Amazon

IThe business

Amazon sells goods and services to shoppers, advertising to the sellers and brands who want those shoppers, and computing to companies. In the twelve months to September 2026 net sales were $797.5B, up 15%, and operating income was $104.4B, a margin of 13.1%.³⁹ From 2016 to 2025 net sales grew 20.3% a year and operating income 38.8% a year.²¹¹

The three reporting segments are three different businesses. AWS, 20% of net sales, earned a 37.8% operating margin and 58% of the company’s operating income. North America earned 8.2% and International 3.3%.³⁹ I call the two retail segments together Stores. They hold the online and physical stores, third-party seller services, advertising and Prime. Amazon does not report the profit on any of those separately, and I do not invent it.²

How it makes money

• Online stores. $286.1B in the twelve months, up 2% in Q3 against a Q3 last year that held Prime Day. Amazon records these sales gross.R (reported)³⁹
• Third-party seller services. $187.7B, up 9% in Q3: commissions, fulfillment and shipping fees. Sellers account for 61% of paid units, which grew 12%.R (reported)³⁹
• AWS. $161.4B, up 39% in Q3 to a $184B annual run rate.R (reported)³⁹
• Advertising. $78.0B, up 11% in Q3 (12% excluding exchange rates), from sponsored ads, display and video. Its profit is not reported.R (reported)³⁹
• Subscriptions. $54.3B, up 11% in Q3: Prime fees and digital media. Amazon does not disclose how many members pay.R (reported)³⁹
• Physical stores and other. $23.1B for physical stores, up 2% in Q3, and $6.8B for the rest, which includes shipping services, healthcare and the co-branded card.R (reported)³⁹⁴⁰

Prime Day fell in Q2 this year and in Q3 last year, which lifted the Q2 retail growth rates and held Q3 down.³⁹ Amazon says Q3 growth would have been about four points higher without the shift.³⁹

What changed after 2023

Amazon is spending several times what it spent three years ago. Net capital spending went from $48.1B in 2023 to $210.7B in the twelve months to September 2026, 26.4% of net sales.C (calculated)²³⁹ AWS net property and equipment went from $110.7B at the end of 2024 to $313.0B at 30 September 2026.C (calculated)²⁴⁰ Contracted revenue not yet recognized rose from $177B at the end of 2024 to $578B.⁷⁴⁰ Amazon held its guide for 2026 cash capital spending at about $220B, up from about $200B in the April shareholder letter, and the 10-Q says it expects cash capital expenditures to meaningfully increase in 2027.¹⁵³⁹⁴⁰

AWS earns most of the profit, and the depreciation on what it will build sets the value.

IIThe moat

I think Amazon has two moats of different types. The first is scale, in cloud infrastructure and in delivery. The second is a network effect in a marketplace that sellers and shoppers both use. The Stores margin has widened, but I have no share data to show that the Stores moat has. The AWS moat holds in profit and is narrowing in share.

Sellers account for 61% of paid units, and the fees they pay for commissions, fulfillment and shipping reached $187.7B in the twelve months.³⁹ Advertising sells the shoppers’ attention back to those sellers and to brands: $78.0B, growing 11% in Q3 after 26% in Q2, both quarters bent by the Prime Day shift.³⁹⁴ The two retail segments together earned 0.9% of sales in 2016, 2.2% in 2019 and 5.8% in 2025, and 6.8% in the twelve months to September 2026, with a loss of 2.4% in 2022 in between.C (calculated)²⁹¹¹³⁹ The margin rose as advertising grew from $12.6B in 2019 to $78.0B. I cannot tell how much of the gain is mix and how much is power over sellers, because Amazon does not report what advertising earns.²⁹³⁹

Shipping costs are the exposure: they rose 11% to $28.3B in Q3 on net sales up 12%, after rising 19% in Q2 on net sales up 20%.³⁹⁴

The AWS moat is scale plus a cost position in chips. Synergy puts Amazon at 28% of cloud infrastructure services in the June 2026 quarter, ahead of Microsoft at 20% and Google at 15%, but Amazon held 30% a year earlier and Google 13%; Synergy’s September-quarter count was not out when I updated this page.²⁸²⁹ AWS still earns the most: $61.0B of operating income on $161.4B of sales over the twelve months, a 37.8% margin, up from 27.1% in 2023.C (calculated)³⁹² Its chips business, Graviton, Trainium and Nitro, passed a $25B annual run rate in the Q2 report, and 98% of the top 1,000 EC2 customers use Graviton.⁴¹⁵ Jassy’s letter says that at scale he expects Trainium to save “tens of billions of capex dollars per year”; I cannot check that.¹⁵

Customers have committed to AWS for years ahead.

The 10-Q names two expansions: OpenAI’s $100B over eight years, on top of $38B, announced in Q1 2026, and Anthropic’s more than $100B over ten, announced in Q2.⁴⁰ Amazon holds stakes in both companies. It has paid $18.0B in cash for Anthropic, $8.0B of convertible notes and $10.0B of preferred stock, and carries the stake at $238.7B; it has funded the whole $50B commitment to OpenAI, $21.3B of it in Q3.C (calculated)⁴⁰¹⁶ Amazon has also agreed to make up to $15.0B more available for Anthropic to draw as AWS reaches compute delivery milestones, as convertible notes or, after an IPO, common stock bought for cash.⁴⁰ The 10-Q does not say whether the expansions sit inside the $578B, or give any customer’s share. If they do, two customers that Amazon funds hold about 35% of the backlog, and 60% of the $334B added since December 2025 if both were booked when announced.C (calculated) Amazon books their commitments as backlog and its stakes at funding-round marks, so the marks lift net income without adding cash, and the backlog is only as good as the customers’ own funding.⁴⁰ The backlog also covers a minority of near-term sales: spread evenly over its 6.1-year weighted life, $578B is about $95B a year, against $161B of AWS sales in the twelve months and $226B in my base’s first year.C (calculated)

