss-making. The company took a $610 million impairment against physical stores in Q4 2025. Whole Foods is a genuinely differentiated asset in grocery; the rest of the physical estate is a collection of experiments that have not scaled. Management has been disciplined about closing what does not work, which is the right approach and a change from the previous decade.

We ascribe no material value to physical stores in our base case. The segment is 3% of revenue, has not demonstrated a path to a meaningful operating margin, and has consumed capital that would have earned a higher return elsewhere.

15 — Adjacent betsLeo, healthcare, and the cost of a call option

Amazon operates several businesses that are immaterial to near-term earnings and potentially material to the long-term shape of the company. We treat them as call options rather than valuation inputs, and we size them honestly.

Amazon Leo

The satellite constellation formerly known as Project Kuiper has been renamed Amazon Leo. It targets a market that SpaceX's Starlink has already proven, using a launch programme spanning Atlas V, Falcon 9, Ariane 6, Vulcan Centaur and New Glenn across 80-plus planned missions. The programme has encountered a meaningful constraint: launch capacity. Reporting through mid-2026 described hundreds of satellites queued awaiting rides to orbit, with Ariane 6 missions representing a bottleneck.

Amazon has committed a reported $10 billion-plus to the programme and holds a Federal Communications Commission licence that carries deployment milestones. The strategic logic is straightforward: a low-earth-orbit network extends AWS connectivity to locations that fibre cannot economically reach, supports the government and defence business, and creates a second revenue stream in a market where the incumbent has demonstrated both demand and pricing power. The execution risk is real, the capital requirement is large, and the competitive position is a distant second. We ascribe no value to Leo in our base case.

Healthcare

Amazon Pharmacy reported new customers more than doubling in the first six months of 2026 and same-day prescription deliveries up nearly fivefold. Amazon One Medical's weight-management programme is now available in 48 states and Washington DC, through both insurance-covered in-person care and cash-pay virtual care. These are real businesses with real revenue, and they are not separately disclosed.

The healthcare opportunity is genuinely large and structurally attractive: high margins, recurring revenue, and an incumbent industry with poor customer satisfaction. It is also an industry with entrenched regulatory capture, complex reimbursement economics, and a long history of technology entrants underestimating the difficulty. Amazon has been in healthcare for over a decade with limited financial impact. We treat it as optionality.

Robotics and automation

Amazon's warehouse automation programme is the adjacent bet with the clearest financial connection to the core business, because it reduces the cost of the largest expense line in retail. The company disclosed a set of new robotic systems in October 2025, including the Blue Jay robotic arm, and then discontinued Blue Jay in February 2026 after less than six months, reassigning the underlying technology to other programmes. That sequence is informative: Amazon is willing to kill projects quickly, which is a positive governance signal, and warehouse robotics at general-purpose scale remains harder than the demonstrations suggest.

The financial significance is nonetheless real. Amazon employs approximately 1.58 million people, the overwhelming majority in operations. A one percentage point reduction in the cost of fulfillment, achieved through automation, is worth roughly $1.5–2 billion of annual operating income. That is why the programme continues despite the setbacks.

16 — Cost structure1.58 million employees, and the productivity question

Amazon employed approximately 1,576,000 people at the most recent count, the second-largest private employer in the United States. Revenue per employee over the trailing twelve months was $492,183, up from $376,900 in FY2023 — a 31% improvement in three years. Net income per employee was $85,838, though that figure is inflated by the Anthropic revaluations and is better stated as approximately $52,000 on operating earnings.

The productivity improvement is almost entirely a revenue story rather than a headcount story. Amazon's headcount has been broadly flat since 2024, while revenue has grown 35%. The company has not meaningfully reduced its workforce; it has grown revenue on a stable base of people.

The corporate headcount reduction

Amazon eliminated approximately 30,000 corporate roles in October 2025 and a further 16,000 in January 2026 — roughly 46,000 positions in total, described by the company as an anti-bureaucracy initiative. Against total headcount of 1.58 million, that is under 3%. Against corporate headcount, which is estimated at roughly 350,000 before the reductions, it is closer to 13%, and it is a materially larger share of corporate cost than of total cost.

