Amazon.com
An online shop with its own delivery network, bolted to AWS, which rents computing power to other companies and earns most of the profit.
The scores
Growth4/4 ↑
Revenue up 20% and operating income up 43%, with the cloud line accelerating for five straight quarters from 20% to 36.7%.
Business quality4/4 →
A competition regulator measured the switching rate below 1% a year, and the cloud line earns a 39.4% segment margin at scale.
Financial position3/4 ↓
Long-term debt doubled in six months and a $17.5 billion term loan was signed, against $1.1 trillion of assets and $123 billion of liquidity.
Execution record4/4 →
Five consecutive quarters of cloud acceleration delivered, both one-off benefits disclosed unprompted, and the capex raise explained by component cost.
Per share capture3/4 →
Shares outstanding rose from 10,731 million to 10,783 million in six months, and stock compensation ran $6,038 million in the quarter.
Potential4/4 ↑
Two counterparties signed more than $200 billion over eight and ten years, and the IEA has the physical market doubling by 2030.
Valuation2/4 →
32 times trailing operating profit with free cash flow negative - the cheap-looking earnings figure is a paper gain, not the business.
Evidence quality4/4 →
The moat comes from a competition regulator, the peer comparisons from three competitors' own filings, and the market size from the IEA.
Risk profile2/4 ↓
$122.3 billion of Level 3 private stakes on which the company declines to publish a sensitivity, funded alongside a debt-financed build.
If it works 1.4× · if it fails you keep about 93%.
The Verdict
We rate Amazon neutral. It runs an online shop and AWS, the cloud arm that earns most of the profit and is growing faster again, so the business is strong. Its earnings look cheap only because of a paper gain on Anthropic; on operating profit buyers pay more than for any rival, while the company spends more cash than it makes. So the price is full, not cheap. As a bet it is defensive: somewhat more than today if AWS keeps its pace, most of the money kept if it slows. The business is as good as the market says; the price already knows it.
The long caseCompanies and AI labs would rather rent computing than build their own, and customers rarely switch clouds, so AWS can keep selling more to the ones it already has.
The evidence under this - 9 claims
The cohort spans four cloud providers and a grocer, and the two cheapest members - Microsoft and Oracle - carry large software annuities at margins Amazon's retail half can never reach. A multiple struck on a company two-fifths of whose profit comes from an 8% margin business is not comparable to one struck on a 47% margin business, and the distance is well short of the 1.5 times the median this repository requires before a score may move.
The same arithmetic supports the bull case. If AWS holds anything near 36.7% growth at anything near a 39% segment margin, its operating profit alone compounds past the whole company's current profit within about three years, and the multiple on that basis is unremarkable. The claim is that the price requires the growth, not that the growth is unlikely.
The company
Amazon sells goods online, runs a marketplace and delivery network for other sellers, and sells advertising to them against the shoppers it draws.
AWS is the other business. It rents computing, storage and databases, and increasingly AI, to companies and AI laboratories, which pay for what they use instead of owning servers and the staff to run them. It even designs its own chips to do it.
The evidence under this - 5 claims
Nothing outside Amazon publishes a price-performance comparison of Graviton against an x86 instance, and the 98% figure is a count of customers touching the chip rather than a share of their spend. Amazon also states it continues a deep partnership with Nvidia and that customers will run on Nvidia 'for as long as we can foresee', which is a concession that the custom silicon has not displaced the merchant part.
Neither figure is a filed one. Amazon reports three segments and no product-level revenue inside AWS, so nothing in the 10-Q corroborates either number, and a run rate is an annualised instant rather than a period. Management raised the chips figure from 'topped a $20 billion revenue run rate' one quarter earlier, so the series is at least internally consistent.
How it makes money
The shop earns on the goods it sells, the fees sellers pay and the advertising, which brought in $19.8B in the June quarter, up 26%. AWS bills for what customers use, like a meter, and more of it is now signed years ahead: $496B of commitments not yet billed, with 6.4 years left on average.
The evidence under this - 4 claims
Where the money comes from, stream by stream
| Stream | Share of sales (FY2025) | How it earns it |
|---|---|---|
| Online stores | 37.6% | Goods Amazon buys and resells online, and digital media sold one item at a time, booked at the full sale price. |
| Third-party seller services | 24.0% | Commissions and fulfilment and shipping fees charged to independent sellers on Amazon's marketplace. |
| AWS | 18.0% | Metered cloud computing, storage, database and AI capacity, paid for as it is consumed or under multi-year commitments. |
| Advertising services | 9.6% | Sponsored listings, display and video ads sold to sellers, vendors and others, reported inside the two retail segments. |
| Subscription services | 6.9% | Prime membership fees plus video, music, audiobook and e-book subscriptions. |
| Physical stores | 3.1% | Goods customers pick up in Whole Foods and other Amazon stores. |
| Other | 0.8% | Shipping services, healthcare, video licensing and the co-branded credit card. |
Who runs it
Andy Jassy, who built Amazon's cloud business, has run the company since 2021; founder Jeff Bezos chairs it and still owns about 9%, nearly all the insiders' stake. Bosses get a small salary, no bonus and shares that vest over five or more years, so their pay rides only on the stock price, not on whether the heavy spending earns a return.