The returns show what the build costs. Each dollar of property AWS added from 2018 to 2021 earned about 36 cents a year after tax. From 2022 to 2025 it earned 14 cents, and from 2023 to September 2026, 12 cents.C (calculated)^2378940 Capacity under construction earns nothing yet, so the latest windows understate what finished capacity will earn. My cost of capital is 10.5%. AWS sales per dollar of net property fell from $1.25 at the end of 2023 to $0.52 in the twelve months to September 2026.C (calculated)^2344039

Amazon’s own account of the economics is payback:

on average, it takes a little less than three years to break even

I turn that into a number. At my 57% margin before depreciation, payback in 2.7 to 3.0 years means $0.58 to $0.65 of annual sales per $1 of server cost, or $0.35 to $0.39 per $1 of all AWS capital if servers are 60% of it and the data center shells earn nothing on their own.E (estimate) My model counts capital net of the year’s depreciation: my base starts at $0.45 and reaches $0.70 per $1 of net new capital. On the gross basis of the payback arithmetic, my base’s first year brings $0.35 of added AWS sales per $1 spent.E (estimate)¹⁵³²

The Stores margin has risen from 0.9% to 6.8% in ten years; the AWS margin now has to carry the depreciation on the $313B it holds and on the $1,021B my base adds.

IIIThe competitive advantage

Amazon’s advantage over Alphabet is profit at scale. AWS earned $17.8B of operating income in the September quarter on $46.0B of sales, a 38.7% margin, or 38.5% without a $0.09B energy-contract gain. Google Cloud earned $11.1B on $30.7B, 36.2%, and AWS’s sales are 1.5 times Google Cloud’s.C (calculated)³⁹⁴⁶ Against Microsoft I cannot measure it. Microsoft does not report Azure’s profit, and from the September quarter it no longer reports Intelligent Cloud either; Azure now sits inside a broader Agents and Infra segment.⁴⁷ Stores earned $43.3B of operating income in the twelve months, 42% of the company’s total, and helps pay for the AWS build.C (calculated)³⁹

Where Amazon leads

AWS keeps 39 cents of each sales dollar as operating income, 38 cents without the energy gain, and Google Cloud keeps 36 cents.C (calculated)³⁹⁴⁶ In chips, OpenAI agreed in February 2026 to consume 2 gigawatts of Trainium capacity, and Anthropic’s expanded commitment includes obligations tied to AWS chips.¹⁶³ Advertising grew 11% in Q3 and Amazon expects growth in the high teens in Q4; its profit is not reported.³⁹⁴⁰

Where it trails

AWS grew 39% in the September quarter, Azure 48% and Google Cloud 103%. Google’s figure includes Workspace and sales of TPU systems, the chips Alphabet designs, so it overstates the gap in cloud infrastructure, but the shares tell the same story as of the June quarter, the latest Synergy count: Amazon’s moved from 30% to 28% and Google’s from 13% to 15%.^3946472829 The 10-Q notes that pricing changes on long-term customer contracts partly offset AWS usage growth.⁴⁰

Amazon is the only one of the three companies with negative trailing free cash flow, and it has covered the gap with debt: $130.0B of long-term debt raised in the twelve months, $48.1B of it in Q3, which took long-term debt from $65.6B at the end of 2025 to $173.1B. The $17.5B term loan was drawn in September, and the sterling notes closed that month. Alphabet sold $49.6B of common and mandatory convertible preferred stock in June; Microsoft’s free cash flow was $60.0B.^394018193047

Against retailers

Against other retailers Amazon earns a higher margin at a similar scale, on different mixes. Stores earned 6.8% on $636.1B in the twelve months to September 2026, with advertising and subscriptions inside it. Walmart earned 4.2% on $713.2B in the year to 31 January 2026, as a company-wide figure.C (calculated)³⁹³¹ The periods differ. The 10-K lists physical, e-commerce and omnichannel retailers, search engines and virtual assistants, logistics providers, and cloud and AI infrastructure companies among its competitors.² Amazon does not disclose its share of online retail and I have no sourced estimate, so I do not judge share in retail.

IVManagement

Andy Jassy has been chief executive since July 2021. He founded and led AWS from 2006, as its chief executive from April 2016, and joined Amazon in 1997.²¹³ Jeff Bezos founded the company in 1994, was chief executive until July 2021, and is Executive Chair. He held 8.8% of the shares in February 2026, worth $231.7B at the close.C (calculated)²¹³ Brian Olsavsky has been chief financial officer since June 2015, Matt Garman has run AWS since June 2024, and Doug Herrington has run Worldwide Stores since July 2022.² Under Jassy net sales rose from $469.8B in 2021 to $797.5B in the twelve months to September 2026, operating income from $24.9B to $104.4B, and AWS operating income from $18.5B to $61.0B.⁸³⁹ The twelve months carry charges and benefits that net to $1.1B of charges, described under Financials.

The stock returned 39%, 6.4% a year, from 2 July 2021, the last close before he became chief executive, to Q4 2026, against 91%, 12.9% a year, for the S&P 500 including dividends. Over ten years it returned 528%, 20.2% a year, against 328%, 15.7% a year.C (calculated)⁴³ The stock trailed the index over his tenure and beat it over ten years.

