Amazon — Bull & Bear Memo
1. Business in one line
Amazon runs two businesses on one balance sheet. Retail earns a thin spread on its own goods but high-margin fees, ads and subscriptions on traffic and logistics it has already paid for. AWS rents compute, storage and AI infrastructure on capital it deploys ahead of demand.
FY2025 revenue was $716.9B. AWS was 18% of revenue but 57% of operating income ($45.6B of $80.0B). Third-party seller services ($172.2B), advertising ($68.6B) and subscriptions ($49.6B) together made up 41% of revenue, and all three monetize the same retail asset base FY2025 10-K.
StalwartBut AWS is now behaving like a fast grower on a cyclical capital build. For a stalwart, the question isn't whether it grows. It's whether the growth earns its cost of capital, and right now that means the returns on the AI capex cycle.
2. Bull case — Peter Lynch pitch
The simple reason this stock could work is that Amazon has already paid for the two hardest-to-copy assets in its industries: the densest consumer logistics network and the largest cloud installed base. Both are now being monetized through higher-margin, contract-backed or capital-light revenue faster than the new AI capex drags down returns.
B1The AWS backlog converts to revenue at a high margin
- What must happen
- The $496B of contracted AWS commitments (6.4-year weighted life) turns into revenue on schedule, and AWS segment margin holds in the mid-to-high 30s as the 2025–26 capex vintage depreciates.
- Why Amazon specifically
- Scale in data centers and power, plus in-house silicon: the new lab contracts include obligations tied to the performance of AWS's own chips. Interpretation: selling its own silicon instead of reselling third-party GPUs keeps more of the margin at Amazon.
- Evidence
- AWS growth went from 17% to 37% (Q2 25 to Q2 26), and margin rose from 32.9% to 39.4% even though AWS D&A rose 67% ($4.8B to $8.1B). Backlog went from $244B to $496B in six months FY2025 10-K; FY2026 Q2 10-Q.
- What to monitor
- AWS margin compared with AWS D&A growth, and backlog growth from customers other than OpenAI and Anthropic.
B2Capex intensity peaks and free cash flow turns positive again
- What must happen
- TTM capex ($169.0B, 21.8% of revenue) plateaus while operating cash flow ($161.4B TTM, up 33% y/y) keeps compounding, so TTM FCF swings back from –$7.6B to positive.
- Why Amazon specifically
- Amazon has done this before. FCF was –$11.6B in FY2022 after the fulfillment build and +$36.8B in FY2023 once capex eased and operating cash flow caught up FY2023 10-K. This time much of the capacity is pre-sold through the backlog.
- Evidence
- TTM OCF went from $121.1B to $161.4B. Management says capex will "increase in 2026" but gives no ceiling FY2026 Q2 10-Q.
- What to monitor
- TTM FCF, and capex as a share of revenue.
B3Retail margins keep expanding because the growth is in ads and 3P fees, not first-party boxes
- What must happen
- Advertising and seller fees keep growing faster than first-party retail volume, and fulfillment cost keeps falling as a share of sales, so North America margin rises from 7.9% toward double digits.
- Why Amazon specifically
- Ads monetize purchase intent on Amazon's own shelf and need little extra capital. Third-party volume runs through a logistics network the first-party business already paid for, and Prime density lowers cost per unit.
- Evidence
- Advertising was $19.8B in Q2 26, up 26%, against 15% growth in online stores. North America margin rose from 7.5% to 7.9%, and fulfillment fell from 15.5% to 14.8% of sales FY2026 Q2 10-Q. Q3 guidance calls for $22.5–26.5B of operating income against $17.4B, although Q3 25 carried $4.3B of FTC and severance charges FY2025 10-K.
- What to monitor
- The gap between ad growth and online-store growth, and North America margin excluding one-offs.
3. Bear case — Munger invert
The most likely way I lose money is that Amazon has turned AWS into a vendor-financed utility. It is borrowing to build capacity for two AI labs it also funds, booking their valuation markups as earnings, and locking in five-year-life assets against revenue that depends on those labs continuing to raise money.
Ordered by severity and permanence, not by the bull sequence.
R1The backlog is concentrated and circularPermanentattacks B1
- How it could fail
- The OpenAI expansion (+$100B over 8 years, on top of $38B) and the Anthropic expansion (>$100B over 10 years) total more than $200B. That's roughly 80% of the $252B backlog increase (our estimate, since the filings don't give exact concentration). Amazon committed $50B to OpenAI and has put about $18B into Anthropic. It also has a $15B undrawn Anthropic facility that opens up as AWS hits compute-delivery milestones, so Amazon's cash partly funds the customer's ability to pay for Amazon's capacity FY2026 Q2 10-Q.