The evidence under this - 3 claims
The 8.8% includes 68,232,131 shares over which Bezos has voting but no investment power, so his economic stake is smaller than the headline; and the Vanguard and BlackRock figures are from Schedule 13G filings as of December 2023, not February 2026.
Because nothing pays on a financial metric, nothing in the plan rewards return on the $220 billion build or free cash flow: executives gain only through the share price, which rewards the build whether or not it earns its cost of capital.
Why it wins, and who could beat it
Why customers come back
What keeps a cloud customer is where its data already sits. Moving one system means moving everything wired to it, and paying to take the data out. Management says AI work follows the data, and more of it lives in AWS than anywhere else.
A UK competition regulator measured it: under 1% of cloud customers switch provider in a year. It also found AWS losing share to Microsoft since 2020, so the lock-in keeps customers without stopping rivals winning new ones.
The evidence under this - 3 claims
This is the incumbent describing its own advantage, and the two rivals grew their cloud lines faster in the same period - Google Cloud at 81.8% and Oracle's cloud infrastructure line at 84.3% against AWS's 36.7%. If data gravity were decisive at the margin, the challengers could not be compounding at twice the rate on bases that are no longer small.
A low gross switching rate is not the same as a high net one. The same investigation found that egress fees are 'not typically the most important factor in customer decisions relating to switching', and both AWS and Microsoft agreed in March 2026 to remove egress fees for UK customers switching, for a period of at least 180 days - so the most cited barrier is being dismantled by agreement.
The geographic market is UK plus the EEA rather than global, the shares are bands the regulator redacted rather than point figures, and 2024 pre-dates the whole AI acceleration - AWS grew 20% in 2025 and 36.7% in the June 2026 quarter. A share series ending in 2024 cannot describe the period this thesis is about.
The threat
Microsoft and Google rent the same computing, and both are growing faster: Google's cloud grew 81.8% in the June quarter and Microsoft's Azure 40% in March, against AWS's 36.7%. Google's committed cloud orders, $513.9B, exceed all of Amazon's, and turn into sales more than twice as fast.
The evidence under this - 4 claims
Alphabet states it changed its backlog definition in Q1 2026 to include contracts with an original expected term of one year or less, which Amazon's disclosure excludes. The two are therefore not struck on identical scopes, and part of Alphabet's lead is definitional rather than commercial.
A longer life is not unambiguously worse - it is a longer contracted revenue stream, and Amazon's stretched precisely because the two AI-lab commitments run 8 and 10 years. The reading that it is slower is only a criticism against a market treating $496 billion as near-term revenue.
Microsoft's figure is a quarter earlier than Amazon's and covers all commercial contracts, not only cloud, so it is not like for like. Amazon's own backlog also grew triple digits year over year on management's account, so the growth rates are comparable even where the levels are not.
AWS is by far the largest base - $42,232 million in the quarter against Google Cloud's $24,768 million and Oracle's $4,888 million - so equal percentage growth would be a much larger absolute gain. Amazon added over $4.6 billion of revenue quarter over quarter, which management says is about 80% more than its largest increase ever.
The rivals: MSFT, GOOGL, ORCL, CRWV, WMT, and 1 more that is private, in-house or listed abroad.
Who else does this, rival by rival
| Ticker | Rival | Where it competes | How far along | Priced against it |
|---|---|---|---|---|
| MSFT | Microsoft (Azure)Priced against it: Yes | AWS | On sale | Yes |
| GOOGL | Alphabet (Google Cloud)Priced against it: Yes | AWS | On sale | Yes |
| ORCL | Oracle (Oracle Cloud Infrastructure)Priced against it: Yes | AWS | On sale | Yes |
| CRWV | CoreWeave and the specialist AI cloudsPriced against it: No usable figure | AWS | On sale | No usable figure |
| In-house | The customers themselves - Anthropic, OpenAI and any laboratory large enough to buildPriced against it: In-house | AWS | Announced | In-house |
| WMT | Walmart and the discount marketplaces, on the retail halfPriced against it: Yes | North America and International | On sale | Yes |
How big it could get
How big this could get
Nobody publishes what a data center earns, so the market is measured in electricity. The world's data centers used 485 terawatt-hours in 2025; the International Energy Agency expects about 950 by 2030, with AI facilities tripling. Announced building plans for 2026 come to $715B, and Amazon's are about a third of that.
The evidence under this - 3 claims
A projection is not a measurement, and the IEA restated its own numbers between its April 2025 and April 2026 reports. Its capacity data comes from a proprietary vendor, and the agency states plainly that operators and trackers 'do not use a consistent scope or definitions' for data-centre capacity - so the electricity is measured better than the capacity it implies.