The financial effect is not separately disclosed. Amazon took $730 million of estimated severance costs in Q4 2025, which is the disclosed portion; the annualised saving from 46,000 corporate positions, at a fully loaded cost that we would estimate in the $180,000–250,000 range, is approximately $8–11 billion before reinvestment. Some of that has been reinvested in AI talent and infrastructure. The net effect on operating income is therefore smaller than the gross saving, and we model a net benefit of roughly $4–6 billion annualised, which is approximately 4–5% of our FY2027 operating income estimate.

The structural cost question

The most important cost development in Amazon's income statement is not headcount. It is depreciation. Amazon's trailing depreciation and amortisation was $75.2 billion, against operating income of $93.7 billion. In other words, 80% of operating income is consumed by depreciation and amortisation before any cash cost is considered. As the $220 billion of 2026 capital expenditure is commissioned, that figure will rise materially — potentially to $100 billion or more on an annualised basis by 2028.

The arithmetic of the next two years is therefore as follows. Amazon must grow operating income fast enough to absorb a depreciation charge that is growing at the rate of the capital programme. If capital expenditure stabilises at $220 billion and the assets are depreciated over six years, the steady-state depreciation charge from that programme alone is approximately $37 billion a year. Combined with the existing base, that implies total depreciation approaching $110 billion by 2029. To hold the current operating margin, Amazon needs roughly $150 billion of operating income by 2029 — approximately 60% above the trailing twelve-month figure.

That is achievable on our base case. It is not achievable without the revenue growth continuing. This is the mechanism by which an infrastructure buildout becomes a fixed-cost problem, and it is the reason the depreciation schedule is the most consequential assumption in any Amazon model.

17 — RegulationThe trial is in March 2027, and the market has stopped pricing it

Amazon faces regulatory exposure on four fronts, of which one is material to the investment case and three are background.

The Federal Trade Commission's monopolisation case

The FTC and a group of state attorneys general filed suit against Amazon in September 2023, alleging that the company illegally maintains monopoly power through a set of interlocking practices: penalising sellers who discount elsewhere, conditioning Prime eligibility on the use of Amazon's fulfillment services, and degrading the search experience in favour of paid placements. After several scheduling changes, a December 2025 order set a bench trial to begin 29 March 2027 before Judge John Chun in the Western District of Washington.

The remedies available in a monopolisation case under Section 2 of the Sherman Act are broad. They range from behavioural constraints — prohibiting specific practices — to structural relief. The FTC has not publicly specified what it would seek if it prevailed, and Amazon contests the allegations.

Our assessment is that this is a genuine risk of indeterminate magnitude that the market is currently pricing at close to zero. Two considerations support that reading. First, the case has been in the docket for three years and has not produced a visible change in Amazon's behaviour or in the share price. Second, antitrust cases against large technology companies have historically been slow, appealable, and frequently settled on terms that impose costs rather than restructuring.

But the practical exposure is worth stating. The advertising business, which we estimate contributes roughly 40% of consolidated operating income, depends on Amazon's control of search placement on its own properties. The third-party seller services business, at $172 billion of annual revenue, depends on the fee structure and the fulfillment tie that the FTC is challenging. A remedy that constrained either would be more consequential than any fine.

The Prime settlement

In September 2025 Amazon agreed to pay $2.5 billion to settle FTC allegations that it used deceptive design to enrol customers in Prime and obstructed cancellation — the largest consumer protection settlement in the agency's history. The settlement comprised a $1 billion civil penalty and $1.5 billion in customer redress covering approximately 35 million affected subscribers. It is a closed matter and financially immaterial, but it establishes a precedent for the enforcement posture the agency takes toward Amazon's interface design.