The board is led by Bezos as Executive Chair and Jamie Gorelick as lead independent director. Nine of the eleven 2026 nominees are independent; the other two are Bezos and Jassy.C (calculated) A shareholder proposal for a mandatory independent chair was on the 2026 ballot and the board opposed it. Kevin Mandia, a cybersecurity executive, joined the board on 8 September 2026.¹³²⁰

Bezos’s holding fell 7.0% between the 2025 and 2026 proxies, by 71.3 million shares. His Forms 4 show he sold 25.0 million shares from 27 June to 24 July 2025 for about $5.65B, under a trading plan adopted on 4 March 2025, and gave 2.5 million shares to non-profit organizations. The proxies show 43.8 million fewer shares over which he has voting power and no investment power. The three add to the full 71.3 million.C (calculated)¹³¹⁴³⁸ He sold a further 1.2 million shares on 3 August 2026 at about $286.³⁸

Capital allocation

The deal record is mixed, and the largest gain so far is an investment, at Amazon’s own marks.

On Amazon’s own marks, $18.0B of cash bought a stake worth $238.7B. Marks are not cash: in 2021 Amazon booked an $11.8B gain on Rivian and in 2022 a $12.7B loss.⁴⁰⁹⁸ The recorded value of the Anthropic, OpenAI and other stakes, $298.8B, is 11.3% of the market value.C (calculated)⁴⁰

From 2016 to 2025 Amazon produced $604.3B of operating cash flow, spent $443.7B net on property and equipment and $126.6B on stock pay, and paid $6.0B for its own shares.C (calculated)²⁸¹¹ It has paid no dividend. The share count rose 12.5%.C (calculated) The one buyback, in 2022, took 46.2 million shares for $6.0B, about $130 a share; the stock is 88% above that price. $6.1B of the authorization remains and nothing has been bought since.C (calculated)⁸⁴⁰ Long-term debt rose from $65.6B at the end of 2025 to $173.1B at 30 September 2026, and interest expense was $1.7B in Q3 against $0.5B a year earlier.²⁴⁰ Amazon drew the $17.5B term loan in September.⁴⁰¹⁸

Incentives

Jassy’s pay was $2.07M in 2025, of which the salary was $365,000. There is no cash bonus. His last equity award, in 2021, was $212.7M on that year’s table, and he has had none since. He holds 2.31 million shares, $564.0M at the close, or 272 times his 2025 pay.C (calculated)¹³ Restricted stock vests over time and no executive’s pay has a performance condition, severance or supplemental pension.¹³ That aligns him with the share price and with nothing else. No measure of return on capital or free cash flow appears, in years when Amazon is spending $211B a year net. Shareholders backed 2026 pay with 94% of the votes for or against, against 78% in 2025.C (calculated)³⁵¹³ Pay does follow the stock: compensation actually paid, which marks unvested stock to market, was $13.2M for 2025 and negative $147.7M for 2022. A hundred dollars of Amazon stock at the end of 2020 was worth $142 at the end of 2025, against $197 for the NYSE technology index and $155 for the S&P retail index.¹³ I rate his operating record high. I cannot rate the Anthropic gain: it is unrealized. Nothing in pay or reporting tests the return on the build.

From 2016 to 2025 stock pay was $126.6B and buybacks $6.0B, with no dividend, and the share count still rose 12.5%.C (calculated)²⁸¹¹ Jassy’s stock vests with time, so it pays him for the share price and size, not for cash returned. Bezos, the largest holder at 8.8%, sold about $5.65B of shares in 2025 and 1.2 million more in August 2026.¹³¹⁴³⁸

VFinancials

Operating income rose 25 times from 2016 to the twelve months to September 2026. Owner free cash flow summed to −$29.4B over 2016 to 2025. Both are true because Amazon puts back more than it earns.C (calculated)²³⁹

I define owner free cash flow as operating cash flow less net cash capital spending, less property acquired under finance leases, less stock pay, as on the Microsoft page. Amazon’s own free cash flow leaves out the last two and, since the FY2025 10-K, shows one measure instead of three. Operating leases never count as capital spending: their rent sits in operating cash flow, $27.2B of operating-lease assets were acquired in the twelve months, and another $164.9B of leases are signed but not started.²⁴⁰³⁹⁷

In the twelve months to September 2026 Amazon added net property worth 196% of its after-tax operating profit. In 2023 it was 64%, and in 2019, before the build, 98% with about half of it leased.C (calculated)²³⁹⁹¹⁰

The twelve months to September 2026 are not a clean base. They carry about $2.4B of charges, the settlements, severance and impairments of Q4 2025; the $2.5B FTC settlement and $1.8B of severance of Q3 2025 have left the window. They also carry about $1.3B of benefits: $0.64B of tariff refunds in Q2 and $0.69B of net gains on energy contracts over the first nine months, $0.09B of it in Q3. Net, operating income is $1.1B higher on a clean basis, $105.4B against the $104.4B reported.C (calculated)²⁴⁰³⁹⁶ In Q3, AWS earned a 38.7% operating margin as reported and 38.5% without the energy gain, and a margin before depreciation of 58.7% as reported and 58.5% without it.C (calculated)⁴⁰³⁹

Operating margin rose from 3.1% in 2016 to 11.2% in 2025 and 13.1% in the twelve months to September 2026. Net sales less cost of sales, which is not a gross profit because Amazon reports none, rose from 35.1% to 51.4% of net sales as AWS and advertising grew.C (calculated)²³⁹¹¹ Technology and infrastructure expense, called technology and content before 2023, rose from 11.8% of net sales in 2016 to 15.9% in the twelve months, which is the build in the income statement.C (calculated)²³⁹¹¹