The per-kilowatt figure is a consultancy's, reproduced inside the IEA report rather than measured by the agency, and it covers non-IT capital only - the shell, cooling, transformers and power supplies, with the servers excluded. The $715 billion is a sum of company announcements, which is the same class of evidence as Amazon's own $220 billion and no more binding.
The two figures are not struck on identical definitions - Amazon's includes fulfilment-network capital as well as AWS, and the IEA's set of 'leading hyperscalers and neo-clouds' is unnamed. The ratio is an order of magnitude rather than a share.
Who has signed
Customers have committed $496B not yet billed, double the figure six months earlier. Two AI laboratories signed most of the increase: OpenAI for eight years, Anthropic for ten. Amazon funds both, and offers Anthropic a $20B credit line it can draw as AWS delivers.
The evidence under this - 5 claims
Where the growth is coming from, business by business
Where the growth is coming from
| Business | Share of revenue | Growth: last quarter | Last half-year | Last year | Share in 3 years at its half-year rate |
|---|---|---|---|---|---|
| North America and International | 79% | +15.7% | +14.8% | +13.3% | 71% |
| AWS | 21% | +36.8% | +32.7% | +27.5% | 29% |
| Whole company | 100% | +19.6% | +18.2% | +15.8% | 100% |
On the filed split through 2026-06-30, if each business keeps its last half-year's growth, the company grows about 19% next year and 20% in year 3, against 15.8% over the last twelve months. AWS carries the case: it grows 33% against 15% for the rest, and would be 29% of revenue by then. A scenario, not a forecast: no business holds its rate.
How we read it: The business carrying the case: AWS grows 33% over the half-year against 15% for the rest, accelerating five quarters running on a backlog doubled to $496 billion. The 4 already rests on AWS.
The results
The numbers
Sales grew 16% over the twelve months to Jun 2026, about the same pace as the 12% in the year to Dec 2025.
Operating margin is 12%, up from 6.4% in the year to Dec 2023.
Running costs take 39¢ of each sales dollar, up from 37¢ in the year to Dec 2021.
Each dollar put into the business earns 14% a year after tax, about where it was (14%) in the year to Dec 2021.
It used $11.6B of cash over the last twelve months, so the $123B it holds lasts about 10.6 years at that pace.
It owes $9.56B more than the cash it holds, about 0.1 years of the cash its operations bring in.
The share count rose 5.9% in 4.5 years, and pay in shares costs 2.5% of sales.
All the numbers
Filed figures from the company's 10-K and 10-Q, read 2026-10-04. Fiscal years end in December.
| Dec 2021 | Dec 2022 | Dec 2023 | Dec 2024 | Dec 2025 | 12 mo to Jun 2026 | |
|---|---|---|---|---|---|---|
| Size and growth | ||||||
| Sales | $470B | $514B | $575B | $638B | $717B | $776B |
| Sales growth | +22% | +9.4% | +12% | +11% | +12% | +16% |
| Profit | ||||||
| Gross margin | 42% | 44% | 47% | 49% | 50% | 51% |
| Operating margin | 5.3% | 2.4% | 6.4% | 11% | 11% | 12% |
| Net margin | 7.1% | −0.5% | 5.3% | 9.3% | 11% | 17% |
| What growth costs, per sales dollar | ||||||
| Sales and marketing | 6.9% | 8.2% | 7.7% | 6.9% | 6.6% | 6.2% |
| General and admin | 1.9% | 2.3% | 2.1% | 1.8% | 1.6% | 1.4% |
| Pay in shares | 2.7% | 3.8% | 4.2% | 3.5% | 2.7% | 2.5% |
| Capital spending | 13% | 12% | 9.2% | 13% | 18% | 22% |
| Cash | ||||||
| Cash from operations | $46.3B | $46.8B | $84.9B | $116B | $140B | $161B |
| Free cash flow | −$14.7B | −$16.9B | $32.2B | $32.9B | $7.70B | −$11.6B |
| Balance sheet | ||||||
| Cash and short-term investments | $96.0B | $70.0B | $86.8B | $101B | $123B | $123B |
| Debt | $50.6B | $71.3B | $67.0B | $57.8B | $68.9B | $133B |
| Net cash | $45.5B | −$1.32B | $19.8B | $43.4B | $54.2B | −$9.56B |
| Returns and shares | ||||||
| Return on capital | 14% | - | 12% | 19% | 16% | 14% |
| Shares (diluted) | 10.3B | 10.2B | 10.5B | 10.7B | 10.8B | 10.9B |
What growth costs
Growth has made Amazon more profitable: warehouses, marketing and overhead grow slower than sales. The cost that grows faster is the cloud build, both the depreciation on the income statement and the cash spent up front, which is why profit rises while free cash has turned negative.
The evidence under this - 2 claims
Amazon tags no research and development expense, so the numbers card shows none: technology and infrastructure is the company's own label and carries the servers' and data centres' depreciation as well as engineering pay, so it is wider than R&D. The line was called technology and content until 2024.
