Amazon (AMZN) — Forty Percent, Paid in Advance

AWS grew 39.3% in the third quarter, its sixth straight acceleration, and management projects more than 40% in each of 2027 and 2028 on capacity that customers reserve in advance. The same quarter spent $75.9 billion on capital, trailing free cash flow fell to −$42.4 billion, and 2027 spending will be "meaningfully more" than 2026's $220 billion. We value Amazon at $175.67, from $142 in July, against $243.79 at the close before the release. Holding our other inputs, the price takes $1.77 of revenue per dollar of new plant, against the $1.06 to $1.23 the prior year's plant has earned, a cost of equity about 1.5 points below ours, or revenue growth about 2.0 times our base case; our bull case clears it at 1.2, on faster growth and wider margins.

TON618 Capital Research · As of 29 October 2026 · Download PDF · Download model

Our July note described Amazon as two machines sharing one invoice: a store that keeps about 6 cents of each sales dollar and a utility, AWS, that keeps about 37. In the June filings, AWS's annualized revenue and Amazon's trailing twelve-month capital spending were both $169 billion. Three months later the run rate is $184 billion and the spending is $210.7 billion. Customers are reserving the utility's capacity before it is built, and every added point of growth is paid for in advance with capital. This note asks what the faster utility is worth once the invoice is counted.

Terms used throughout, defined once. AWS: Amazon Web Services, the cloud segment; retail means the North America and International segments together, advertising included. Capex: purchases of property and equipment net of proceeds from sales and incentives, the "cash capex" Amazon guides. Free cash flow: operating cash flow less capex, Amazon's own definition. Ex-gain: a figure with the cash-flow statement's "non-operating expense (income)" line removed at the 21% statutory tax rate; most of that line is marks on the Anthropic investment. Backlog: revenue under contracts of more than a year not yet recognized, primarily AWS (accountants call it remaining performance obligations). Free cash flow to equity (FCFE): net income plus depreciation of property and equipment, less capex, plus net new borrowing; stock compensation stays an expense. SOTP: sum of the parts, here AWS and retail valued separately on 2027 after-tax operating profit at peer multiples. Re-mark: our July estimate of earnings with servers depreciated over three years. Sales-to-capital: added revenue per dollar of added net property and equipment, the measure of how much plant growth requires. ROIC: after-tax operating income over invested capital (equity plus debt less cash), against the weighted average cost of capital (WACC). Terminal growth: the rate the business grows at forever after the forecast ends, set here from long-run GDP projections.

Net sales were $201.9 billion, up 12% (13% excluding currency); the release says the move of Prime Day into the second quarter cost about 400 basis points of growth. North America rose 6% to $112.9 billion, International 5% to $43.1 billion (8% excluding currency) and AWS 39% to $46.0 billion. Advertising grew 11% to $19.6 billion (12% excluding currency), held back, Jassy said, by an 850-basis-point Prime Day effect; the release expects high-teens growth in the fourth quarter. Operating income rose 61% to $28.08 billion, but that rate flatters the quarter: a year earlier "other operating expense" was $2.9 billion, against $0.1 billion now. Without that line operating income rose about 39%. North America earned $8.7 billion against $4.8 billion, International $1.6 billion against $1.2 billion, AWS $17.8 billion against $11.4 billion.

The miss, in two parts. Diluted EPS was $1.89, against $1.95 a year earlier and a consensus of $1.97 in our brokerage feed. The year-on-year fall is the investment line. A year ago the release booked $9.5 billion of pre-tax gains on Anthropic in non-operating income; this quarter "other income (expense)" was a loss of $703 million. Removing the non-operating line at 21%, EPS was $1.94 against $1.22, up 60%. Against the consensus the marks explain about 5 cents of the 8-cent gap, if analysts modeled none; the compilation does not say. Exhibit 1 shows why the reported line misleads in both directions: ex-gain EPS has run between $1.65 and $1.94 for four quarters, while reported EPS ran between $1.89 and $5.75.

