Amazon.com, Inc.

01 — Executive summaryTwo announcements, read together

On 20 April 2026 Anthropic promised to spend more than $100 billion over ten years buying computing from Amazon. Amazon, in the same announcement, promised to invest up to $33 billion in Anthropic. Read together, the two announcements describe something with no precedent in American corporate history: Amazon is now both the landlord and a major shareholder of the two artificial-intelligence laboratories that will decide what the next decade of computing costs.

The company reported $200.61 billion of net sales in the quarter to June 2026, up 20%, and $27.46 billion of operating income, up 43%. AWS grew 36.8%, its fastest in eighteen quarters, on a contracted backlog of $496 billion. Advertising grew 26% to a $79 billion run rate.

Then there is the rest of it. Reported net income was $62.65 billion, or $5.75 a share, of which $53.4 billion was other income — and $50.49 billion of that was an upward adjustment to private-company equity investments, attributed primarily to Anthropic preferred stock. That holding, plus a second Anthropic instrument carried as convertible notes, is worth $190.4 billion: 17.4% of the group's assets and 34.5% of its book equity.

Meanwhile 2026 capital expenditure is guided to $220 billion, and trailing-twelve-month free cash flow is negative $7.6 billion. The advertising auction — the highest-margin thing Amazon sells, worth roughly $79 billion a year — is the subject of a suit filed on 31 August 2026 by the Federal Trade Commission and twenty-two states.

02 — How to read this reportThree ledgers, and the discipline of not adding them together

Amazon is now three things with three different accounting treatments, and most of the confusion in the debate comes from reading one through the arithmetic of another.

03 — What Amazon is nowA retailer that stopped being one

Revenue grew 20% in the June quarter, the fastest since 2021. The guidance for the current quarter — $197 to $202 billion — looks like a deceleration to 9% to 12%, and it is not: Amazon moved Prime Day from July into June, so the third quarter laps a quarter that contained the event. Excluding the shift, growth "would be nearly 400 basis points higher".

The composition is where the story is. In the June quarter AWS produced 21.1% of net sales and 60.5% of operating income — the ratio that explains why the capital is being spent as it is and why the valuation is so sensitive to one segment's growth rate.

04 — The rent roll: AWSEighteen quarters of deceleration, then this

For three years the bear case on Amazon was one number. From a peak of 40% in 2021, AWS growth slid quarter after quarter into the high teens while Azure and Google Cloud grew faster. AWS had won the first cloud war and was losing the second.

Then it stopped. Six consecutive quarters of acceleration, from 17% to 36.8%. Revenue is annualising at $169 billion and the contracted backlog reached $496 billion — roughly 2.9 times the run rate.

The re-acceleration is real and it is showing up in profit as well as revenue: AWS operating income rose 64% to $16.62 billion, with the margin expanding from 33.1% to 39.4%. And the growth rate is still the slowest of the three hyperscale clouds.

Both charts are true and they point in opposite directions. AWS earns a better margin than Google Cloud on a larger revenue base — the strongest evidence that it is the best-run infrastructure business in the industry. It is also growing at less than half Google Cloud's rate. That is not a reason to avoid the shares: 39% margins on a $42.2 billion base compound faster in absolute dollars than 82% on a base one-third the size.

05 — The tenantsThe two largest AI labs are both customers now

In February 2026 Amazon announced a strategic partnership with OpenAI — strange news on its face, because Amazon had spent three years and $8 billion building a relationship with Anthropic, OpenAI's principal rival. "Today, we now have the two largest AI labs who are both significantly betting on Trainium," Jassy told CNBC. Two months later Anthropic answered with a ten-year commitment of its own.

The structure of both deals is the same. Each laboratory commits to buy computing from AWS over a decade-scale term; Amazon commits capital to the laboratory. Anthropic: more than $100 billion of purchases over ten years, spanning Graviton and Trainium2 through Trainium4, against an investment of $8 billion already made, $5 billion added in April 2026, and up to $20 billion more available. OpenAI: $100 billion over eight years against up to $50 billion, of which $15 billion is committed and $35 billion contingent.

