Amazon.com, Inc.

I have called this company's founder "the most remarkable business person of our age," and I have called myself an idiot for never buying a share. Both statements were accurate. So let me describe, with the humility of a man who watched it from the very first day, what Jeff Bezos built in a Bellevue garage in 1994. Amazon is three great businesses wearing one trench coat. The first sells you nearly everything on Earth and delivers it tomorrow. The second is a toll booth: two million independent merchants pay Amazon roughly half of every sale for the privilege of using its shelves, its warehouses and its front door. The third — Amazon Web Services — rents computing power to much of the modern economy, and it alone produces more than half the company's profit.

The numbers are staggering even by the standards of our giant era. In the year ended December 2025, Amazon booked $716.9 billion in revenue — more than any company we have ever examined — and $77.7 billion in net income. Roughly 240–250 million households (approx.) pay it an annual membership fee just to shop there more conveniently. Its logistics network delivers more parcels in America than UPS or FedEx. And it is now spending money on artificial-intelligence infrastructure at a pace — about $200 billion guided for 2026 — that has no precedent in the history of private enterprise.

"Your margin is my opportunity." — Jeff Bezos

That one sentence explains everything in this report. Where other companies defend fat margins, Amazon deliberately runs its retail store near breakeven — pricing so low, and delivering so fast, that competing with it is unattractive — and then earns its real profits around the store: on the tolls it charges sellers, on the advertisements it sells against its own traffic, and on the cloud infrastructure it rents to everyone else. It is the flywheel Bezos famously sketched on a napkin, and it has been spinning, faster every year, for three decades.

The road matters more here than for any company on our board, because Amazon's history is a repeating pattern: accept years of losses Wall Street hates, build something nobody can replicate, then harvest. It has run this play at least four times. Knowing the pattern is essential to judging the fifth attempt, now underway.

Two lessons for an owner. First, the pattern has always paid: the market called the fulfillment build-out wasteful in 2016 and the AWS build-out speculative in 2012, and both became the widest moats in their industries. Second — the caution — the pattern once nearly killed the company, and as recently as 2022 the same "demand justifies doubling capacity" logic produced an overbuild that took two painful years to digest. The fifth bet is by far the largest. History says trust the builder; history also says check the price of the bricks.

The store, any ten-year-old understands — better than most analysts, in fact, since the ten-year-old actually uses it. The consolidated machine takes more work. Here is the flywheel:

Full disclosure: this business sat outside my circle of competence for twenty years — not because I couldn't understand the store, but because I couldn't price the reinvestment. That has changed in one important way: Amazon now reports enormous, visible profits ($80 billion of operating income in FY2025), so the earning power is no longer a matter of faith. What remains a matter of faith is the next $200 billion. That is the honest boundary of the circle today.

Amazon reports seven revenue lines, but read them as one machine with a loss-leader at the front and toll booths behind it. FY2025 revenue by line:

Who subsidises whom. For most of the 2010s, AWS carried the company while retail ran at breakeven — the cloud paid for the warehouses. Today the honest map is subtler: within retail, the seller tolls, the advertising and the Prime fees are the profit, while first-party retail is closer to a traffic-acquisition cost. And the arithmetic of the tolls is remarkable: between referral fees, fulfillment, storage and increasingly indispensable advertising, a typical independent seller hands Amazon roughly half of every dollar of sales (approx. — Marketplace Pulse, Bloomberg). That rising take rate is the heart of the FTC's antitrust case (Part XI) — and of the profit recovery: North America segment margins reached record levels in 2025–26 on regionalised logistics and ad mix. One more beautiful detail for the connoisseur: Amazon operates on a negative cash-conversion cycle of about −51 days — it sells your goods and holds the cash for almost two months before paying its suppliers. Its customers, in effect, finance the machine.

Amazon's moat is filled from four springs, and it is worth being precise about each, because they age differently:

Scale and cost — the fulfillment network. The closest thing the modern economy has to a railroad. Amazon delivers more US parcels than UPS or FedEx, from a network of warehouses, planes, trucks and over a million robots that took thirty years and hundreds of billions to build. Nobody — not Walmart, not anyone — can replicate it at national scale, and unlike a railroad it is self-funding: the sellers pay fulfillment fees for the privilege of riding it. Trend: widening.

Network effects — the marketplace. More sellers → more selection → more buyers → more sellers. Roughly 62% of units sold are third-party, and more American product searches begin on Amazon than on Google. Trend: stable, but contested at both ends — Temu and Shein attacked the cheap end (blunted, for now, by tariffs), and AI agents threaten the discovery end, which is where a marketplace's gravity lives.

Switching costs — AWS and Prime. Moving production workloads off AWS is expensive, risky and slow; the AI era deepens the lock (data gravity, Bedrock, Trainium-optimised models). Prime is subtler: a paid habit, renewed by ~240 million households, that converts shopping from a decision into a default. Trend: widening.

Brand. Not aspiration — reliability. The promise is "it arrives tomorrow and returns are painless," kept billions of times a year. The caveat: customers love Amazon; many of its counterparties — sellers squeezed by tolls, unions, regulators — resent it, a Standard Oil profile that invites the political attacks of Part XI.