Operating margin has risen with scale, and depreciation is now cutting into it. On the capital in my base case, AWS depreciation rises from 19.3% of AWS sales in the twelve months to September 2026 to 22.3% in the year to September 2027, 29.1% in the year to September 2029 and 31.5% in the year to September 2030.E (estimate) Of the $144.5B of AWS depreciation in the year to September 2030, about $37B is on the property AWS holds today and about $107B on the $1,021B my base adds in four years.E (estimate) To hold its 37.8% margin, AWS would need a margin before depreciation of about 69% in 2030; I assume 57%, below the 58.5% of Q3 without the energy gain.E (estimate)⁴⁰³⁹

Accounting to watch

1. Asset lives. Amazon has changed server lives four times in six years: from three to four years in January 2020, from four to five (and networking from five to six) in January 2022, from five to six in January 2024, and back to five for a subset in January 2025. The 2024 change cut depreciation by $3.2B and added $2.5B to net income, mainly in AWS; the 2025 reversal added $1.4B of depreciation.¹⁰⁹⁷² From January 2025 Amazon also lengthened heavy-equipment lives from ten to thirteen years, which it estimated would add about $0.9B to 2025 operating income, mainly in Stores. In Q4 2024 it booked about $920M of accelerated depreciation on servers it chose to retire early, and expected about $0.6B more in 2025.⁷ Jassy says servers have a useful life of at least five to six years.³² In my model a 4.5-year life for AWS’s short-lived assets lowers the base value to $155 and a 4.0-year life to $150; a 6.5-year life raises it to $175. The short-lived share matters as much: at 50% of AWS capital my base is $188, and at 70% it is $139. My 9-year life for Stores equipment is shorter than the 10 to 13 years Amazon uses.E (estimate)
2. Cash timing. Property bought but not yet paid rose $26.8B in the twelve months, so cash capital spending of $210.7B ran behind $237.4B incurred, and free cash flow on that basis was about −$69.1B.C (calculated)⁴⁰³⁹ Operating cash flow of $168.3B includes $36.3B of deferred income taxes and $13.3B from higher payables. The investment gains do not lift it: net income carries them and the tax on them, and the cash flow statement takes both out. Cash taxes were $6.8B against about $22.0B of tax at 21% on pre-tax income excluding other income.E (estimate) Three things cut them: 100% accelerated depreciation on qualified property and immediate expensing of domestic research costs under the 2025 Tax Act, and IRS Notice 2026-7, which the 10-Q says applies retroactively to 2025 and lowers 2024 and 2025 cash taxes; the 10-Q does not size it. The depreciation deduction reverses if capital spending slows.⁴⁰³⁹³²
3. Investment gains. GAAP net income of $134.8B includes $69.5B of pre-tax other income, 40% of pre-tax income, mostly marks on Anthropic. Without it, earnings per share were about $7.59 against the reported $12.37. The price is 32.1 times the first and 19.7 times the second.E (estimate)³⁹⁴⁰ Another $140.3B of unrealized gain on the Anthropic notes sits in other comprehensive income. The FY2026 consensus mean of $8.00 sits below the $10.42 earned in the first nine months, so it cannot include the gains.^4045397

VIValuation

First, how this goes badly wrong

I asked what would most likely hurt an owner, and worked back. Three things can. The build outruns demand: net capital spending was $210.7B in a year, and at least $238B of the $578B backlog comes from OpenAI and Anthropic, both funded by Amazon.⁴⁰ The new capital earns too little: AWS has added $0.29 of sales per dollar of property since the end of 2023, and depreciation then lands on slower sales.C (calculated)²³⁹ Or the fixed charges stay when the revenue does not: debt and leases come to $610.2B, 5.8 times operating income.C (calculated)⁴⁰³⁹

What an owner can take out is far below what is reported. Net income for the twelve months to September 2026 was $134.8B, with $69.5B of pre-tax income mostly marks on Anthropic. Operating cash flow less net capital spending, finance leases and $19.9B of stock pay left −$66.8B. That is $201.6B apart.C (calculated)³⁹⁴⁰

Amazon does not say how much spending only keeps AWS in place, so I cannot know what it earns standing still. I show a range. If new AWS capital never earns more than today’s $0.45 of sales per dollar, my base is $120, not $166.E (estimate) The lab stakes are $78.1B at cost or $245.8B at Amazon’s mark, and I cannot judge which is closer.⁴⁰

What the price implies

Run my model backwards and change only the growth of AWS. The price then needs AWS sales to grow 28.0% a year for ten years, to about $1.89 trillion in the year to September 2036, against 17.7% and $822B in my base.C (calculated) AWS grew 29.9% a year from 2016 to 2025, from $12.2B. It now starts from $161.4B, 13 times larger, and capital spending has risen from 5% of Amazon’s net sales in 2016 to 26%.C (calculated)²³⁹¹¹

The price needs AWS to grow about 28% a year for ten years, to about $1.89 trillion, on my margins and capital needs.

If AWS kept its 37.8% operating margin every year instead of paying for the depreciation the capital brings, my base would be worth $192, $26 more, and the price would still need AWS to grow 22.6% a year, to about $1.24 trillion.C (calculated)⁴⁰³⁹

What it takes to reach the price

Capital and margins matter as much as growth. Swapping one bull input group at a time into my base adds $55 for capital needs, $40 for margins, $17 for growth and $5 for growth after year 15.E (estimate) Growth adds least because, at my base capital needs, extra growth consumes cash almost as fast as it earns. Taken together, in the order the chart shows, the groups reach my bull case at $321.⁴⁰³⁹

The capital lever rests on one ratio, the AWS sales each new dollar of capital brings, and it is the least evidenced input in the model.