Anthropic's numbers are moving faster than any of the contracts. Its run-rate revenue passed $30 billion in April 2026, up from roughly $9 billion at the end of 2025 — a tripling in about four months. More than 100,000 customers run Claude on Amazon Bedrock, and Anthropic's own announcement is candid about the strain: growth "places an inevitable strain on our infrastructure".

06 — The circularity$200 billion is $22.5 billion a year

Here is where the analysis usually stops and the adjectives begin. Two hundred billion dollars of committed demand is an enormous-sounding number, and it has been repeated as though it were revenue. It is not, and on an annual basis it is a good deal smaller than it sounds.

Spread over their stated terms, the two commitments are worth roughly $22.5 billion a year — about 13.3% of AWS's current annualised run rate. That is material. It is also not the reason AWS grew 37% this quarter: AWS added roughly $11.4 billion of quarterly revenue year on year, while the lab commitments fully drawn would contribute about $5.6 billion a quarter.

The circularity is real but narrower than the excitable versions suggest. Amazon has created a financing facility for Anthropic of up to $20 billion, of which $15 billion remains after the Series H investment, so Anthropic can receive capital from Amazon and spend it on AWS. That is a loop, but it is not a fiction: the compute is delivered and the invoices are real. The answer is not to exclude the revenue — it is to refuse to pay a premium multiple for the portion that depends on Amazon's own capital.

07 — The markA third of the book is one private company

On 30 June 2026 Amazon's balance sheet carried two Anthropic instruments: nonvoting preferred stock at a fair value of $92.5 billion, and convertible notes at $97.9 billion. Together, $190.4 billion.

The two are not the same asset wearing different labels. The preferred stock is marked through other income when observable prices change, so a revaluation lands directly in reported earnings — as $50.49 billion of it did in the June quarter. The convertible notes run unrealised gains through other comprehensive income instead, changing book equity without touching the income statement. Both are Level 3.

There is a legitimate bull argument here. If Anthropic lists, or sells secondary stock at a higher price, Amazon's mark rises and the market's confidence is validated by a real transaction. The counter-argument is not that the mark is wrong: it is that a Level 3 mark on a company whose largest customer is also its largest shareholder is not the same quality of asset as cash.

08 — What the earnings wereThe gap between $27.5 billion and $62.6 billion

Amazon's June quarter produced the single largest gap between operating profit and net profit in its history as a public company: operating income of $27.46 billion against net income of $62.65 billion. The headline growth rate of 245% is not a statement about the business.

The bridge is short. Interest income of $1.30 billion and interest expense of $1.31 billion roughly cancel. Other income contributes $53.42 billion, $52.30 billion more than a year earlier; of that, $50.49 billion is upward adjustments to private-company equity investments, attributed primarily to Anthropic preferred. Income before tax was $80.86 billion and the tax provision $18.20 billion, an effective rate of 22.5%.

None of this makes the quarter bad: operating income grew 43%. But a reported $5.75 of earnings per share that annualises to $23 is not a meaningful valuation input for a company whose operating earnings annualise closer to $8.

09 — Normalised earningsWhat the business actually earned

Normalising Amazon's earnings is not a matter of adjusting for one-off items in the ordinary sense. It is a matter of removing an entire line that reflects the price of a private company. We strip the whole of other income, retain the net interest position, and apply the blended 22.5% tax rate.

The result is that the first half of 2026, reported at $8.57 a share, was closer to $3.53. Amazon does not disclose the tax effect of each investment item separately, so these are estimates and we label them as such. On the same basis we model full-year 2026 normalised earnings at approximately $7.40 a share — which, at $262.43, is 35.5 times current-year normalised earnings. That is not the number most Amazon investors carry in their heads.

10 — The bill: capital spendingTwo hundred and twenty billion dollars

Amazon's capital expenditure has gone from $48.1 billion in 2023 to a guided $220 billion in 2026 — a 4.6-fold increase in three years, and the largest corporate capital programme in the world outside the sovereign wealth funds.