European regulation

The Digital Markets Act designates Amazon as a gatekeeper and imposes obligations on self-preferencing, data combination, and access to the marketplace for business users. Amazon has been adjusting its European practices in response. The direction of travel in Brussels is toward more constraint on the practices that make retail media economics attractive, and enforcement has become more assertive across the sector.

The financial exposure is smaller than the US exposure because the European businesses are smaller. The strategic exposure is that European remedies often become templates for other jurisdictions, including the United States.

Export controls and geopolitical exposure

AWS operates a global infrastructure footprint and sells to sovereign customers, which exposes it to export control regimes, data localisation requirements, and the possibility of restrictions on serving specific customers. The company's $50 billion commitment to US federal agencies is politically favourable but increases its exposure to changes in government procurement priorities. Its Chinese business is minimal.

We do not assign probability-weighted value to any specific regulatory outcome. We note that the regulatory environment for large technology platforms has become less predictable in both the United States and Europe, and that this argues for a modest discount to the multiple rather than a scenario adjustment.

18 — ValuationThe multiple depends entirely on which earnings you use

Amazon's valuation is unusually sensitive to a definitional choice, and the choice is not a matter of preference. The company reports two earnings numbers: a GAAP figure that includes non-operating investment gains, and an operating figure that does not. At the current price those two numbers produce multiples that differ by 60%.

On reported GAAP earnings the shares trade at 20×, which for a company growing revenue at 14% and operating income at 33% looks undemanding. On operating earnings they trade at 32×, which for a company with negative free cash flow and $129 billion of long-term debt looks full.

Our position is that the operating number is the correct one for valuing the business, and that the Anthropic position should be valued separately as an asset rather than as a multiple applied to a mark. There is a coherent argument for the opposite treatment — if the market is willing to capitalise Anthropic gains, the company has genuinely created value and the shareholder is better off. But capitalising a mark-to-market gain on an unlisted security at a full earnings multiple requires an assumption that the gain is permanent and repeatable, and we are not willing to make that assumption. Our model therefore values the operating business on operating earnings and adds the Anthropic stake at its carrying value in the base case.

Cross-checks

The cross-checks do not resolve the question, and we present them because they frame it. On enterprise value to revenue, Amazon trades below its own five-year average — a measure on which the shares look inexpensive. On enterprise value to EBITDA it trades in line with large-cap technology — a measure on which the shares look fairly valued. On operating earnings it trades at a premium to that group — a measure on which the shares look expensive. All three are correct; they are measuring different things, and the reason they diverge is that EBITDA and revenue do not capture the depreciation that the capital programme is about to produce.

Scenario valuation

The table below sets out our three cases in full. Every input is stated so that a reader can disagree with a specific assumption rather than with the conclusion.

Weighting the three cases at 25/50/25 produces a scenario value of $286 per share. We then apply a discount for time value and for the width of the outcome distribution to arrive at a 12-month target of $270, approximately 9% above the current price.

Against a consensus 12-month target of approximately $328 across 62 covering analysts, our $270 sits below. We are comfortable with that divergence. Consensus ratings on Amazon are overwhelmingly positive — the aggregate is a Strong Buy — and the target dispersion is narrow, which is itself a signal. When every analyst agrees on a large-cap company with a contested capital cycle, the disagreement is usually about timing rather than about fundamentals, and the market has already absorbed the consensus view.

19 — The bull caseSix arguments we take seriously

We present the bull case at full strength, not as a straw man. If we cannot state the strongest version of the opposing argument, our own position is not worth publishing.

20 — The bear caseSix arguments we cannot dismiss

21 — Risk matrixRanked by probability times impact

The matrix below scores each identified risk on a 1–5 scale for probability and impact. The score is the product. Risks scoring 12 or above are the ones we would expect to drive a material re-rating in either direction.

Cells read left to right as impact rises from 1 to 5, top to bottom as probability rises from 1 to 5. Superscripts identify risks from the register above. R11 is not shown; it falls below the plotted range.