The last report (30 Jul)
Quarter ended June 2026, filed figures.
On the surface the June 2026 quarter was a triumph: $5.75 a share against a $1.82 forecast. But $50.5B of it was Amazon marking up its stake in Anthropic, a private company it funds. Nothing was sold, and 66% of pre-tax profit came from outside the business. Analysts expect $1.96 for September, which leaves the gain out.
The part that was earned was strong. AWS grew 36.7% and kept 39.4 cents of each sales dollar as operating profit. The cash went the other way: over the year to June, free cash flow was an outflow of $7.6B, against an inflow of $18.2B a year earlier.
What it did to the thesisIt strengthened the business case, with AWS speeding up and keeping more of each sale. It also confirmed why the view is Neutral: the low price against earnings comes from a paper gain on Anthropic, and heavy spending turned cash flow negative. Whether demand keeps pace is open.
The quarter in detail
Sales growth from the previous quarter, filed figures.
Earnings per share $5.75 against $1.82 expected; $5.75 as filed, with no adjustments.
The evidence under this - 8 claims
A consensus estimate is a vendor aggregation of an unknown number of analysts on an unknown basis, and some may be modelling adjusted rather than GAAP earnings, which would produce a similar gap for a different reason. The inference that analysts exclude the mark is strongly supported by the arithmetic but not stated anywhere by anyone.
What management did with the money
It borrowed to pay for about $220B of building this year: $82.4B of new debt in the first half against $4.7B a year earlier, then a $17.5B bank loan in June. Management says more is coming; the FT reports it is sounding out investors on leasing back about $8B of chips (unconfirmed).
The evidence under this - 4 claims
A capex plan is a plan, and management said in the same breath that servers are bought a few months before service and that 'if the demand isn't there, we won't spend the capital'. So the figure is a ceiling that flexes, and the flexible part - servers and networking - is the larger half. The data-centre half, spent two years before monetisation, is the part that cannot be pulled back.
Declining to name the sources is itself the answer analysts did not get, and the options are not equivalent: more unsecured debt is cheap at this balance sheet, but asset-backed or joint-venture structures for data centres would move capital off the reported capex line without reducing the obligation. Nothing in the filings says which is intended, and the term loan signed in June is the only one disclosed. On 2 October 2026 the Financial Times reported, citing people familiar, that Amazon has sounded out investors on moving about $8 billion of Nvidia Grace Blackwell chips into a special-purpose vehicle that would borrow from outside investors and lease the chips back to Amazon, with up to a 10% equity stake offered; Amazon has not confirmed it and no filing carries it, so it is a reported option and not a disclosed arrangement.
- 2026-10-02editedcounter-evidence extended: the asset-backed structure the claim named as a possibility is now reported as being explored, about $8 billion of chips, unconfirmed
The price, and where we disagree
In short. Full. The price suits a business that keeps compounding and leaves no room for one that slows. The low price against earnings comes from a paper gain on its Anthropic stake.
The June quarter made the shares look cheap on earnings for the first time in years, at 20.8 times against a six-year median of 44.6. That cheapness is the paper gain.
The test is the September quarter, reported on 29 October. The company already guides total sales growth down to 9% to 12%, though it says nearly 4 points of that is Prime Day moving dates.
The evidence under this - 3 claims
The denominator contains the marks. Two of the four trailing quarters carry non-operating gains of $16.8 billion and $53.4 billion before tax, so the ratio is dividing today's price by a year of earnings that will not repeat unless Anthropic and OpenAI keep raising money at higher prices.
Operating income is not clean either: it carries approximately $1.2 billion of disclosed one-off benefits in the quarter - tariff refunds and an energy-derivative fair-value gain - which is 4.4% of it. And a case can be made that the Anthropic gain is real economic value accruing to shareholders whether or not it is sold, in which case excluding it understates what the company is worth rather than correcting it.
Prime Day moved into Q2 in 2026 for most of the largest countries, so the Q2 figure is flattered and the Q3 figure depressed by the same shift; adding back the company's own 400 basis points puts underlying growth near 13% to 16%. Operating income is guided to $22.5 billion to $26.5 billion against $17.4 billion, which is 29% to 52% growth and not a deceleration at all.
How we read the multiple
Buyers pay about 32 times last year's operating profit for the whole company, debt included, more than for any rival. That profit leaves out the Anthropic paper gain, so it is the figure to trust.
It already carries the cost of the new data centres, but almost none of the sales they should bring. Amazon spent more cash than it brought in last year, so cash cannot value it yet.
The price works only if AWS keeps growing near today's rate for several more years.
Against its own history. Cheaper than its own past on earnings, dearer on operating profit. The paper gain distorts the first reading; the second is the one to trust.
The evidence under this - 4 claims
The cohort spans four cloud providers and a grocer, and the two cheapest members - Microsoft and Oracle - carry large software annuities at margins Amazon's retail half can never reach. A multiple struck on a company two-fifths of whose profit comes from an 8% margin business is not comparable to one struck on a 47% margin business, and the distance is well short of the 1.5 times the median this repository requires before a score may move.