The guidance was raised twice. In February, Amazon said $200 billion; in July, Jassy lifted it to $220 billion and attributed the increase to higher memory prices — the same shortage that drove DRAM prices up 80% to 90% and pushed Amazon's own Echo Dot retail price up 60%. The largest line in its forecast moved because of a component cost it does not control.

The cash-flow consequence is visible. Free cash flow was $11.2 billion for the whole of 2025, down from $38.2 billion in 2024. Over the twelve months to June 2026 it is negative $7.6 billion — the first sustained period of negative free cash flow since 2022, and unlike 2022 not because revenue is falling.

You cannot sell compute you have not built, and demand exceeds capacity through 2027. But the question is not whether the spending is justified; it is who bears the risk while it is spent. Amazon is converting a capital-light business into a capital-intensive one, funded partly with debt. If demand arrives, the returns are enormous. If it arrives two years late, the depreciation arrives anyway.

11 — The bill: how it is paidThe gap is filled with bonds

In the twelve months to 30 June 2026 Amazon generated $161.4 billion of operating cash flow — up from $121.1 billion a year earlier — and spent $169.0 billion on capital items. The difference is negative $7.6 billion. In the June quarter alone the company issued $37 billion of US dollar notes, plus further issues in euros, Swiss francs and Canadian dollars, and long-term debt reached $149.6 billion.

This is not distress. Amazon holds $86.8 billion of cash and short-term investments, carries investment-grade ratings, and generates more cash in a year than most listed companies are worth. But the direction of travel is unambiguous: the company is borrowing to build while the ten-year Treasury yields above 5%.

12 — The toll: advertisingThe best business Amazon owns is not the cloud

There is an argument that Amazon's most valuable asset is not AWS and not Anthropic but the advertising business that sits invisibly inside its retail segments. It has the economics of a toll booth: Amazon controls the point at which a shopper has already decided to buy something, charges suppliers for the right to be seen at that moment, and incurs almost no incremental cost.

Advertising services revenue was $19.81 billion in the June quarter, up 26% — six points faster than the group. It passed $68 billion for the whole of 2025, and on an annualised basis it is now a $79 billion business: roughly 47% the size of AWS while requiring a fraction of the capital. Amazon does not disclose its operating profit, but the direction of the evidence is not in doubt: retail margins have expanded while advertising grew at more than twice the rate of retail sales, and most analysts who attempt the decomposition land on a margin north of 40%. Which is why the next section matters more than any other risk here.

13 — The toll under attackThe auction is the lawsuit

On 31 August 2026 the Federal Trade Commission and twenty-two states filed suit against Amazon over how it priced advertising. The complaint is about the mechanism by which the most profitable thing Amazon sells is priced, and it alleges that mechanism was a misrepresentation for seven years.

The mechanics, as alleged, are specific. Amazon told more than 500,000 advertisers that it ran a "second-price" auction, in which the winning bidder pays one cent more than the second-highest bid rather than their own. The FTC alleges that from 2019 Amazon quietly changed the auction, adding a hidden surcharge it called a "soft reserve price" and introducing what an internal document described as an "invented auction participant" — a synthetic bidder used to push the clearing price above what real competition would have produced. Instead of paying a cent above the runner-up, Sponsored Products advertisers were charged their own full winning bid close to 80% of the time.

Two things temper the alarm. The first is time: this case will not be resolved quickly. The second is precedent: the earlier Prime enrolment case ended in a $2.5 billion settlement, of which more than $845 million had been paid to consumers by September 2026. That was a fine, not a restructuring. The misconduct alleged here is about the price of the product.

14 — The counter: retailThe part nobody watches

For most of the last decade the retail business was the reason to be embarrassed about owning Amazon: enormous, low-margin, labour-intensive and periodically written down. The interesting argument was always AWS.

