Amazon is a good company, but that does not make it a good stock at every price. Its commerce platform, advertising and installed cloud business are mature and profitable, and near-term AI demand is strong. What remains unproven is whether the money now going into chips and data centres will still earn an attractive return after industry supply catches up. This is a combination of good businesses and a large investment awaiting proof—not a stock that can be owned without regard to price.
01 · The answer in thirty seconds
If you read only one section, read this one. The rest of the report simply shows how these four answers were reached.
Two businesses, one capital wallet
One side runs retail, Prime, third-party merchants, logistics and advertising. The other, AWS, rents computing, storage, databases and AI capabilities to enterprises. The customers differ, but one management team reinvests the cash from both.
AWS is the profit centre
In Q2 2026, AWS supplied about 21% of revenue but roughly 60% of operating profit. Commerce produces most revenue; advertising and third-party merchant services are probably its economically richer layers.
Demand has arrived; returns have not
Orders and cloud revenue show that AI compute has users. But every major supplier is building. If capacity catches demand, price and utilisation can fall. Shareholder return depends on what the new capital earns, not on how impressive the order book looks.
Meaningful upside, modest base reward
Using $262.65 on 14 August 2026 as the starting point, illustrative annualised outcomes to end-2028 are about −15% bear, +11% base and +39% bull. These are stress tests, not price targets. The real question is whether +11% adequately pays for an unproven capital return.
02 · What kind of company is Amazon?
The most common mistake is to put AWS inside the retail flywheel. It is a largely separate business. Commerce and AWS mainly share capital, engineering culture and management.
Amazon buys inventory and sells it itself. Customers pay for the goods; Amazon bears inventory, warehousing and delivery costs, so margins are thin.
Other sellers use the marketplace. They pay platform, fulfilment and advertising fees, while carrying most inventory risk themselves.
Prime raises purchase frequency and loyalty. Advertising charges brands and merchants for influence at the moment a shopper is choosing what to buy.
Enterprises pay by usage for servers, storage, databases and AI compute. Once systems, data and permissions are built there, moving often means rewriting and migrating them.
Why Amazon's accounts are confusing: profit is reported through North America, International and AWS, while revenue is also split into stores, third-party services, ads and subscriptions. These classifications overlap and cannot be added as if each were a separate business.
03 · Where does Amazon really make money?
Revenue shows scale. Profit shows which part of the company drives its value.
| Q2 2026 | Share of revenue | Share of operating profit | How to read it |
|---|---|---|---|
| North America + International commerce | About 79% | About 39% | Enormous scale across first-party retail, merchant services, ads and subscriptions. Profit is not disclosed for each activity. |
| AWS | About 21% | About 60% | Only one-fifth of revenue, yet the largest profit source. Quarterly growth was about 37% and operating margin about 39.4%. |
Merchandise creates traffic; services probably create more of the commerce profit
Advertising revenue was about $19.8 billion in Q2 2026, up roughly 26%. Amazon does not disclose ad profit, so “advertising is most of retail profit” is not a company fact. But even a conservative margin sensitivity suggests ads could contribute a large share of North America and International profit. The more careful conclusion is that first-party retail may not be very profitable by itself, but it creates shoppers, merchants, order density and advertising opportunities.
That is why Prime, logistics and first-party sales should not be dismissed simply because their stand-alone margins are modest. They are the location, membership system and foot traffic of the mall. The richer counters are marketplace services, fulfilment and advertising.
Why AWS earns more
Once an enterprise builds its databases, permissions, software and operations on AWS, moving is not a website change. It involves data migration, code changes, security and compliance work. That friction makes installed customers sticky. Amazon also designs Trainium and Graviton chips to lower the cost of each unit of computing and help protect margins during price competition.
The limit matters: AWS is strong at retaining installed workloads, but it lacks an enterprise distribution channel on the scale of Microsoft Office and Copilot. Its stickiest databases and managed services are also different from today's fastest-growing activity—large AI labs renting blocks of compute. The former resembles compounding platform revenue; the latter resembles capital-intensive leasing.
04 · What is the investment case?
The case is not “Amazon is strong, therefore buy it.” It is a sequence. If one link fails, the stock outcome weakens.
Management expects roughly $220 billion of cash capital expenditure in 2026, with most directed toward AWS and AI infrastructure. That full-year plan exceeds trailing operating cash flow of about $161.4 billion. The periods do not match exactly; the comparison is only meant to show scale. This is not a small accounting adjustment. It is one of the largest capital-allocation decisions in Amazon's history.
- Enterprises and AI labs keep increasing compute use
- New data centres fill reasonably quickly
- Price and margin survive the supply buildout
- Demand broadens from a few large customers into stickier enterprise services
- Free cash flow per share grows after depreciation and construction spending
Amazon is no longer betting on whether anybody will use AI. It is betting that money spent today will become enough durable profit a few years from now. Near-term orders are here. Long-term returns have not yet been delivered.