Operating income is not clean either: it carries approximately $1.2 billion of disclosed one-off benefits in the quarter - tariff refunds and an energy-derivative fair-value gain - which is 4.4% of it. And a case can be made that the Anthropic gain is real economic value accruing to shareholders whether or not it is sold, in which case excluding it understates what the company is worth rather than correcting it.
A negative free cash flow driven by capacity that is contracted for is not the same as one driven by losses - operating cash flow rose 33% to $161,403 million. Management's case is precisely that the outflow is the investment and that revenue growth outpaces capex growth in time. But the ratio is unusable today whatever the reason, so the valuation has to rest on the operating line.
The same arithmetic supports the bull case. If AWS holds anything near 36.7% growth at anything near a 39% segment margin, its operating profit alone compounds past the whole company's current profit within about three years, and the multiple on that basis is unremarkable. The claim is that the price requires the growth, not that the growth is unlikely.
Priced against its rivals
The competitors its own annual report names, each confirmed against the rival's filing. Measured on company value for each $1 of operating profit, cheapest first:
| Ticker | Company | company value for each $1 of operating profit |
|---|---|---|
| MSFT | Microsoft | 24.0× |
| ORCL | Oracle | 25.8× |
| WMT | Walmart | 28.2× |
| GOOGL | Alphabet | 28.9× |
| AMZN | Amazon.com | 31.6× |
The middle of the other 4 is 27.0. Amazon.com, Inc. is dearest of 5.
Rivals' figures come from a data provider, so read the spread of the group rather than any single pair.
Buyers pay more for each dollar of operating profit here than at any rival, so the high price is Amazon's own rather than the industry's.
Private or foreign rivals that cannot be priced: Temu (PDD Holdings) and Shein, Alibaba Group, Anthropic and OpenAI.
The bet: what it pays if it works
This bet is on one business only: AWS, the cloud arm that earns most of the profit. The stores, sellers, advertising and subscriptions are counted separately, at the price buyers pay for Walmart's profit.
If AWS keeps its recent pace for three more years, the stock is worth only somewhat more than today, so this is an expensive bet. If AWS were counted at nothing, the rest would be about a third of today's price: the price already pays for AWS growing.
| AWS operating profit in two to three years | AWSs at the June-quarter run rate needed | At the stated multiple |
|---|---|---|
| $54.68B | about 1 at 100% kept | 0.9× |
| $110.89B (the big win) | about 2 at 100% kept | 1.4× |
| $129.56B | about 2 at 100% kept | 1.6× |
To pay 2× (a Fair bet) it would need about $175.73B of AWS operating profit (about 3 AWSs at the June-quarter run rate); to pay 5× (a Cheap bet), about $503.31B of AWS operating profit (about 8 AWSs at the June-quarter run rate), at the same multiple.
Each row is AWS alone, plus the rest of the company (North America and International) counted at about $1.10T.
The steps it has to take
- Done AWS growth accelerated five quarters running, from 20% in 2025 to 36.7% in the June 2026 quarter
- Done Backlog doubled to about $496 billion on OpenAI's eight-year and Anthropic's ten-year expansions
- In progress September 2026 quarter: AWS growth at or above 33%, segment margin at or above 36%, backlog higher again due 2026-10-29
- Not started Trailing free cash flow positive again while the backlog keeps growing due 2027-08
- Not started AWS 2027 net sales about $227 billion, the half-year rate held for two years, with margin above 30% due 2028-02
The detail behind the label
The evidence under this - 4 claims
The cohort spans four cloud providers and a grocer, and the two cheapest members - Microsoft and Oracle - carry large software annuities at margins Amazon's retail half can never reach. A multiple struck on a company two-fifths of whose profit comes from an 8% margin business is not comparable to one struck on a 47% margin business, and the distance is well short of the 1.5 times the median this repository requires before a score may move.
Where we disagree with the market
AWS is speeding up again, and the shares are cheap on earnings for the first time in years.
The business is as good as they say, but the cheapness is a paper gain. On operating profit the shares are dear.
84 of 94 analysts rate the shares a buy. They read the headline; we read the note explaining where the profit came from.
The evidence against us is real. The company guides September-quarter operating profit up 29% to 52%, has computing capacity reserved into 2028, and has two customers signed for eight and ten years.
What analysts say, and the evidence
The evidence under this - 5 claims
Alphabet states it changed its backlog definition in Q1 2026 to include contracts with an original expected term of one year or less, which Amazon's disclosure excludes. The two are therefore not struck on identical scopes, and part of Alphabet's lead is definitional rather than commercial.
What could go wrong, and how we would know
The risks
- Orders from its own borrowers. The two commitments that doubled its committed orders came from AI laboratories Amazon had just financed.