That framing is now out of date. North American operating income rose from $7.5 billion to $9.1 billion in the June quarter — a 7.8% margin on $116.2 billion of sales, in a quarter that included the shifted Prime Day. International, long the loss-making sibling, produced $1.7 billion of operating income on $42.2 billion of revenue, a 4.0% margin that has quietly tripled in four years.

Three forces are doing the work, and only one is under management's control. Advertising grew 26% against retail sales growth in the mid-teens. Fulfilment has genuinely improved: regionalised inventory, a same-day network, and a delivery density that makes the marginal parcel cheaper to deliver than five years ago. And the comparisons are flattering — the retail segments of 2026 are measured against a 2025 that carried the cost of overbuilding.

That third point is the one to watch. Retail margins at Amazon have expanded before, in 2018 and again in 2020, and both times they reverted. What is different this time is the advertising mix and the regional fulfilment network, which are structural rather than cyclical. What is not different is the exposure to consumer confidence, to tariff policy — Amazon booked roughly $640 million of IEEPA tariff refunds in the June quarter, a benefit that will not repeat — and to the labour market.

15 — The chipsThe business that is bigger than it looks and smaller than it sounds

Amazon's custom silicon programme is the most interesting thing in the company that almost nobody models. It began as a cost-reduction exercise and has become, by Jassy's own description, one of the top three data-centre chip businesses in the world.

The numbers require care because Amazon presents them on three different bases. On the run-rate basis the chips business is over $20 billion and growing triple digits year over year. On the market-price basis — Jassy's framing, in which the unit sold its output at what a merchant chipmaker would charge — the run rate would be approximately $50 billion. On the commitment basis, Amazon has over $225 billion of contracted Trainium revenue, which is not annual revenue.

The programme matters not for its revenue line, which is still small against $200 billion of quarterly sales, but for its margin. Every workload that migrates from a purchased GPU to a Trainium chip converts a pass-through cost into an owned one.

The risk to that argument is temporal rather than technological. Trainium4 is eighteen months away, and CUDA remains the default for a generation of engineers who have never written for anything else. The $225 billion of commitments is evidence that the port is being made. It is not evidence that it has been made.

16 — The long betsTwelve per cent deployed, and the capital is already spent

Amazon Leo, the satellite constellation that spent four years called Project Kuiper, is the largest of the company's unmodelled ventures — and the one where the gap between the ambition and the deployment is most visible.

By July 2026 Amazon Leo had completed 14 missions and put more than 375 satellites in orbit, making it the third-largest constellation in operation. The design calls for 3,236; roughly 12% is deployed, and the capital required to reach the rest sits inside the same capital expenditure figure that carries AWS. The strategic logic is the logic that produced AWS — Amazon moves enormous quantities of data and would rather own the pipes. The competitive problem is that the market has a first mover with more satellites, more spectrum and more subscribers.

We credit Amazon Leo at $15 billion in the sum of the parts, deliberately conservative: a constellation 12% deployed, competing against a rival at scale, with no disclosed revenue, is an option rather than an asset. None is large enough to move a $2.86 trillion valuation.

17 — The cost sideThirty thousand roles, and where the money went

The most politically visible thing Amazon has done in the last twelve months is remove people. The corporate reductions run to roughly 30,000 roles since October 2025 — about 14,000 announced that month and 16,000 confirmed in January 2026 — with a further tranche of under 1,000 in early October 2026.

Management has not attributed the reductions to artificial intelligence, and the more defensible reading is the older one: Amazon over-hired through the pandemic, then spent three years correcting. What matters for an equity holder is not the cause but the destination. The savings have not been returned to shareholders. They have gone into data centres.

The other cost pressure is the one Amazon has least control over. Memory prices rose 80% to 90% on artificial-intelligence demand, and the company cited that specifically when it raised 2026 capital expenditure guidance from $200 billion to $220 billion. Consumer hardware prices followed.