Why strong demand can still lose money
Air travel can grow for decades while airline investors still suffer if every carrier buys too many aircraft. Data centres are similar. AWS, Microsoft, Google, specialist clouds and large AI companies all see demand, and all are building. Demand determines how many users want compute; supply determines what each machine can charge.
During shortage, customers urgently need compute, so AWS has more room to hold price and margin. When supply loosens, prices can fall while the chips and buildings already purchased continue to create depreciation for years. That is why demand growth is not the same as shareholder return growth. The evidence should become clearer around 2027–2028, when new capacity enters service and more depreciation reaches the income statement.
Why Amazon is still worth serious attention
A thick foundation
Prime, marketplace services, advertising, logistics and installed cloud workloads are mature businesses. Amazon is not a pure data-centre operator dependent on one order.
The bet fits its capabilities
Amazon's strongest historical results have tended to come from platforms and infrastructure. Cloud, logistics and chips fit its engineering and cost discipline better than a chase into an unfamiliar category.
Cost helps on the way down
Custom silicon and scale cannot prevent a downturn, but they can help AWS suffer less than higher-cost rivals in a price war. The company may be resilient even when the stock is not.
05 · What is proven, and what is not?
Separating fact from judgement prevents one strong quarter from becoming an unsupported ten-year forecast.
| Question | Current view | Evidence and boundary |
|---|---|---|
| Is there near-term AI compute demand? | Yes, it is strong | AWS grew about 37% and its AI business exceeded a $25 billion annual run rate. Contracted but unrecognised revenue is also large, though customer mix and term structure are not transparent. |
| Can AWS hold its economics during shortage? | Current signs are positive | AWS operating margin was about 39.4%. That shows strong current economics, but does not by itself prove cost pass-through or permanent pricing power after supply loosens. |
| Is growth broad-based across ordinary enterprises? | Not yet clear | Publicly announced mega-contracts are concentrated among a few frontier AI companies. Amazon does not split lab compute rentals from broader enterprise platform use. |
| Will the buildout become shareholder cash? | Not proven | Trailing operating cash flow was about $161.4 billion, up 33%, so the business is healthy. Free cash flow was about −$7.6 billion, mainly because AI investment surged. |
| Can commerce keep funding the company? | For now, with risks | North America and International margins improved, while ads outgrew retail. Tariffs, regulation, merchant economics and AI shopping agents can hit its richer layers first. |
Why Anthropic needs its own paragraph
Anthropic is a frontier model company. Amazon both owns a stake and rents it AWS capacity; when Anthropic's valuation rises, Amazon also recognises an investment gain. In Q2 2026, Amazon reported about $53.4 billion of non-operating pre-tax other income, primarily from revaluing that investment. It is neither fraud nor operating cash. The issue is common risk: if AI financing weakens, the investment value, AWS customer demand and data-centre utilisation could come under pressure together.
Measurement note: the claim that ads may contribute much of commerce profit, and estimates of large AI-customer concentration in contracted revenue, are analytical inferences from public information—not line items disclosed by Amazon.
06 · Valuation and return: is a good company worth buying now?
Neither reported P/E nor current free cash flow tells the whole story. Investment marks lift reported profit and make reported P/E look lower, while the construction peak depresses cash flow. The table below is a transparent stress test.
Start with the assumptions
The exercise starts at $262.65 on 14 August 2026 and runs to the end of 2028. To avoid treating Anthropic's revaluation as operating profit, it uses the research-date 2027 consensus EPS of $10.34 as a bridge, equal to about 25.4 times earnings. Consensus still embeds assumptions about future margin and capital returns, so it is not ground truth.
| Scenario | What happens | 2028 assumption | Scenario result | Annualised from start |
|---|---|---|---|---|
| Bear | Supply catches demand; AWS growth falls into the teens; compute discounts, depreciation and investment write-downs pressure profit. | $10 EPS × 18 | About $180 | About −15% |
| Base | AWS and ads keep growing; the AI investment broadly earns its cost, but cash returns more slowly than bulls expect. | $13.27 EPS × 25.4 | About $337 | About +11% |
| Bull | Enterprise AI broadens; high-margin platform and adjacent services take over; AWS sustains high growth and investment becomes cash. | $16.5 EPS × 35 | About $578 | About +39% |
How to read this table: these are neither price targets nor probability forecasts. Each row pairs a business outcome with EPS and the multiple investors might pay. The $13.27 base-case EPS is the 2028 market consensus captured on 12 August 2026; the bear and bull figures are stress values around it. The bear case is not bankruptcy. It is a still-important company suffering both an earnings downgrade and a multiple downgrade because capital returns weaken.
Bull upside is much larger than the moderate downside, but the base case returns only about 11% a year after already assuming that the giant buildout broadly works. For a risk whose answer may not arrive until 2027–2028, that reward is not especially generous. The higher the entry price, the more success is prepaid.
Trailing free cash flow was about −$7.6 billion, so current cash flow cannot support a dependable valuation. The base case above effectively requires free cash flow to turn positive after the construction peak and then keep rising. If free cash flow per share is not improving steadily by 2028, the base case has not worked.
07 · How does the investment win or lose?