- A gain that can reverse. Amazon values its Anthropic stake by estimate, since no market prices a private company. Only $7.4B of its $229.7B in stakes trades publicly.
- Cash going out, debt coming in. Spending of about $220B this year turned the cash negative, and management says even that will not buy enough capacity.
- Power running short. Texas paused new data-center building in August and cut its forecast for 2027 electricity demand growth from 14% to 6%.
The evidence under this - 7 claims
Observable-price adjustments under ASC 321 are triggered by actual transactions in the issuer's securities at observable prices, so this is not a model output - Anthropic raised Series G and Series H rounds in the quarter and Amazon bought $5.0 billion of each. The estimate is anchored to a real primary-market price, and Amazon's own participation in those rounds is disclosed.
A capex plan is a plan, and management said in the same breath that servers are bought a few months before service and that 'if the demand isn't there, we won't spend the capital'. So the figure is a ceiling that flexes, and the flexible part - servers and networking - is the larger half. The data-centre half, spent two years before monetisation, is the part that cannot be pulled back.
One state, one revision, and a pause rather than a prohibition. Amazon does not disclose its Texas capacity and the filings name no state, so nothing connects this to AWS's own build. It matters as the first case of the constraint being political rather than physical, not as a quantified hit.
What we still do not know
- What a data center earns once built: the missing link between the spending and the return.
- Whether the laboratories' commitments are priced at today's margins, or discounted for a decade of volume.
- Whether OpenAI and Anthropic would have signed without Amazon's money behind them.
The evidence under this - 5 claims
Every number in it is management's and none is filed. A three-year server break-even on a five-year life leaves a narrow margin if useful lives shorten, and useful life is an accounting estimate Amazon sets itself - it has extended server lives before. The 30-year data-centre life is the assumption doing the most work and the least testable.
A longer life is not unambiguously worse - it is a longer contracted revenue stream, and Amazon's stretched precisely because the two AI-lab commitments run 8 and 10 years. The reading that it is slower is only a criticism against a market treating $496 billion as near-term revenue.
What would change our view
- Neutral→AttractiveThe cash returns with AWS intact
It needs all of:
- ○Free cash flow over the past twelve months turns positive again, while committed orders keep growing every quarter.checked at the next results, due by 29 Oct
- ○AWS keeps growing at 25% or more, with its operating margin above 30%.checked at the next results, due by 29 Oct
- Neutral→UnattractiveAWS decelerates or the stakes are written down
It needs any of:
- ○AWS growth drops below 25% in any quarter, or its operating margin below 30%.checked at the next results, due by 29 Oct
- ○Amazon writes down its Anthropic or OpenAI stakes.watched as it happens
We would also be wrong if
- Committed orders fail to grow in a quarter: customers have stopped signing further ahead.
- Free cash flow is still negative a year from now.
What is coming
- 29 October 2026 Third-quarter 2026 results: Whether AWS growth holds in the thirties with margin near its June level, the backlog rises again, and the trailing free-cash-flow outflow stops widening.
Every dated event ahead
- 29 October 2026 Third-quarter 2026 results: Whether AWS growth holds in the thirties with margin near its June level, the backlog rises again, and the trailing free-cash-flow outflow stops widening.
- February 2027 Fourth-quarter 2026 results and the 2027 capital spending plan: Whether 2027 spending rises again on top of about $220 billion, and whether full-year free cash flow turns back toward positive.
- February 2027 FTC monopolisation trial begins: How much the marketplace fee and advertising model is at legal risk; a remedy on seller fees would hit third-party services and advertising, about a third of sales.
Since we wrote this
- 26 OctNewsAmazon's cloud capacity is largely booked through 2027; Verizon adds a $2B deal. View heldJassy says AWS capacity is largely sold out for 2027; Verizon signs a $2B+ AI deal with AWS.
- 19 OctNewsAmazon backs an Indiana nuclear plant and opens its AI chips to PyTorch code. View heldAmazon, X-energy and AEP plan an Indiana reactor project; Amazon releases TorchNeuron for Trainium.
- 5 OctThesis reviewThe bet now names what it ridesThe owner asked that every bet name the macro trend behind it, after BWXT's bet was missing its chief executive's case for nuclear over gas.
- 4 OctThesis reviewOwner's checklist: revenue streams, cost of growth, management, catalystsThe owner's checklist found the method silent on who runs the company, what growth costs, where the money comes from and what is dated ahead; this pass fills those four blocks from the proxy and filings.
- 3 OctNewsAmazon pledges $1 billion to data-centre host towns and fights moratoriums; no claim moved. View heldAWS chief's post against data-centre moratoriums, with a $1 billion community pledge.
- 2 OctNewsReport: Amazon may lease back $8B of chips via a vehicleA press report, unconfirmed, says Amazon is exploring moving chips off its books. It would hide spending, not remove the debt, so it raises the bar for reading cash flow at the October results.
- 2 OctBet moveBet re-struck: Priced for success to Safe, little upsideRe-priced at the 1 October close.