18 — CompetitionAccelerating, and still the slowest

AWS grew 36.8% — the fastest in eighteen quarters — and finished third of three. Azure grew 43% in Microsoft's fiscal fourth quarter. Google Cloud grew 82%. That last number should keep an Amazon shareholder awake, because it is not a small base: Google Cloud is now roughly $25 billion a quarter, comparable to AWS in the middle of the last decade, growing at four times AWS's rate.

Google owns the model, the silicon, the data and the distribution, and it has been willing to sign the largest capacity contracts in the industry. Microsoft owns the enterprise relationship and the OpenAI distribution, and its partnership with Anthropic means it now has a claim on both frontier labs too. Amazon owns the largest installed base, the best margins and the most complete silicon roadmap. What it does not own is a frontier model: Nova exists, but the head of the San Francisco AGI lab left in February 2026 and the group now reports into the cloud organisation. Amazon has decided to be the landlord rather than the tenant — defensible, because landlords do not need to be right about which tenant wins, but it caps the multiple.

In retail the picture is the reverse: Amazon's share of US e-commerce has been stable for three years, its logistics network is now a genuine moat, and its nearest competitor by online volume is a Chinese marketplace with regulatory problems in three continents.

19 — Regulation beyond advertisingThe other case, and the one that will not settle

The advertising suit is the newest regulatory problem and not the largest. The Federal Trade Commission's long-running monopolisation case against Amazon's marketplace — alleging that the company punishes sellers who discount elsewhere and conditions Prime eligibility on using its fulfilment network — is still live, and on 9 October 2026 a Washington federal judge refused the commission's request to postpone the trial.

The distinction between the two matters is worth holding onto. The marketplace case is structural and usually ends in behavioural remedies and fines. The advertising case is a pricing case: quieter, faster, and it changes economics directly. For a shareholder the advertising case is the more dangerous of the two.

Beyond antitrust, three exposures belong in the register. Tariff policy remains volatile, and the pass-through of duties to third-party sellers is a live question about Amazon's liability for the pricing decisions of others. Labour regulation is a persistent cost risk across a network with more than a million employees. And tax policy is an unusually large lever on a company that earns much of its profit in jurisdictions with active digital-tax debates.

20 — ValuationFour methods, one median, and one refusal

Amazon resists a single valuation method more than any other large company, because the three ledgers do not respond to the same arithmetic. A price-to-earnings ratio is meaningless on reported earnings; an enterprise-value-to-EBITDA ratio treats a $190 billion private-company mark as though it were a factory. We therefore run four methods and take the median.

2027E operating income of $133 billion is our estimate, built from reported first-half results, the third-quarter guidance range of $22.5 to $26.5 billion and a fourth-quarter estimate of $31 billion, grown 28% into 2027. 2027E EBITDA adds estimated depreciation and amortisation of $112 billion — the most sensitive input in the table. Net debt is $149.6 billion of long-term debt less approximately $90 billion of cash and marketable securities.

The sum of the parts is worth reading line by line, because it shows where the disagreement with the market lives. We value AWS at 25 times 2027 estimated EBIT — a premium to the group multiple, justified by 37% growth and a 39% margin, and already generous. We value the retail and advertising businesses at 13 times. We credit the Anthropic position at 75 cents on the dollar, to allow for tax on an eventual realisation and for the illiquidity of a Level 3 asset. We credit the OpenAI stake at cost, because the second tranche is contingent. And we deduct $60 billion of net debt. The result is $271 a share; the three other methods cluster between $251 and $287; the median is $275.

21 — What $262 requiresTwenty-two times, which is what the market already pays

On 2027 estimated operating income of $133 billion and 10.903 billion diluted shares, each turn of the exit multiple is worth about $12 a share. At $262.43 the market is paying approximately 22 times that estimate — give or take a turn, where Amazon has traded for most of the last five years. This is the finding that makes the shares a hold rather than a buy. The market is not paying a bubble multiple; it is paying the multiple it has always paid, on an estimate that already assumes 28% operating-income growth in 2027. That assumption is not unreasonable, but it is in the price.