AWS revenue was $45.98 billion, up 39.3%, and $3.8 billion more than in the second quarter (Exhibit 2). Its operating margin was 38.7%, against 39.4% in the second quarter and 34.6% a year earlier. Management's projection of growth above 40% in 2027 and 2028 is not guidance, and chief financial officer Brian Olsavsky said Amazon is "not guiding at all on our margin for AWS." The call gave three supports for it. The backlog: $578 billion at 30 September, against $496 billion at 30 June, with a weighted-average remaining life of 6.1 years, from 6.4, in the 10-Q. The reservations: the largest blocks of capacity are reserved well before delivery, so management can see "'27 and '28 and some of '29." The plan: capacity planned "over the next, call it, 8 quarters or so, which is about as far out as we get 8 to 10 quarters."

The backlog is large, but it does not set the growth rate. Spread over its 6.1-year life, $578 billion is about $95 billion of revenue a year, about half of AWS's current run rate. Capacity carries the projection. Revenue in a capacity-constrained cloud is capacity delivered times use, and Jassy said customers take capacity "almost as fast as we install" it. So growth in 2027 and 2028 is mainly a statement about how much capacity Amazon will deliver, and that is the bill in §3.

What we carry. Our base case takes AWS growth at 40% in 2027, inside the eight-to-ten-quarter horizon the plan covers, and 36% in 2028, at its edge, where some of the year rests on reservations not yet made. The bull carries 44% and 42%; the bear 30% and 22%, a case in which capacity arrives late or inference prices fall faster than use grows. July carried 30% in 2027 and then the consolidated growth rate; on July's method the faster path is worth +$18.11 a share. The extra revenue earns a 35% operating margin in the base case, below the quarter's 38.7%, because Olsavsky said the "outsized growth of AI puts margin pressure on the business as a whole" and named component prices as a further headwind; AWS also carries its own depreciation, 20% of its revenue in the first nine months. Retail grows 9% in 2027 and 8% in 2028 in the base case, below its trailing 11.7%, at a 7% operating margin (bull 11% and 10%, bear 6% and 5%). Taking management's projection literally, 40% in both years, moves fair value only to $183.16: in our model the growth arrives with the capital that builds it.

Capex was $75.9 billion in the third quarter, 37.6% of revenue, against $34.2 billion a year earlier. Operating cash flow was $42.4 billion, so the quarter's free cash flow was −$33.5 billion. Over twelve months, operating cash flow rose 29% to $168.3 billion and capex rose 82% to $210.7 billion, so trailing free cash flow fell from −$7.6 billion in June to −$42.4 billion (Exhibit 3). The release attributes the increase "primarily" to "investments in artificial intelligence." Trailing capex per dollar of ex-gain net income rose from 2.34 times in July to 2.62 times: Amazon now spends $2.62 on capital for every dollar it earns.

The 2026 guide leaves an odd fourth quarter. The release repeats "approximately $220 billion" of 2026 cash capex. Nine months have spent $172.2 billion, which leaves $47.8 billion for the fourth quarter, about 37% less than the third. Either spending falls sharply for one quarter or 2026 ends above $220 billion; the call did not say which. Jassy said Amazon expects "to spend meaningfully more in 2027." He gave no figure. We carry $280 billion in the base case, about 27% above 2026 and below the $304 billion annual pace of the third quarter; the bear carries $310 billion, the bull $265 billion, because in the bull more revenue arrives for the same build. On July's method the 2027 figure costs $8.29 a share.

The funding. Long-term debt is $173.1 billion, from $128.9 billion at June and $65.6 billion at the start of the year, after $48.1 billion of new notes in the quarter. Interest expense was $1.69 billion, from $0.54 billion. Net of cash and marketable securities, Amazon now owes $58.9 billion. Property and equipment, net, rose $62.8 billion in a quarter to $508.9 billion. Depreciation of that base is the invoice's second payment: depreciation of property and equipment was $15.1 billion in the quarter, 39% above a year earlier, and AWS's allocated share $9.2 billion. The 10-Q discloses no change in useful lives.