From the end of 2023 to September 2026 AWS added $70.6B of sales on $240.3B of added net property, $0.29 per dollar, and that counts capacity not yet in service.C (calculated)²⁴⁰³⁹⁸ Measured with a one-year lag, as the model does, 2025 sales added per dollar of 2024 net property added were $0.56, against $1.61 to $2.36 a year in 2017 to 2021, before AI data centers.C (calculated)^2397891011 That ratio differs from the $0.52 of total sales per dollar of the whole property stock in the moat section. My base averages $0.57 and is worth about the same as holding $0.56 every year. I expect the ratio to rise from $0.45 for three reasons: data center shells, 40% of AWS capital in my model, are built once and filled by successive generations of servers; Amazon is putting its own chips into the next generation; and the backlog turns on capacity already under construction.E (estimate)¹⁵³² This is the largest single lever between my base and the bull case, $55 a share, and the bear case holds the ratio at $0.30 for three years.

What the price pays for Stores

What the price pays for Stores gives the same answer from the other side. I value AWS at multiples of its twelve-month after-tax operating profit, at a 21% tax rate, and read off what is left for Stores.E (estimate) The enterprise value is the market value plus $90.3B of net debt less $245.8B of stakes after tax, $2,481B.C (calculated)⁴⁰³⁹

At the middle row the price pays 37.2 times Stores’ twelve-month after-tax profit.C (calculated) Twelve-month AWS profit is a poor guide because AWS is growing: on my base year-one profit of $62.0B for AWS and $34.5B for Stores, the enterprise value is 25.7 times the total.C (calculated)⁴⁰³⁹

Bear, base and bull

The model runs fifteen years, the years to September 2027 through September 2041: ten explicit years, then five in which each engine’s growth fades in a straight line to the terminal rate. The table under How the model works lists the choices.

The bear case is an overbuild. The labs’ funding runs short, AWS growth falls from 27% to 3.5%, and price competition takes share. Capital already committed arrives anyway, so new AWS capital brings $0.30 of sales for three years and earns 6% to 12% after tax on AWS net property, below my 10.5% cost of capital. AWS margin sinks to about 18%, and after-tax operating profit is flat near $79B to $86B for five years.E (estimate)⁴⁰¹⁶ Because the labs are the customers that run short, the bear case carries the stakes in them at cost, $78.1B, and not at Amazon’s $245.8B after-tax mark. At the mark the bear is worth $73 and with the stakes at zero $50.E (estimate)

My bull case sits close to consensus profit per share for three years, then rests on my assumptions. Its after-tax operating profit per share is +2.1%, −2.1% and +4.9% against consensus earnings per share for the years to September 2027, 2028 and 2029: $10.20 against $9.99, $12.93 against $13.21, and $16.04 against $15.30, the last resting on three estimates.C (calculated) I calendarized the fiscal-year means; the provider and basis of the consensus are not stated, and I read them hours after the Q3 report, so some may not reflect it. Years 4 to 15 are mine: AWS operating margin rising to 42%, $0.60 rising to $0.95 of new AWS sales per $1 of capital, and growth after year 15 of 3.5%. The capital assumptions alone are worth $55 of the $155 between base and bull, and 60% of the bull’s value comes after year 15.C (calculated) The bull spends less capital in year one, $245B, than the base, $266B, while growing faster, because it assumes more sales per dollar. My profit per share is before interest, which was a net $0.3B of income in the twelve months and a net $0.5B of expense in Q3 alone, and the expense grows as debt rises.⁴⁵³⁹

In the base, AWS earns 13% after tax on its year-end net property in year one, 10% in year four and 14% in year ten, against 15% in the twelve months to September 2026.E (estimate)⁴⁰³⁹ The value after year 15 is 68% of my base, 56% of the bear and 60% of the bull.C (calculated)

The discount rate

Most of the gap between the price and my base is the discount rate. My base earns 11% less than consensus per share in the year to September 2027 and 21% less in the year to September 2029, the rate is 1.5 points above the 9.0% on the Microsoft page, and the margin of safety comes off a base that is already below consensus.C (calculated)⁴⁰⁴⁴⁴⁵

I discount at 10.5%. The capital asset pricing model gives 10.5%: the 4.98% ten-year Treasury plus Damodaran’s 4.20% implied equity risk premium for 1 October 2026, times Amazon’s ten-year monthly beta against the S&P 500 of 1.32.C (calculated) Its five-year beta of 1.46 gives 11.1%.C (calculated) The Microsoft page uses 9.0% at a beta of 0.97 (1.10 over five years), measured over the same 120 months, monthly, against the same index. The beta is the only input that differs. I do not adjust Amazon’s beta to Microsoft’s, and the grid below gives the value at each rate in between.²⁴²⁵²⁶

The 8.8% the price needs implies an equity premium of about 2.9 points at the ten-year beta.C (calculated) Damodaran’s 4.20% was measured against a 5.29% risk-free rate on 1 October, and the ten-year yield was 4.98% in Q4 2026. On his own pair the rate is 10.8% and my base is $155.C (calculated)⁴⁴²⁶

Only the 8.5% row reaches the price; the 9.0% row tops out at $243.C (calculated) One point on the discount rate moves my base by 18% to 24%: $136 at 11.5% and $206 at 9.5%.C (calculated)⁴⁰⁴⁴

How the model works

The 60% for AWS comes from the property note: servers and networking equipment at cost rose $59.3B in 2025, equal to 61% of AWS’s $96.5B of net property additions. Some of that equipment serves Stores, and the change is net of retirements.C (calculated)²

