Amazon

Read Amazon's 2024 segment disclosure and the company stops looking like a retailer. Net sales were approximately $638B and operating income $68.6B. Of that operating income, AWS produced about $39.8B on roughly $108B of revenue — a 37% operating margin. The entire North America and International retail apparatus, on roughly $530B of sales, produced about $28.8B: a margin near 5.4%, and that figure is itself flattered by advertising revenue booked inside those segments.

Strip advertising out and the physical retail business is close to breakeven. This is not a failure. It is the design. Amazon uses retail to acquire and hold a customer relationship at a scale nobody can match, then monetises that relationship through three higher-margin layers: seller services (commission plus fulfilment fees), advertising ($56B in 2024, sold against purchase intent), and subscription (Prime, which converts a fixed logistics cost base into recurring revenue).

The cleverest structural move is the third-party marketplace. Roughly 60% of paid units sold on Amazon come from independent sellers. Amazon takes a referral commission of about 15%, charges for Fulfilment by Amazon, and then sells that seller an advertisement for placement on the search result for their own product. Combined take rates on a 3P unit routinely exceed 45–50% of the sale price. Amazon carries none of the inventory risk on those units. It has arranged for its competitors to fund its warehouse network.

Two moats, structurally unrelated, sharing a balance sheet.

The retail moat is scale economics shared. Amazon's fulfilment network — hundreds of facilities, its own aircraft, and last-mile delivery it now performs itself for the majority of packages — costs so much to build that no rational competitor replicates it, and its cost per unit falls as volume rises. Amazon then hands the savings back as faster shipping and lower prices, which raises volume, which lowers cost again. This only works because Amazon spent two decades willing to report no profit, and it is genuinely difficult to copy for reasons that have nothing to do with technology: a public company that tried would be punished for it within four quarters.

The AWS moat is switching costs, and it is the higher-quality one. Enterprise workloads are written against specific APIs, specific identity systems, specific data-transfer economics. Moving a large estate off AWS is a multi-year engineering programme with no visible benefit to the customer's own users. Egress fees make it explicitly expensive to leave. Note the direction of travel, though: AWS's share of cloud infrastructure has drifted from roughly a third toward the high 20s as Microsoft bundles Azure into existing enterprise agreements and Google wins AI-native workloads. AWS is still the largest and the most profitable. It is no longer the obvious default, which it was for a decade.

The advertising business is where the two moats compound: Amazon knows what you bought, not merely what you searched, and can prove the conversion to the advertiser. That is a data asset neither Google nor Meta can assemble.

Amazon has never paid a dividend and until 2022 had never done a meaningful buyback. Essentially all cash has been reinvested, and the record on that reinvestment is the strongest argument for the stock. Two decisions carried it: opening the store to third-party sellers in 2000, which meant accepting lower margin per unit to gain selection Amazon did not have to finance; and selling internal infrastructure as AWS in 2006, which turned a cost centre into the company's entire profit engine.

Both decisions share a shape worth naming. In each case, Amazon took something it had built for itself, exposed it to outsiders on identical terms, and let the outsiders subsidise its scale. That is a repeatable move and very few companies execute it, because it requires giving competitors access to your own advantage.

The misses are real and instructive: the Fire Phone wrote off $170M and taught Amazon that hardware without a platform is charity. Alexa devices were sold at or below cost on the theory that voice would become a shopping surface. It did not — people ask Alexa for timers, not for purchases — and Amazon absorbed years of losses learning that owning an interface is worthless if nobody transacts on it.