- 24 SepNewsAmazon borrowed again and the FTC sued over ad pricing; no claim moved. View heldA sterling bond sale, an FTC suit over advertising prices, and a Qualcomm chip collaboration.
- 25 AugThesis writtenNeutral, medium convictionWritten after the June quarter, when a paper gain on Anthropic tripled reported earnings and made the shares look cheap.
Also checked 10 times with nothing new to act on.
Sources 22
Sources - 22
Every claim above cites one of these, or states the reasoning that derived it. The quote is frozen at the date it was read: a link can rot, and an agent fetch returns a reading of a page rather than the page. Backs counts the whole claim ledger, so it can name a claim this page does not surface.
- S01primarycompany filing2026-02-06Item 1 on what the company sells and the ten classes of competitor it names, the FY2025 statements by segment, and the December 2025 backlog of $244 billion at a 4.1-year weighted-average lifebacksC03C08C53C54C55C58C59
- S02primarycompany filing2026-07-31the Q2 2026 income statement line by line, the other-income note that decomposes the $53.4 billion, the Anthropic and OpenAI investment note with carrying values and the $20 billion facility, the $496 billion of unrecognised customer commitments, the free-cash-flow reconciliation and the contractual-commitment tablebacksC01C02C03C07C08C09C10C11C12C13C16C17C18C19C20C26C27C30C32C34C39C40C41C48C53
- S03primarycompany release2026-07-30the segment table, the third-quarter guidance, the supplemental revenue split by product group, the headcount, and the company's own disclosure of both the Anthropic gain and the free-cash-flow outflow
Second quarter 2026 net income includes non-operating pre-tax other income of $53.4 billion, primarily from our investments in Anthropic.
backsC01C02C04C05C15C16C22C26C30C33C35C36C44C45C47 - S04primarycompany release2026-02-05the full-year 2025 base the 2026 acceleration is measured from, the Q4 2025 special charges, and the first-quarter 2026 guidance
AWS segment sales increased 20% year-over-year to $128.7 billion.
backsC33 - S05primarycompany release2026-04-29the middle rung of the free-cash-flow descent and the first of the two Anthropic marks, so the June figures can be read as a trend rather than an event
First quarter 2026 net income includes pre-tax gains of $16.8 billion included in non-operating income from our investments in Anthropic. ... Free cash flow decreased to $1.2 billion for the trailing twelve months
backsC33 - S06primarytranscript2026-07-30Amazon.com second-quarter 2026 earnings call, 30 July 2026data/cache/transcripts/AMZN-2026Q2.jsonthe capital-cycle explanation, the raised capex plan, the two one-off benefits inside operating income, the reserved-capacity commentary for 2027 and 2028, and management's own trillion-dollar framing for AWS
We now believe we will spend approximately $220 billion in cash CapEx in 2026. The higher cost of memory pushing this number up from our prior estimate of about $200 billion. ... Remember, by the way, that 85% of the global IT spend is still on premises. That equation is going to flip in the next 10 to 20 years.
backsC04C05C06C28C29C32C34C46C51 - S07primarycompany filing2026-04-29Microsoft Form 10-Q, quarter ended 31 March 2026 Microsoft Corporationthe nearest competitor's own filed figure for the two numbers Amazon's case rests on - contracted backlog and cloud growth - struck on its own books rather than on a vendor's
Commercial remaining performance obligation increased 99% to $627 billion. ... Azure and other cloud services revenue increased 40%.
backsC14C44C45 - S08primarycompany filing2026-07-23Alphabet Form 10-Q, quarter ended 30 June 2026 Alphabet Inc.the same quarter's backlog and cloud revenue from the competitor closest to Amazon in structure, and the conversion speed that makes the two backlogs comparable rather than merely similar in size
As of June 30, 2026, we had $519.5 billion of remaining performance obligations (“revenue backlog”), of which $513.9 billion related to Google Cloud. ... We expect to recognize just over 50% of the revenue backlog as revenues over the next 24 months
backsC12C13C15C44C45 - S09primarycompany filing2026-03-11Oracle Form 10-Q, quarter ended 28 February 2026 Oracle Corporationa third provider's separately disclosed cloud-infrastructure revenue line, which nearly doubled year over year - evidence that the demand Amazon describes is being met by more than one seller
Cloud infrastructure | 4,888 | 2,652 | 12,314 | 7,240
backsC15C44C45 - S10primarycompany filing2026-03-02CoreWeave Form 10-K, fiscal year ended 31 December 2025 CoreWeave, Inc.the pure-play's own description of the product, which is the cleanest outside evidence that AI cloud capacity can be sold by a company with none of Amazon's scale, distribution or installed base
We designed and built a carrier-grade network backbone engineered specifically for the ultra-low latency, high-throughput requirements of our customers' training, data replication, inference, and agentic AI initiatives.