Amazon holds two kinds of Anthropic securities: nonvoting preferred stock, carried at the last observable price adjusted for its rights and a discount for lack of marketability, and convertible notes, carried at fair value with the gains in other comprehensive income. At 30 June the preferred was carried at $92.5 billion and the notes at $97.9 billion. In the third quarter Amazon's unrealized gain on its available-for-sale debt securities rose by $48.2 billion before tax; the 10-Q puts the notes at $146.2 billion and the preferred at $92.5 billion at 30 September. Amazon's cost for all of it is about $18 billion. Beside it sit $50 billion of OpenAI preferred stock and smaller warrants and stakes.

Our July model valued the investment book once, as an asset, at Amazon's carrying value net of the tax the gains would bear. Anthropic filed to list its shares on 26 October, and on 27 October we published our own valuation: $50.72 a share, 14% below the Series H price of $58.90. The 10-Q gives the count for the first time: 164 million preferred shares, and about 184 million more from the notes, each convertible into ten nonvoting common shares, so 3,480 million common-equivalent shares, subject to Amazon's 12% ownership cap on the notes; the S-1's 1,840 million Class N shares for the notes match. That is roughly 21% of Anthropic's 16,427 million economic shares. On that count Amazon's books carry Anthropic at about $68.59 a share (the notes at $79.46, the preferred at $56.40), 35% above our fair value and 16% above the last private round (Exhibit 4).

Which mark the fair value uses, and why. We use our published fair value. Amazon's mark would import a price 35% above that view into this one, and the Series H price is a private round that predates the registration, the S-1's financial statements and our model of them. This changes July's method, at −$4.50 a share. At Amazon's carrying value fair value would be $180.16; at the Series H price, $177.73; with the stake at zero, $162.65. The listing does not change the rule. Amazon expects a customary lock-up, and the 12% cap could leave part of the notes unconverted, so a first-day price is a data point for the next update and not a mark (tripwire 6).

The model. Free cash flow to equity in three scenarios weighted 25/50/25, stock compensation expensed, a depreciation add-back limited to property and equipment, borrowing modeled and its interest charged, and the investment book added once. Every input and method below was set as the most defensible choice on today's evidence. To show what moved, we first re-ran July's configuration, which reproduces its fair value, $141.84 (published as $142): free cash flow to equity $117.70, the re-marked sum of the parts $178.05, blended 60/40. Each factor in Exhibit 5, and in the attribution table later in this section, re-runs that blend with one more change, so the factors sum to the move.

The quarter's data, on July's method: $165.10. Rolling the valuation date to 29 October, with the borrowing already done, adds +$0.55. The investment book at its 30 September carrying value adds +$6.16. Booking nine months of results with the fourth-quarter guide midpoint adds +$0.37: revenue of $827.6 billion and operating income of $110.4 billion for 2026. AWS and retail as their own lines add +$18.11 (§2). The 2027 capex figure takes −$8.29 (§3). Updating the sum of the parts for this quarter, and scaling the depreciation re-mark to the capex now planned, adds +$6.37.

The growth gets its own runway: +$25.23. July faded one consolidated rate and ended in 2032 at 7.0% growth against a 3.5% terminal; a single rate fading in a straight line to terminal growth would have cut AWS from 36% in 2028 to about 13% in 2029. AWS now falls in a straight line from its 2028 rate to terminal growth by 2033, one more forecast year, and retail from its own 2028 rate; margins blend by the resulting mix, so AWS reaches 42% of revenue by 2033.