- What would confirm failure
- A flat or down funding round at either lab, a rephased or renegotiated commitment, the backlog stalling while its weighted life lengthens, or a negative mark in "other income".
- Damage to economics
- Revenue visibility, the $229.7B investment carrying value and the returns on capex all get hit at once. H1 26 investment gains ($69.1B) exceeded operating income ($51.3B), so a reversal would swing reported earnings hard. Capacity built around one customer's chip contract is hard to resell at the same price.
R2Capex is a treadmill, not a humpPermanentattacks B2
- How it could fail
- In 2025 Amazon cut server lives from six years to five, citing the pace of AI. So AI hardware has to be replaced about every five years, and replacement capex grows with the installed base. AWS net PP&E rose from $190.1B to $263.8B in six months. AWS revenue per dollar of net PP&E fell from about 0.68x to 0.56x (TTM revenue over period-end PP&E, our calculation). The 2022 precedent cuts both ways: that FCF snapback came from pulling back on a fulfillment overbuild.
- What would confirm failure
- Capex still outgrowing OCF into 2027, and more debt. Long-term debt went from $65.6B to $128.9B in six months, plus $25.0B of notes in July (coupons up to 6.25%) and a $17.5B term loan. Total commitments went from $439.7B to $650.0B FY2025 10-K; FY2026 Q2 10-Q.
- Damage to economics
- A permanent shift from compounder to capital-intensive utility. Quarterly interest expense has already risen from $0.5B to $1.3B, and incremental returns fall even if revenue grows.
R3AWS margin is peaking before the depreciation landsPartly fixableattacks B1
- How it could fail
- AWS D&A has already risen to 19.1% of AWS revenue, from 15.7% a year earlier. AWS net additions to property and equipment were $48.6B in Q2, three times the prior year, and most of that hasn't started depreciating yet. The 10-Q attributes pricing pressure to "long-term customer contracts", which suggests big negotiated contracts carry lower prices (our interpretation).
- What would confirm failure
- AWS margin falls toward the low 30s while growth stays above 30%, which would mean the growth is being bought.
- Damage to economics
- Every 1 point of AWS margin is about $1.5B of operating income on TTM AWS revenue of $148.4B. Partly fixable through repricing at renewal, but structural if a few negotiated lab contracts dominate the mix.
R4Retail margin expansion is partly one-off and exposed to regulatorsFixable, except for a regulatory remedyattacks B3
- How it could fail
- Q2 cost of sales included about $640M of IEEPA tariff refunds, mostly in North America. That's roughly 0.55 points of North America's 7.9% margin, so the underlying figure is about 7.3%, slightly below last year's 7.5%. Management also flagged 2026 spending on Amazon Leo, quick commerce and sharper international prices FY2025 10-K; FY2026 Q2 10-Q. The $2.5B FTC settlement shows the ads, fees and fulfillment bundle is under legal scrutiny.
- What would confirm failure
- North America margin excluding one-offs stays flat for two more quarters, fulfillment stops falling as a share of sales, or a remedy splits the marketplace from fulfillment.
- Damage to economics
- Cyclical cost pressure is fixable. A structural antitrust remedy would be permanent because it breaks the cross-subsidy that funds the highest-margin lines.
Bull vs bear — assumption by attack
The $496B of contracted AWS commitments (6.4-year weighted life) turns into revenue on schedule, and AWS segment margin holds in the mid-to-high 30s as the 2025–26 capex vintage depreciates.
Evidence: AWS growth went from 17% to 37% (Q2 25 to Q2 26), and margin rose from 32.9% to 39.4% even though AWS D&A rose 67% ($4.8B to $8.1B). Backlog went from $244B to $496B in six months FY2025 10-K; FY2026 Q2 10-Q.
The OpenAI expansion (+$100B over 8 years, on top of $38B) and the Anthropic expansion (>$100B over 10 years) total more than $200B. That's roughly 80% of the $252B backlog increase (our estimate, since the filings don't give exact concentration). Amazon committed $50B to OpenAI and has put about $18B into Anthropic. It also has a $15B undrawn Anthropic facility that opens up as AWS hits compute-delivery milestones, so Amazon's cash partly funds the customer's ability to pay for Amazon's capacity FY2026 Q2 10-Q.
Confirms failure: A flat or down funding round at either lab, a rephased or renegotiated commitment, the backlog stalling while its weighted life lengthens, or a negative mark in "other income".
AWS D&A has already risen to 19.1% of AWS revenue, from 15.7% a year earlier. AWS net additions to property and equipment were $48.6B in Q2, three times the prior year, and most of that hasn't started depreciating yet. The 10-Q attributes pricing pressure to "long-term customer contracts", which suggests big negotiated contracts carry lower prices (our interpretation).