read, but no claim rests on it - S11primarycompany filing2026-06-10the size of the bank facility signed alongside the note issues, and the identity of the twenty-five lenders funding a build the operating business no longer covers
As of the Closing Date, the Aggregate Commitments equal to US$17,500,000,000.
backsC31 - S12primarycompany filing2026-03-13Walmart Form 10-K, fiscal year ended 31 January 2026 Walmart Inc.the retail half of the cohort, confirmed against its own filing, and the reason a single median cannot price this company
Walmart Inc. (“Walmart,” the “Company” or “we”) is a people-led, technology-powered omnichannel retailer dedicated to helping people around the world save money and live better by providing the opportunity to shop in both retail stores and through eCommerce
read, but no claim rests on it - S13primaryregulator2025-07-31UK Competition and Markets Authority, cloud services market investigation - summary of final decision UK Competition and Markets Authoritya competition regulator's own measurement of how sticky this product is, and its finding on which of the alleged barriers actually bind - egress fees and technical differentiation yes, committed spend discounts no
Very few customers switch between clouds: less than 1% of customers switch provider each year. ... A key commercial barrier is the presence and magnitude of egress fees required to transfer data between cloud providers for the purposes of switching and/or multi-cloud.
backsC23 - S14primaryregulator2025-07-31UK Competition and Markets Authority, cloud services market investigation - final decision report UK Competition and Markets Authoritythe share figures, banded and redacted by the regulator itself, and the one sentence in the record that says which of the two leaders has been losing ground
The IaaS market is highly concentrated and the two largest providers, Microsoft and AWS, each has a high share of supply at [30-40]% in 2024. Microsoft has grown its share while AWS' share has decreased since 2020.
backsC24 - S15primaryregulator2026-03-31UK Competition and Markets Authority, actions on cloud and business software through the UK digital markets competition regime UK Competition and Markets Authoritywhat the regulator did with its own finding - voluntary commitments rather than a strategic-market-status designation, with a Board review of progress six months out, which is the dated test of whether the switching barrier is coming down
Action by both parties to remove egress fees from UK customers' contracts when they want to switch cloud provider. They will apply this for a switching period of at least 180 days, and as soon as a customer decides to switch at least one service.
backsC25 - S16primaryvendor data2026-08-25Research pack for AMZN, frozen beside this thesiscontent/AMZN/research-pack.jsonthe price, the market capitalisation and enterprise value, the twenty-day range, the analyst split, the insider filings, the six-year multiple medians, the cohort multiples, and the consensus estimate for the September quarterbacksC21C22C37C38C39C40C41C42C43
- S17primaryindustry data2026-04-16International Energy Agency, Key Questions on Energy and AI International Energy Agencythe size of the build in a physical unit measured by an institution that sells no compute, the rate it is growing at, and the total capital the industry has announced for the year Amazon is spending $220 billion in
Our updated projections see electricity consumption from data centres roughly doubling from 485 TWh in 2025 to 950 TWh in 2030, accounting for around 3% of global electricity demand by that date. Electricity consumption from AI-focused data centres grows much faster than overall data centre electricity consumption, tripling in this period. ... Leading hyperscalers and neo-clouds have announced a 75% increase in capital expenditure for 2026, amounting to USD 715 billion.
backsC49C50C51 - S18primaryindustry data2025-04-10International Energy Agency, Energy and AI (World Energy Outlook special report) International Energy Agencythe capital cost of a gigawatt of data centre, which is the conversion this prize needs and which no filer publishes - note the per-kilowatt figure is a consultancy's, reproduced inside the IEA report rather than measured by it
The combined cost of construction and installations ranges from USD 6 000 per kilowatt of IT load in some emerging market and developing economies to over USD 10 000 in many advanced economies (Turner & Townsend, 2024). ... An additional 64 GW of greenfield data centre IT load was built over the past decade, causing annual investment to grow from around USD 100 billion in 2015 to over USD 500 billion in 2024.
backsC50 - S19primarygovernment data2026-08-11U.S. Energy Information Administration, Short-Term Energy Outlook, August 2026 U.S. Energy Information Administrationthe first dated, official mark against the build - a state pausing new development and a federal agency cutting its own load forecast by more than half in one revision
On August 3, the Texas governor announced a pause on new data center development, and as a result, we have lowered our forecast for electricity demand in Texas. We expect electricity load in Texas will grow by 6% in 2027, in contrast to our forecast of 14% growth in the previous STEO.
backsC52 - S20secondarynews2025-06-04the month the FTC's monopolisation case against Amazon goes to trial; the full article is behind registration and the 10-Q's legal note gives no date
A Washington federal judge granted today the US Federal Trade Commission's proposal to move the trial date for the agency's antitrust case against Amazon to February 2027.
backsC60 - S21primarycompany filing2026-04-09the beneficial ownership table at 24 February 2026, the compensation discussion and analysis, and the summary compensation tablebacksC56C57
- S22primarycompany filing2023-02-03the 2021 and 2022 operating expenses as a percentage of net sales, including technology and content before it was renamed technology and infrastructurebacksC59