Where it could crack. One fault line matters above all: if AI shopping agents become where purchase decisions happen, Amazon's storefront could be disintermediated even while its warehouses still deliver the box — the tolls, the ads and the Prime habit all live at the front door, and a company reduced to logistics-for-someone-else's-agent would be a far poorer business. Amazon knows it: it has blocked the crawlers, sued the agents, and is racing to build the shopping agent itself (Part VI). I score the moat a 9, not a 10: physically the widest we have ever examined, but guarding a front door that technology is trying to move.

Here is the number that dominates everything: Amazon spent $131.8 billion on capital expenditure in 2025 — and guided to roughly $200 billion for 2026 (approx. — Feb 2026 guidance), the vast majority for AWS and AI. For scale: that single year approaches double the peak annual spend of the Apollo program, inflation-adjusted. Free cash flow — the metric Bezos taught two generations of investors to worship — has gone negative under the weight (Part IX). What is it buying?

The Anthropic axis. Amazon has committed roughly $33 billion (approx.) to the AI lab Anthropic, which in turn has committed $100+ billion over ten years to AWS and trains its Claude models on more than a million of Amazon's custom Trainium chips — including Project Rainier, an ~$11B datacenter campus in Indiana. The custom silicon is the strategy in miniature: owning the chips converts Nvidia's 70% margins into AWS's cost advantage. Your margin is my opportunity, applied to the most profitable company on Earth.

An honest accounting note before the scoreboard. FY2025's $77.7B net income includes roughly $16.8 billion of non-operating investment gains — largely paper marks on those same AI stakes (operating income was $80.0B against $96.8B pre-tax). The circularity deserves naming: Amazon invests in Anthropic → Anthropic spends it on AWS → AWS books revenue → Amazon books investment gains. Real business flows through that loop — but so does some self-referential accounting, and a careful owner reads the operating line, not the headline.

On the consumer side the fight is defensive and offensive at once: Amazon blocked OpenAI's crawlers (removing ~600M products from ChatGPT shopping), sued Perplexity over its agent shopping on the storefront — and won a blocking order in March 2026 — while merging its own Rufus assistant into "Alexa for Shopping" for every customer. The framing I keep returning to: Amazon owns fulfillment and the purchase graph; agents may come to own intent. Whoever owns intent taxes the other. Amazon is spending $200 billion, in part, to make sure it is not the one being taxed.

Read the cloud row with both eyes open, because both things it says are true. AWS grew 28% last quarter — its fastest in fifteen quarters — an astonishing re-acceleration for a $150B-run-rate business. And yet its share of the cloud market has drifted from the low-30s to about 28%, because Azure and Google Cloud are growing faster still in the AI cycle, where every workload is contested from scratch. In retail the picture is calmer: Walmart is a permanent, disciplined #2 whose real edge is grocery; Temu and Shein turned out to be regulatory arbitrage that Washington closed; and the tariffs that hurt them also squeeze Amazon's own China-based sellers — margin pressure inside the toll booth. The agents are the wildcard, and they are covered where they belong: at the moat's front door (Parts V–VI).

Two men define this company's management story — one running it, one watching over it with nine percent of it.

Now the part a dividend-minded reader must hear plainly. Amazon has never paid a dividend — not once in twenty-nine years as a public company — and its buyback record is a token: a single $10B authorisation in 2022, of which $6B was ever used, while stock-based compensation has pushed the share count up ~12% over the decade (9.5B → 10.8B shares). Contrast Apple, which retired a third of itself. Every dollar Amazon earns is reinvested — into warehouses, then cloud, then satellites (Project Leo), robotaxis (Zoox), and now $200B of AI plant. For thirty years that reinvestment has passed Buffett's one-dollar test with the highest marks in modern history: each retained dollar became roughly $14 of market value (approx. — $281B of cumulative retained earnings against a $2.66T market cap). But a fresh dollar retained today goes into the largest, least-proven bet the company has ever made — and the institutional-imperative flags (satellites, robotaxis, a $13.7B grocery chain still unsolved) are real. I score management 8: superb operators and honest letters, with empire-sized appetites that demand watching.

Study the trajectory before the snapshot. For its first twenty years Amazon chose to show no profit; the past three years are what the machine looks like when management lets margins breathe — operating income roughly quadrupled from the 2022 trough to $80 billion, and margins hit all-time highs. Then read the two ruby rows, because they are the whole debate: capital expenditure now consumes 101.7% of operating cash flow, so on a trailing basis this $2.7 trillion company generated less than nothing in free cash. In 2025 it was $7.7B (on $139.5B of operating cash flow); in 2021–22 it was negative; at ~$200B of 2026 capex it will be deeply negative again. The company that wrote the book on free cash flow per share is asking you to ignore the book for a few years. Whether you should is the valuation question.

We have now seen this movie three times — Microsoft's ~30%-of-revenue capex, Alphabet's $185B — but Amazon is its extreme form: the only company on our board where the AI build-out pushed free cash flow below zero. So the automated DCF produces its most broken answer yet: an "intrinsic value" of $75, implying the stock is 70% overvalued. Ignore it with confidence — a model that capitalises a construction year as a permanent condition is measuring the scaffolding, not the building. Wall Street, looking through to earnings, lands at a $308 mean target (+25%), with 83 of 94 analysts saying buy. The sensible anchor is the earnings path: $8.83 next year, $13.04 by FY2028 (consensus), which prices today's $247 at ~19× two years out — a fair, unheroic price for a business compounding revenue at 10%+ and earnings at ~20%. Fair, mind you — not cheap: the margin of safety only appears on weakness.