And then pays for its plant: −$22.22. From 2028, capex replaces what depreciates and adds plant at one dollar of revenue per dollar of new net property and equipment in the base and bear cases, 1.2 in the bull. Amazon's own record brackets that level (Exhibit 6). The 2025–26 readings (0.76 and $0.58) understate it because plant is built before it bills: the revenue added in 2026 (the fourth quarter at the guide midpoint) is $1.06 per dollar of plant added in 2025, and on a trailing basis the plant added in the year to September 2025 earned $1.23 on the following year's revenue. That lagged reading is about to fall: 2026 adds 1.8 times 2025's plant, and the base case's 2027 revenue gain is $0.70 per dollar of it. The 2017–24 readings (1.3 to 4.4) overstate it for what comes next, because they came from retail-led growth, and AWS carries about $1.94 of net plant per dollar of trailing revenue, including plant not yet billing, against retail's $0.29 (10-Q segment property and equipment). The base takes 1.0, just under the lagged $1.06 to $1.23 and above the $0.70 the 2026 build earns in 2027. The bull takes 1.2 for Trainium's cost advantage and a faster payback; the bear keeps 1.0 because it is a demand case, not a capital-efficiency case. Depreciation runs on the rolled-forward plant at a rate that rises from 2026's 13.0% to 16%, close to the 17% Amazon ran in 2020–21, and the terminal takes the plant per dollar of revenue the path actually lands ($0.87 in the base case) instead of July's outside $0.55. That leaves a terminal return on capital of 15.6% in the base, 22.5% in the bull and 9.6% in the bear, about the 9.9% stable cost of equity: a bear in which growth no longer creates value.

Four smaller changes to how the model ends. Terminal growth is set from long-run nominal GDP: the Federal Reserve's projections of 16 September 2026 give 2.0% real growth and 2.0% inflation in the longer run, so 4.0% in the base and bull cases and 3.0% in the bear, against July's 3.5%, 3.5% and 3.0%: +$2.87. The terminal adds the debt that grows with the business, rather than holding it flat forever: +$4.53. It is taxed at 22%, the trailing effective rate, while the forecast years keep 19%, the cash rate while full expensing of new equipment shelters income: −$3.94. The 2026 stub counts only fourth-quarter flows, not a share of a year whose borrowing already sits in cash: −$0.69. With the runway and the plant, these six changes add +$5.77 net.

The cost of equity: 11.3% for the forecast years (−$5.77), 9.9% for the terminal (+$22.87). We argue the rate for the company and show the standard reads beside it. The starting point is the 10-year Treasury, 5.05% on 28 October against 4.67% on 29 July, plus a beta times a 4.5% equity risk premium. For 2027–28 we use a beta of 1.38: Amazon's own 60-month beta against the S&P 500, 1.43, blended slightly toward the 1.09 median of Microsoft, Alphabet and Walmart in proportion to its noise. The estimate is precise (standard error 0.20) and stable (the 36-month beta is 1.46), so the data give no reason to pull it toward 1.0, as the Blume adjustment would (10.9%, worth $178.09), and a build financed partly with new debt argues against a lower read. Relevering the beta for that debt (the Hamada adjustment) would add about 0.2 points by 2033, inside one standard error of the beta, about 0.9 points either way, so we show it ($174.55) rather than add it. The rate goes from July's 10.84% to 11.3%. The terminal value describes a mature company, so we discount it at 9.9%, reached in a straight line over 2029–33 as growth fades. AWS is about 65% of segment value in the sum of the parts on 2027 profit. Weighting the average beta of Microsoft and Alphabet, 1.14, by that share, with retail at the market's 1.0, gives 1.09, within 0.1 point of the 9.9% we use. That rate sits between two bounds: Walmart's beta for all of retail gives 9.3% ($193.70) and Alphabet's beta for the whole company, a cloud-only terminal, gives 10.4% ($163.20). One rate throughout, 11.3% with no stable rate, would give $145.03.

How long each part of the business grows faster than normal. Each stage's length comes from what drives it, not from a default.

Return on capital. Trailing return on invested capital is 16.0% against a 10.85% WACC, and the regime is fading: the spread fell 6.3 points in four quarters, and the return is down from a 27.9% peak in Dec ’24 (Exhibit 7). Operating income rose 37% on the year; average invested capital rose faster, because the build is debt-funded and because the investment marks sit in equity while earning nothing in operating income; without the investment book the return is about 25%. The 2026–27 build has not yet reached the numerator. Which arrives first, its depreciation or its revenue, sets the next turn.