22 — The bull caseSeven things that are genuinely going right

Taken together that is a company with three independent engines — cloud, advertising and silicon — each growing faster than 25%, plus the most valuable private-equity portfolio in corporate America, at 35 times normalised earnings. Our disagreement with consensus is not that Amazon is a bad business. It is that consensus has already been paid for it, and then asked for the Anthropic stake on top.

23 — The bear caseSix things, none of them dramatic

The strongest version of the bear case is not that any of these is fatal. It is that they are correlated: a slowing AWS, a compressing multiple and a marked-down Anthropic position all flow from the same source, the market's willingness to pay for artificial-intelligence capacity in advance of the cash returns.

24 — Risk matrixTwelve risks, ranked by probability and impact

25 — CatalystsWhat would change our mind, and when

26 — ConclusionA good company at a full price

Amazon in October 2026 is the best-operating version of itself in five years. AWS is growing at its fastest rate in eighteen quarters, with a backlog nearly three times its run rate and the best margin in the industry. Advertising is compounding at 26% inside a business with the economics of a toll booth, and the retail segments are contributors rather than drags.

None of that is enough to make the shares cheap. The company is spending $220 billion this year to earn a return that will not be visible for two of them, its free cash flow is negative, and a third of its book equity is a mark on a private company rather than a fact about a business.

Our estimate of fair value is $275 a share, the median of four methods and a 4.8% return from the 9 October close — 17% below the street's $330. The entire difference comes from two decisions: we credit the Anthropic position at 75 cents on the dollar rather than at face value, because a Level 3 estimate in a private company whose largest supplier is also its largest shareholder is not the same asset as cash; and we do not apply an artificial-intelligence premium multiple to the operating business at the same time as we credit that stake.

What we would not do is capitalise a valuation adjustment, apply a growth multiple to revenue partly funded by the company's own balance sheet, and assume the highest-margin profit pool survives a federal pricing case — all three at once, which is what the current price requires.

We rate the shares Neutral with a balanced skew. We would become buyers below $230, where the market would be paying nothing for the Anthropic stake, and sellers above $340, where the price would require the bull case to be delivered rather than merely believed.

27 — AppendixFinancial summary

FY2023 and FY2024 figures are as reported in Amazon's Form 10-K for 2024; FY2025 is as reported in the Form 10-K for 2025 and the fourth-quarter release of 5 February 2026. H1 2026 covers the six months to 30 June 2026. FY2026E net sales applies the third-quarter guidance midpoint of $199.5 billion and a fourth-quarter Farstar estimate of $237 billion to the $382.1 billion already reported; FY2026E operating income applies the third-quarter guidance midpoint of $24.5 billion and a fourth-quarter estimate of $31 billion to the $48.5 billion of the first half. "n/d" means not disclosed on a comparable basis. Normalised EPS is a Farstar estimate that removes the whole of other income and applies a 22.5% tax rate; it is not a company measure.

28 — SourcesPrimary and secondary references

Primary — company filings, counterparties and regulators

Amazon.com, Inc. Form 10-K for the fiscal year ended 31 December 2025, filed 6 February 2026; Form 10-Q for the quarter ended 30 June 2026, filed 31 July 2026; Form 8-K on the OpenAI strategic partnership and investment; and the earnings releases and call remarks of Andy Jassy for the fourth quarter 2025, the first quarter 2026 and the second quarter 2026. Anthropic, "Anthropic and Amazon expand collaboration for up to 5 gigawatts of new compute", 20 April 2026. OpenAI and Amazon joint announcement, 27 February 2026. Federal Trade Commission press release and complaint, "FTC and States Sue Amazon Over Secret Ad Surcharge Scheme", 31 August 2026, and Amazon's response.

Secondary — market and industry data

Third-party market data aggregators for the share price, market capitalisation, diluted share count, analyst price targets and peer valuation ranges, as at 9 to 11 October 2026. Industry memory-price datasets on the 2026 DRAM and NAND shortage. Reported results of Microsoft Corporation and Alphabet Inc. for the comparable quarters.

29 — DisclosureIndependence and method