The base reconciles to Amazon’s Q4 guidance and its capital spending language, and Q3 narrowed the stretch. It does not reconcile to Jassy’s projection for AWS in 2027 and 2028, which has its own section below. Year-one operating income of $122B is 16% above the $105.4B of the last four quarters before one-time items. After Q4 guidance of $28.5B to $33.5B it needs about $30.4B a quarter for the next three, against $28.1B in Q3, $28.0B before the energy gain.E (estimate)⁴⁰³⁹ Q4 net sales are the highest of the year, and Prime Day moved out of Q3 this year, which lowered Q3 sales growth by about four points; guidance assumes no remeasurement of energy contracts.²³⁹ AWS sales of $226B in year one need sequential gains of about $4.2B a quarter from Q3’s $46.0B; Q3 itself added $3.8B and Q2 $4.6B.C (calculated)³⁹ AWS depreciation in year one is $50.3B, after $7.3B in Q1, $8.1B in Q2 and $9.2B in Q3, which needs a rise of about $1.3B a quarter, against $1.2B in Q3.E (estimate)⁴⁰ Capital spending incurred in year one is $266B, against a 2026 guide of about $220B in cash, of which $172.2B was spent in the first nine months. That leaves $48B for Q4, against $75.9B in Q3, and implies $218B for the first three quarters of 2027, about $73B a quarter; the 10-Q says 2027 cash capital spending will meaningfully increase.C (calculated)⁴⁰³⁹

My base adds 3.7 points to Stores’ margin before depreciation over ten years. The mix is the support: advertising was 5.1% of Stores’ sales in 2019 and is 12.3% in the twelve months.C (calculated)⁹³⁹ Amazon does not report what advertising earns, so the margin gain is my assumption. The first-year Stores margin in my base is 6.2%, against 6.7% in the first nine months of 2026 and 6.8% in the twelve months. Stores depreciation on the equipment Amazon holds ends in year 5 and steps down in year 6, which lifts Stores’ margin by about two points that year in every case.E (estimate)³⁹

What management projects, against my cases

Amazon’s Q3 release quotes Jassy on AWS: “projected revenue growth rates of over 40% both in 2027 and in 2028.” It also quotes him on AWS and Stores together: “There has never been a trillion dollar revenue business, and we may have two of them in the next 5-10 years.”R (reported)³⁹ On the call he said “we believe” AWS can do both, and that nobody can predict five to 10 years out with certainty.⁴⁸ Neither statement is in the 10-Q, and the release’s guidance section covers Q4 and 2026 capital spending.⁴⁰³⁹

My 17.7% a year sits far below his 40%. The two cover different spans, so the fair comparison is year by year. In calendar 2027 my base has about 39%, just under him. In calendar 2028 it has about 31% against over 40%, which is about $319B of sales against about $345B. Even my bull has about 36% in calendar 2028, though its faster 2027 takes it to about $342B. After that my base fades to 7% by year ten and reaches $822B in the year to September 2036. It passes $1 trillion only in the year to September 2041, about five years past the slow end of his range. A trillion within ten years needs 20.0% a year. My bull gets there in the year to September 2035.E (estimate)³⁹

I hold the base below him as a judgement, and I kept it after Q3. Growth of 40% in 2028 needs AWS to add about $6B to $8B of sales a quarter, against Q3’s $3.8B, and Q3 itself came in under the $4.2B a quarter my lower path needs.E (estimate)³⁹ Two customers that Amazon funds may fill about a third of the backlog, so part of the demand behind the projection rests on Amazon’s own money; Jassy calls the backlog well-diversified, and the 10-Q gives no customer’s share.⁴⁰⁴⁸ AWS’s cloud share fell from 30% to 28% in the year to the June quarter, and AWS sales per $1 of net property have fallen from $1.25 to $0.52.^2829239 The evidence against me is real. Jassy says Amazon has “very good visibility into 2027 and 2028” because the largest capacity blocks are reserved well in advance.⁴⁸ He can see those reservations and I cannot, so my 2028 may prove too low.

If he is right, the value moves much less than the sales, because AWS sales arrive after the capital that builds them. Put his two years into my cases and change nothing else, and the base moves from $166 to $167 and the bull from $321 to $319. The bear falls from $57 to $44, because it buys that growth at $0.30 of sales per $1 of capital; and with 40% through 2028 the bear’s story of the labs running short would have to wait until after 2028.E (estimate) Now give the base his two years and then my bull’s growth, which reaches $1,083B in the year to September 2036, on my base margins and capital. AWS capital over four years rises $268B to $1,289B, and the base is $176 at 10.5%, with the price 39% above it. At the 9.0% I use for Microsoft the same path is $253, and the price is 4% below it.E (estimate) So at 10.5% his sales do not reach the price on my margins and capital, and at 9.0% they do.

At 10.5%, after his two years, the price still needs AWS to grow 25.1% a year for the next eight, which ends near $1.9 trillion, nearly double the trillion he describes.C (calculated) The bull’s $321 needs his sales and more: operating margins of 34% to 42%, and $0.60 rising to $0.95 of sales per $1 of capital. Management gives no AWS margin guide. Olsavsky said on the call that AI’s lower margins and component prices are pressures, and Jassy called the capital this growth needs meaningful.⁴⁸ I have not moved the cases to close this gap. It stands on the page, and 2028 will test it.

Outside the cases

The stakes in Anthropic, OpenAI and others sit outside the cash flows. The base and bull cases carry them at their after-tax mark and the bear case at cost. Other items below also sit outside the cases.