What the price requires. Holding every other base input, the base case reaches $243.79 with costs of equity of 9.8% for the forecast years and 8.4% for the terminal, each about 1.5 points below the rates we use; or with $1.77 of added revenue per dollar of added plant from 2028, against 1.0 in the base case, about the 1.8 Amazon averaged over 2021–24 (2.0 over 2020–24) and well above the $0.58 of the last twelve months; or with revenue growing about 2.0 times faster than the base case in every year to 2032, a compound rate of 24% against 12%. Terminal value is 89% of the base case's operating value (85% in the bull, 89% in the bear), so these long-run inputs carry most of the number. Of the three scenarios, only the bull, $282.02, is above the price.

Our July note set five tests, and our October note on Amazon's $8 billion chip leaseback set four. The print test fired: trailing free cash flow deepened, while AWS beat its comparison. Four are pending to the fourth-quarter call, the 10-K or a 2027 date. The capex concession is not passing, the price test was not tripped, and the leaseback has not appeared in the release, the call or the 10-Q.

Each tripwire has its reading rule set now, so the next update scores it rather than reinterprets it. Effects are on fair value at today's other inputs.

1. Fourth-quarter capex (Q4 release, about 4 February 2027). The guide implies $47.8 billion. At or below that, the build is pacing to the guide and the base 2027 figure moves halfway toward the bull's $265 billion. Between that and $60 billion, the base holds. Above $60 billion, 2026 ended at least $12 billion above the guide and the 2027 build has started early; the base 2027 figure moves halfway toward the bear's $310 billion.
2. The 2027 capex guide (same call). At or below $250 billion, every scenario is re-run at $250 billion, fair value about $184. Between $250 billion and $310 billion, the base holds. At or above $320 billion, about $167. No number on the call: the base holds, and the next update says so.
3. AWS growth in 2027 (FY2027 release, February 2028; waypoints every quarter). The base carries 40%. Below 32% for the year, the bear's pace becomes the base, about $147. At or above 42%, the bull's pace, about $191. The fourth-quarter house figure is $50.0 billion; each quarter's growth and the backlog are recorded against the path.
4. AWS margin through the build (Q4-26 and Q1-27). Both quarters below 33%: the incremental AWS margin moves to the bear's 31% in every scenario, about $162. Both at or above 38.7%: the Q3 margin in every scenario, about $185.
5. The sum-of-the-parts weight (each quarter's 8-K, read from the 2028 releases). The observable is trailing sales-to-capital on the model's same-year definition: the change in twelve-month revenue over the change in net property and equipment across the same four quarters, $0.58 at September (lagged one year, $1.23). It is read from 2028, the year the model's 1.0 applies; until the 2026–27 build bills, the same-year reading is depressed by construction and the lagged one flattered by it. At or above 1.0 for two consecutive quarters, the sum of the parts returns at 20% (about $179.58 today). At or above 1.2 for two consecutive quarters, about the level at which the sum of the parts and the cash-flow model agree, with trailing free cash flow positive, it returns at July's 40% (about $183.48). Either reading also re-sets the base case's sales-to-capital to the observed level; at 1.2 in every scenario fair value is about $192, at 0.8 about $130.
6. Anthropic prices and lists. The stake stays at the house fair value until the house updates it, with the first-day close and Amazon's next carrying value recorded beside it. A change in the house value flows through at about $0.25 a share of Amazon per dollar of Anthropic.
7. Price. A close at or below $175.67 resolves the gap. $132 is a 25% margin of safety to fair value. Replaces July's $107.

We value Amazon at $175.67 a share, from $142. The utility ran faster than July assumed, and that alone would have lifted value by about $23. The invoice ran further: trailing free cash flow is −$42.4 billion, the company is borrowing to build, and 2027 will cost more. A model in which the plant grows with the revenue it serves gives the faster AWS real value, but not the price. Holding our other inputs, the price takes $1.77 of revenue per dollar of new plant, a capital efficiency Amazon last reached in 2023 (its 2022–25 aggregate is 1.3), costs of equity about 1.5 points below ours, or revenue growth about 2.0 times our base case; our bull case clears the price at 1.2, on faster growth and wider margins. The fourth-quarter call, with the 2027 capex number, is the next test. The Fund holds no position in AMZN.