Confirms failure: AWS margin falls toward the low 30s while growth stays above 30%, which would mean the growth is being bought.
TTM capex ($169.0B, 21.8% of revenue) plateaus while operating cash flow ($161.4B TTM, up 33% y/y) keeps compounding, so TTM FCF swings back from –$7.6B to positive.
Evidence: TTM OCF went from $121.1B to $161.4B. Management says capex will "increase in 2026" but gives no ceiling FY2026 Q2 10-Q.
In 2025 Amazon cut server lives from six years to five, citing the pace of AI. So AI hardware has to be replaced about every five years, and replacement capex grows with the installed base. AWS net PP&E rose from $190.1B to $263.8B in six months. AWS revenue per dollar of net PP&E fell from about 0.68x to 0.56x (TTM revenue over period-end PP&E, our calculation). The 2022 precedent cuts both ways: that FCF snapback came from pulling back on a fulfillment overbuild.
Confirms failure: Capex still outgrowing OCF into 2027, and more debt. Long-term debt went from $65.6B to $128.9B in six months, plus $25.0B of notes in July (coupons up to 6.25%) and a $17.5B term loan. Total commitments went from $439.7B to $650.0B FY2025 10-K; FY2026 Q2 10-Q.
Advertising and seller fees keep growing faster than first-party retail volume, and fulfillment cost keeps falling as a share of sales, so North America margin rises from 7.9% toward double digits.
Evidence: Advertising was $19.8B in Q2 26, up 26%, against 15% growth in online stores. North America margin rose from 7.5% to 7.9%, and fulfillment fell from 15.5% to 14.8% of sales FY2026 Q2 10-Q. Q3 guidance calls for $22.5–26.5B of operating income against $17.4B, although Q3 25 carried $4.3B of FTC and severance charges FY2025 10-K.
Q2 cost of sales included about $640M of IEEPA tariff refunds, mostly in North America. That's roughly 0.55 points of North America's 7.9% margin, so the underlying figure is about 7.3%, slightly below last year's 7.5%. Management also flagged 2026 spending on Amazon Leo, quick commerce and sharper international prices FY2025 10-K; FY2026 Q2 10-Q. The $2.5B FTC settlement shows the ads, fees and fulfillment bundle is under legal scrutiny.
Confirms failure: North America margin excluding one-offs stays flat for two more quarters, fulfillment stops falling as a share of sales, or a remedy splits the marketplace from fulfillment.
4. Signals to monitor
Each chip jumps to the assumption or risk the signal tracks.
6. Bottom line
- Amazon could work because it has already paid for the densest logistics network and the largest cloud base, and both are now earning higher-margin revenue: AWS at a 39.4% margin with a $496B contracted backlog, and retail through advertising growing 26%.
- What must go right is that the backlog converts on schedule at mid-to-high-30s AWS margins while capex plateaus, so free cash flow swings from –$7.6B TTM back to positive without long-term debt doubling again.
- The most important way the thesis breaks is circularity: if OpenAI or Anthropic stumble, Amazon loses revenue, investment value and capex returns at the same time while the debt stays on its books, because it has committed about $68B to the two labs and they drive most of the backlog expansion.
- I would turn more bullish on backlog growth from customers outside the two labs plus a positive-FCF quarter with capex still rising, and more bearish on any down-round mark in other income or AWS margin slipping toward the low 30s.
Sources
Annual reports 2
- Form 10-K FY2025 (filed 2026-02-06) — Segment and product revenue, FY25 op income, RPO at Dec-25, server useful-life change, FTC/severance charges, total commitments, Anthropic history
- Form 10-K FY2023 (filed 2024-02-02) — FY2022–23 FCF precedent
Quarterly reports 2
- Form 10-Q FY2026 Q2 (filed 2026-07-31) — Q2/H1 26 segment results, AWS D&A and PP&E, RPO $496B, OpenAI/Anthropic commitments and investments, TTM FCF, debt, commitments, tariff refunds, guidance
- Form 10-Q FY2025 Q2 (filed 2025-08-01) — Prior-year comparatives
Our calculations (not disclosed) 3
- Backlog concentration ~80% — (> $200B disclosed lab expansions) / ($496B – $244B). Assumes both expansions sit fully in RPO.
- AWS revenue / net PP&E — FY25 $128.7B / $190.1B = 0.68x; TTM $148.4B / $263.8B = 0.56x
- NA margin ex tariff refund — ($9,123M – ~$640M) / $116,177M ≈ 7.3%; refund allocation to NA is approximate
All figures from company filings in sources/. Section 5 (external challenge) not run.