A business win and a stock win are different. AWS can keep growing while shareholders lose if growth falls short of what the price already assumed.
Demand and return arrive together
Ordinary enterprises succeed a few frontier labs; new data centres stay well used; AWS margin holds after depreciation rises; capital spending slows while cloud growth remains high and free cash flow recovers.
A sound company whose price prepaid success
AWS, ads and retail efficiency improve, but cash returns slowly and profit merely meets prior expectations. Operations are fine, yet the stock delivers only ordinary market-like returns.
Supply catches up and revenue quality is weak
Growth remains dependent on a few AI labs; financing or contracts weaken; suppliers discount long commitments; price, utilisation and residual values fall together. Earnings and the earnings multiple decline at once.
One easily misread signal: free cash flow comes back
| Same headline | What happens at the same time | Meaning |
|---|---|---|
| Free cash flow turns positive | AWS still grows around 25% or faster and margin holds | More constructive: investment is harvesting while demand remains strong. |
| Free cash flow turns positive | AWS growth falls into the teens and capex is forced down | More negative: building stopped because demand cooled; the cycle may be peaking. |
Twenty-five per cent is not a magical exact line. It is a practical marker between “still growing quickly” and “a clear step-down.” The trend over several quarters matters more than one print.
08 · How each risk hurts shareholders—and what we watch
A useful risk list must explain what happens, how the shareholder is hurt, and where the evidence appears first.
| Priority | Risk | Shareholder damage | What we watch |
|---|---|---|---|
| 1 | Too much AI capacity | Compute price, utilisation and return fall while existing assets keep depreciating; earnings and valuation weaken together. | Large discounts for long contracts; easing AI-chip, power and data-centre supply; capex rising while AWS growth slows. |
| 2 | Common concentration in AI labs and Anthropic | One financing downturn can hit cloud orders, investment marks and data-centre utilisation at once. | Contract delays or renegotiation, weaker lab financing, and whether Amazon begins to show broader enterprise growth. |
| 3 | AWS retains the past but misses high-value growth | Infrastructure still grows, but richer software, data and application value goes to Microsoft, Google or independent platforms. | Whether growth moves from lab machine rental into AWS's enterprise AI platform (Bedrock), databases, data management and compliance services; margin after depreciation. |
| 4 | Commerce's richer layers weaken | Regulation, tariffs, merchant costs or shopping agents can pressure marketplace, fulfilment and advertising together. | Ad growth persistently below merchandise sales; stalled North America margin; whether court-ordered changes directly affect ads or marketplace rules. |
| 5 | Capital allocation fails while shares dilute | Amazon pays no regular dividend, repurchases relatively little and issues shares for compensation. Weak reinvestment leaves no cash-distribution cushion. | Per-share profit versus total profit, share count, repeated capex increases and management's explanation of expected returns. |
If AWS growth and margin step down after capacity enters service; free cash flow improves only because building is cut sharply; AI revenue remains concentrated in a few labs; and ads plus North America margin lose momentum, we would move from “returns await proof” to “the capital cycle has deteriorated.” We would not explain it away with “the long term will be fine.”
Regulation now touches the richer layers more directly. In addition to the existing marketplace antitrust case, the US Federal Trade Commission and 22 states filed a new complaint on 31 August 2026 concerning alleged advertising-auction surcharges. These are pending allegations, not established facts, but they show that scrutiny reaches Amazon's advertising and marketplace economics—not merely merchandise retail.
The final view
We accept Amazon's business quality and the reality of near-term AI demand. We do not automatically translate either into shareholder return. We would neither own it forever simply because it is a good company nor treat it as a short-term exercise in calling the top. A better approach is to leave room for uncertainty in the price, then let 2027–2028 capacity, pricing, margin and cash flow test the case. If “many orders” does not become “more free cash flow per share,” the logic has not completed.
09 · Sources, estimates and notice
Financial data are based on Amazon's Q2 2026 disclosures; regulatory developments are updated through 31 August 2026; valuation scenarios use a 14 August 2026 price snapshot. Market prices and consensus estimates change, so the scenario outcomes cannot be applied mechanically to today's price.
- Amazon Q2 2026 results: segment revenue and profit, cash flow, advertising and AWS growth, and the Anthropic revaluation.
- Amazon Q2 2026 Form 10-Q: full financial statements, risk factors, capital investment and accounting detail.
- AWS AI and custom-chip materials: Bedrock, Trainium, Graviton and enterprise AI products.
- US Federal Trade Commission Amazon case page: current marketplace, advertising and other enforcement matters.
Consensus convention: the $10.34 2027 EPS and $13.27 2028 EPS figures came from S&P Global Market Intelligence via Koyfin, captured on 12 August 2026. They are historical snapshots of subscription data, not Amazon guidance, and should be refreshed before use.
What is estimated: advertising's contribution to commerce profit, concentration of contracted revenue among AI labs, the valuation scenarios and normalised cash-flow judgement are not line items disclosed by Amazon. They are identified as estimates or scenarios in the text. This note is general research and information only. It is not investment or financial product advice, a recommendation, an offer or solicitation.