Amazon.com,
Inc.

The trailing multiple looks cheap against a five-year average roughly three times higher, and nearly all of that gap is an artefact: two enormous Anthropic revaluations sit inside trailing GAAP earnings and flatter the denominator. On the adjusted trailing figure in the panel the shares trade in the low thirties, a discount to their own recent history and close to the forward multiple.

Amazon runs three segments and one profit engine. Cloud contributes under a fifth of revenue and well over half of operating income, and the ratio has been widening. North America is a large, low-margin machine that has become steadily more efficient, and International is a thinner version of the same thing with a longer runway and a weaker margin base. The reporting structure hides the two businesses that matter most to the margin trajectory: advertising sits inside the retail segments and carries incremental economics closer to cloud than to logistics, and the custom silicon programme sits inside AWS and is the main lever on cloud gross margin.

The funding chain runs in one direction. Retail generates working capital float, because suppliers and third-party sellers are paid after customers pay, and that float plus cloud operating income funds the capital programme. It no longer funds all of it. The capital outlay in the anchor quarter alone exceeded the entire operating cash flow of that quarter, which is why free cash flow turned negative, and the shortfall has been bridged with a step change in long-term debt issuance rather than with retained cash. That is a deliberate choice, and it is the single most important fact about the current financial structure.

Operating leverage originates in three places, and the order matters. First, in cloud, where each incremental workload runs across infrastructure whose cost per unit of compute the company is actively lowering through its own chips. Second, in advertising, where inventory is already paid for by the retail business and the incremental cost of a sponsored placement is close to zero. Third, in fulfilment, where regional network redesign, robotics and delivery density lower cost per unit shipped without adding revenue. The first two show up in margin quickly. The third shows up slowly and is the reason North America margin has roughly doubled over three years. Against all of that sits the depreciation schedule on the current buildout, which has begun to land and will keep landing through the end of the decade.

Rate the mechanism, not the brand. Fulfilment density earns the top mark because it is arithmetic: the cost of the next delivery falls as the number of deliveries in a postcode rises, and no competitor has the volume to match the denominator. Walmart is the only firm close, and it is closer in groceries than in the long tail. The advertising pillar earns the same mark for a different reason. Amazon observes what people buy rather than inferring what they might want, and an advertiser paying for a placement adjacent to a completed purchase is buying a different product from an impression on a social feed. Neither pillar depends on being liked.

Cloud switching costs are strong and slowly weakening. Data gravity is real, committed-spend contracts are real, and the cost of rewriting an application estate around a different set of managed services remains prohibitive for most enterprises. What has changed is that model access has become portable. A customer choosing between frontier models can increasingly run the same model on more than one cloud, which moves the competitive question from lock-in to price and capacity availability. That is a worse fight to be in, and it is why the silicon programme matters beyond its own revenue line.

Custom silicon gets three stars rather than four because the advantage is contingent. It works if the chips stay a generation behind the merchant leader on performance while sitting far ahead on cost per unit of useful work, and if the two largest customers keep committing capacity at scale. Both conditions currently hold. Neither is structural. The Prime bundle earns three for a related reason: retention is high, but the bundle has been repeatedly widened with new content and services, which is a sign that the pricing power is being maintained by adding value rather than by charging for it.

Sequencing decides everything here, and the sequence has one gate. Silicon must work first. If Trainium delivers acceptable performance at a materially lower cost per unit of useful work than merchant accelerators, then the capital programme buys more sellable capacity per dollar, cloud gross margin holds while volume compounds, and the backlog converts at margins that justify the depreciation. If it does not, Amazon is buying the same accelerators as everyone else at the same prices, the buildout becomes a commodity capacity race, and the returns compress toward the cost of capital. Everything downstream depends on which of those is true, and the evidence so far points the right way without being conclusive.

Advertising is the second gate and a much easier one. It is already large, already growing faster than the company, and already the reason consolidated margin can expand while retail volumes grow at single digits. It needs no capital. Its practical constraint is inventory quality, which is why the push into live sport, Prime Video and third-party placements matters more than it looks.

Leo and Zoox are options, not drivers. Both consume cash now and will not contribute meaningfully inside the horizon of this report. They earn the emerging tag because each has cleared a real gate this year, orbit count for one and a federal exemption for the other, and because each attaches to distribution Amazon already owns. Rate them as call options funded from the retail float, and price them at close to nothing until the capital programme has visibly turned.