The margin of safety

• Price$243.79Close, Q4 2026 [42]
• Base value$166My estimate E (estimate)
• Margin of safety below$10835% under my base value C (calculated)

I set the margin of safety at 35% below my base value, which gives $108. My rule for a discount to base value is 25% for a business with net cash and no customer above a tenth of revenue, 30% when one cycle or two customers drive results, and 35% when fixed charges such as leases sit ahead of the shares. Amazon meets the second and third clauses, and the higher governs. Two customers that Amazon funds may fill about a third of the backlog, as the moat section shows. Fixed charges are large: debt with interest is $294.3B and leases, including $164.9B signed and not yet started, are $315.9B, together $610.2B, 5.8 times twelve-month operating income. Total contractual obligations are $883.6B.C (calculated)⁴⁰³⁹ The bear case I describe is the one in which those charges stay and the revenue does not. At 30% the line would be $116.

The line sits 10% below $120, the value of my base if new AWS capital never earns more than today’s $0.45 of sales per dollar: it pays for the whole improvement in capital productivity failing to arrive, and a little more.E (estimate) The price is 126% above that line, so the stock would have to fall about 56% to reach it.C (calculated) The lowest close of the past year was $198.79 on 13 February 2026, still 84% above it; the stock is 14.2% below its $284.02 close of 3 August 2026.C (calculated)⁴³⁴¹ The close used here came before Amazon released Q3 after the market closed; the stock traded near $239 in the next session.⁴¹

Where this could be in 5 and 10 years

Each figure is my model's value at that date: the present value, in that year's dollars, of the owner cash that comes after it, at 10.5%. The stakes stay at today's marks and net debt at today's level. I divide by today's 10.813B shares.C (calculated) Owner cash is after-tax operating profit less net new capital. Amazon pays none today, so a negative year means the business needs funding.

Owner cash is per share, summed to September 2036, with the negative early years netted out. Return a year is the year-10 value plus owner cash, against the Q4 2026 price of $243.79, over ten years.C (calculated)

For the base to hold at five years, AWS has to reach $534B of sales by September 2031, from $161B now, with Stores at $981B. AWS grows 40% in the first year and about 18% a year over ten. At ten years AWS is $822B at a 32% operating margin. The base needs $1.6T of capital spending in the first five years, so owner cash is negative until year 5, a total of −$21 a share.E (estimate)

The bear has AWS growing 8% a year to $354B at a 24% margin. The bull has 21% a year to $1,089B at 42%. They also differ on how much AWS revenue each new capital dollar buys, $0.50, $0.70 and $0.95 by year 10. The rate matters as much: my base is $231 at 9% and $124 at 12%. About 72% of the base's year-10 value is the terminal value after 2041.E (estimate)

The most likely way the base is wrong is capital. If a new dollar buys 20% less AWS revenue than I assume, the year-10 value is $343, not $398.C (calculated)

These are estimates of value, not forecasts of the share price. The market price can sit above or below value for years.

Other uses of the same dollar

The ten-year Treasury pays 4.98%.⁴⁴ My base here returns 6.1% a year, about 1.1 points more.E (estimate) The closest comparable on this site is Microsoft: 4.5% on its base, −3.5% on its bear and 9.6% on its bull.⁶⁰ Amazon’s spread is wider, from −4.5% to 13.0%. Its page uses a 9.0% rate, not 10.5%.

How good the business is and how good the price is are two questions. AWS earned a 38.7% margin last quarter. Yet owner free cash flow was −$29.4B over 2016 to 2025, and at $243.79 the price is 47% above my base. The price needs AWS to grow 28% a year for ten years.³⁹²

The case for the price needs several things at once: two years of 40% AWS growth, $0.60 or more of sales per dollar of new capital, and a Stores margin near 12%. My guesses are 50%, 35% and 35%. Multiplied, that is about 6%.E (estimate) They overlap, so the true figure is higher, but not high.

VIIThe case for the price

What most people miss

The backlog outside OpenAI and Anthropic is large. The 10-Q names at least $238B from those two. Take that out of the $578B and up to about $340B is left, if those deals are fully inside it. That is nearly double the whole $177B backlog at the end of 2024.⁴⁰⁷

Jassy said on the Q3 call that short-term capacity on older Trainium chips now costs more than in the first half.⁴⁸ A 6.5-year server life, against my 5.5, adds $9 a share in my model.

Amazon’s chips are further along than headlines say. Revenue commitments for Graviton and Trainium each rose more than 20 times in a year, and most of Bedrock, its AI service, already runs on Trainium. Amazon plans to offer the chips to outside customers from late 2027.⁴⁸

The bull case

In their August letter Bill Ackman and Ryan Israel of Pershing Square make the clearest case. Worry about data center spending, they say, understates Amazon’s resilience. Retail units rose 15% in Q1, the fastest since 2021. Longer term, they believe Amazon can “compound earnings at a 20%-plus rate” on AI, online shopping and wider retail margins.⁴⁹ Pershing’s 13F listed 8.56M shares at 30 June, about 25% fewer than in March.⁵⁰⁵¹

Gavin Baker of Atreides said at the Sohn conference in May that Trainium is the most underrated rival to Nvidia. His reason is the network linking the chips: he said only Nvidia and Trainium have that kind working. I read this in reports, not a transcript.⁵⁷⁵⁸

After Q3, JPMorgan said management can see enough demand to commit to 40%-plus growth for two years.⁵⁹

Altimeter listed 2.43M shares, ARK 1.59M and Atreides 0.97M at 30 June.⁵²⁵⁴⁵³ I found no 2026 argument from Altimeter or ARK. Berkshire listed 2.28M at the end of 2025 and none at 30 June.⁵⁵⁵⁶

Against reality

The 28% a year on this page changes AWS growth and leaves everything else at my base. A bull has a shorter route. Give AWS Jassy’s two years of 40%, then my base fade³⁹: 18.3% a year for ten years, to $867B. Add my bull capital, $0.60 rising to $0.95 of sales per new dollar. Add my bull Stores margin, about 12% by year ten. That path is worth about $247 at 10.5%, with AWS margin bottoming near 31%, or $231 with the stakes at cost.E (estimate) If Amazon hits those three targets, the price is possible.