22 — CatalystsWhat we are watching, and when

23 — ConclusionNeutral at $270, on a business that is performing and a price that already assumes it will

Amazon's operating businesses are performing better than at any point in the last decade. AWS grew 37% — the fastest in eighteen quarters — while expanding margin 650 basis points. Advertising grew 26% on a business approaching $80 billion of annualised revenue. Consolidated operating margin reached 13.7%, a record. The company holds $496 billion of contracted backlog and has stated that most of its 2027 capacity is already reserved.

Amazon is also now a materially different financial entity from the one that most investors hold in mind. It reported net income 44% above its operating income over the trailing twelve months. It carries $190.4 billion of a private AI company on its balance sheet, representing 34.5% of book equity, without disclosing how much of that company it owns. Its free cash flow is negative for the first time since 2022. It has added $63 billion of long-term debt in six months and roughly $104 billion in twelve. And its return on invested capital is currently indistinguishable from its cost of capital, at the exact moment it is deploying capital faster than any company in history.

The tension between those two paragraphs is the investment case. Our scenario-weighted value is $286 per share, and our 12-month target is $270, approximately 9% above the current price. We rate the shares Neutral.

24 — AppendixFinancial summary tables

A. Annual income statement, $ billions

B. Segment detail, $ billions

C. Revenue by product and service line, $ billions

D. Key operating and quality-of-earnings metrics

E. Valuation summary

25 — SourcesPrimary and secondary references

Company filings and releases

• Amazon.com, Inc., Form 10-K for the fiscal year ended 31 December 2025, filed with the SEC.
• Amazon.com, Inc., Form 10-Q for the quarter ended 30 June 2026.
• Amazon.com, Inc., Form 10-Q for the quarter ended 31 March 2026.
• Amazon.com, Inc., Second Quarter 2026 Results, 30 July 2026.
• Amazon.com, Inc., First Quarter 2026 Results, 29 April 2026.
• Amazon.com, Inc., Fourth Quarter and Full Year 2025 Results, 5 February 2026.
• Amazon.com, Inc., 2025 Annual Report to Shareholders and Andy Jassy's shareholder letter, 9 April 2026.
• Amazon.com, Inc., Form 424(b)(5) prospectus supplement, July 2026 bond offering.
• Amazon.com, Inc., quarterly earnings call transcripts, Q4 2025 through Q2 2026.
• Amazon Newsroom, product and corporate announcements, October 2025 – September 2026.

Third-party research and industry data

• Synergy Research Group, global cloud infrastructure services market share and market size, Q2 2026, as reported by CRN, 6 August 2026.
• CRN, AWS vs. Microsoft vs. Google Cloud Earnings Q2 2026 Face-Off, 3 August 2026.
• CNBC, Amazon Q2 2026 earnings coverage and bond issuance reporting, July 2026.
• Business Insider and Reuters Breakingviews reporting on Amazon's Anthropic investment and convertible note structure, 2026.
• Anthropic Series H funding announcement, 28 May 2026.
• The Motley Fool, analysis of Amazon's Anthropic position and the $2 trillion IPO scenario, 15 September 2026.
• Moody's Ratings and other credit research on hyperscaler capital expenditure and leverage, 2026.
• Third-party market data aggregators for consensus price targets, valuation multiples and balance-sheet aggregation, September 2026.

Regulatory and legal

• Federal Trade Commission and State Attorneys General v. Amazon.com, Inc., No. 2:23-cv-01495 (W.D. Wash.), scheduling order of December 2025 setting trial for 29 March 2027.
• Federal Trade Commission, Amazon to pay $2.5 billion to settle Prime enrolment allegations, September 2025.
• European Commission, Digital Markets Act gatekeeper designations and compliance reporting, 2025–2026.
• US Bureau of Industry and Security export control framework, 2025–2026.

Market data

• Share price, market capitalisation, share count and valuation multiples as of the close on 15 September 2026.
• Consensus price targets and coverage counts as compiled by third-party market data aggregators, September 2026.

26 — DisclosureConflicts, limitations, and revision policy