Today’s numbers fall short on each. AWS grew 39%, and 40% needs about $50B in Q4, $4.2B more than Q3. Stores earns 6.8%. The widest gap is sales per new dollar: $0.29 since 2023, $0.56 with a one-year lag. The path allows about $245B of capital in its first year, about $61B a quarter, less than my base’s $266B. Q3 was $76.8B, counting unpaid bills, and the last twelve months $237.4B.³⁹⁴⁰ The 2026 guide leaves $48B for Q4, so the path fits about $260B in 2027. The 10-Q says only that 2027 rises meaningfully.⁴⁰ One year at Q3’s pace costs about $10 a share. Stuck at $0.60, the path is $216. On my base capital, $186.E (estimate) My base compounds profit per share about 17% a year for three years, below Pershing’s 20%-plus. This path does 25%.

What they have to hit

Thresholds come from this path; stricter signposts are noted. Filings are free on SEC EDGAR (opens in a new tab). February dates are expected.²

Where I come down

The chip evidence moved me most, because it could lift sales per dollar of capital, the $55 lever between my base and bull. The spending did not move me: no number Amazon reports yet shows 40% growth on less capital than my base. Two years at $0.60 or better, with spending near the $260B this path allows, would raise my values. Until then they stay at $57, $166 and $321.

VIIIRisks, and what would change my view

The risk I weight most is the one management states itself: demand for AWS capacity is ahead of supply, and the build to meet it drains cash. Jassy said on the Q2 call that even at $220B of spending Amazon “will still not have enough capacity” for 2026’s demand, and the filings cite energy shortages and volatility in resources, including memory chips, as limits.³²⁴⁰

1. The build and its funding. Net capital spending was $210.7B in the twelve months against $168.3B of operating cash flow, and the 2026 guide rose by $20B in Q2 on the cost of memory and held at about $220B in Q3. Leases not yet started ($164.9B) and purchase obligations ($254.4B, $159.6B of it for property and equipment) come on top, and total contractual obligations are $883.6B.³⁹⁴⁰³² Jassy expects free cash flow headwinds until the data centers under construction are monetized.³² In my base, free cash flow is negative for four years and the cumulative outflow reaches about $252B in the year to September 2030, financed with debt.E (estimate)
2. Two customers that Amazon funds. The moat section sets out the numbers. A commitment is only as good as the customer’s own funding. Rivian shows what a mark can do.⁹⁸
3. Depreciation and margins. AWS depreciation was $9.2B in Q3 against $5.6B a year earlier, and my base takes it to 31.5% of AWS sales by September 2030. A faster obsolescence cycle in AI hardware would raise it further.E (estimate)⁴⁰²
4. Share and price in cloud. Amazon’s share of cloud infrastructure fell from 30% to 28% in a year while Google’s rose from 13% to 15%.²⁸²⁹
5. Regulation and the rest. Private plaintiffs, state attorneys general and the FTC have brought antitrust and consumer cases in the United States, Canada and the United Kingdom. In August 2026 the FTC and 22 state attorneys general sued over reserve prices in Amazon’s sponsored-ad auctions, and Texas filed its own case. The FTC settlement cost $2.5B in Q3 2025, EU fines include EUR 746M under appeal in Luxembourg and EUR 752M in Italy, and Kove IO won $525M plus $148M of interest against AWS, on appeal. The 10-Q says China-based sellers account for significant portions of third-party seller services and advertising revenue.⁴⁰²

What would move me

What Q3 showed against the tests

Amazon reported Q3 on 29 October, after the close, and filed its 10-Q on 30 October.⁴⁰ Net sales were $201.9B against a guide of $197.0B to $202.0B, at the top of the range. Operating income was $28.1B against a guide of $22.5B to $26.5B, $1.6B above the top, up from $17.4B a year earlier; the margin was 13.9%, the highest Amazon has reported. AWS sales were $46.0B, up 39%, and AWS operating income was $17.8B, up from $11.4B.³⁹⁴

Against my base, Q3 landed close. AWS added $3.8B of sales from Q2 where my base needs about $4.2B a quarter, so the sales path is a little behind the base and the margin is well ahead: AWS earned 38.7% against 34.7% in my first year. AWS depreciation rose $1.2B from Q2 against about $1.3B needed, on pace. The backlog rose $82B to $578B with a 6.1-year life, short of the $600B signpost but past its 6-year life test. AWS grew 39% with the capital spending guide held at about $220B, so the bear signpost of a raised guide with growth below 30% was not met, and the bull signpost of a flat 2027 was not met either: the 10-Q says 2027 cash capital spending will meaningfully increase. Stores earned 6.8% over the twelve months, between my 6.2% and 7.4% signposts. Synergy’s cloud share for the September quarter was not out, and the recorded value of the Anthropic, OpenAI and other stakes rose by $69.1B to $298.8B: $21.3B of cash into OpenAI and a $48.3B mark-up of the Anthropic notes through other comprehensive income, less about $0.5B net on other holdings, with no Q3 adjustment to the preferred stock reported.³⁹⁴⁰

Q3 did not change my assumptions, and it did bring a projection my base does not match: AWS growth of over 40% in 2027 and in 2028, against about 39% and about 31% for those calendar years in my base.E (estimate)³⁹ It moved the starting point forward a quarter, raised the stakes and the net debt together, and left the verdict where it was: the price needs AWS to grow 28.0% a year for ten years, and my base is 17.7%. Free cash flow fell to −$42.4B over the twelve months, and net debt rose to $90.3B from $32.3B in the quarter. My base value moved from $163 to $166.