Amazon's biggest quarter of profit was mostly a number it marked itself.

The number the market repeated after Amazon reported on July 30 was $62.6 billion — net income for a single quarter, a figure large enough to end most arguments. The stock rose more than 11% the next session. The same day, Apple, which had also beaten estimates, fell hard, and the split was written up as investors finally "picking AI winners." It is a clean story, and clean stories move fast. This one is also mostly about a number Amazon calculated itself, on an asset it cannot easily sell, and it deserves a slower read than a one-day move allows.

Start with the composition of the profit, because the composition is the story. Of that $62.6 billion, $53.4 billion was non-operating, pre-tax "other income," and by Amazon's own disclosure the bulk of it was a mark-to-market gain on its stake in Anthropic — the AI lab it has backed with billions, and whose paper value rose over the quarter. A mark-to-market gain is exactly what it sounds like: the holding did not change, its estimated price did, and the accounting rules require the increase to run through the income statement as though it were earnings. No cash arrived. Nothing was sold. Bank of America had flagged the gain as likely weeks before the print, which tells you the sophisticated money knew to look through it. Not everyone who saw the headline did.

Strip the mark out and Amazon's operating income — the money the business actually made moving goods and renting servers — was $27.5 billion, up 43% from a year earlier. That is a genuinely strong number and would have carried the quarter on its own. It is also less than half the headline. The distance between the $27.5 billion the business earned and the $62.6 billion the market cheered is the distance between a cash result and a valuation, and Amazon reported both under the same word: income.

A mark-to-market gain is a price, not a payment. The market cheered $62.6 billion of profit; $27.5 billion of it was the business.

The circular part

The Anthropic mark has a feature worth naming, because it is the kind of thing that is invisible while it works. Amazon is both a major investor in Anthropic and one of its largest infrastructure suppliers — the lab runs heavily on AWS and on Amazon's own Trainium chips. Anthropic's rising private valuation, the thing that just added $53.4 billion to Amazon's reported profit, is itself driven in part by the scale of the compute it has committed to buy. The gain that flatters Amazon's earnings and the demand that fills Amazon's cloud trace back toward the same set of AI companies. None of that is improper; the accounting is by the book, and strategic investment in a key customer is an ordinary move. It is an observation about concentration. The profit, the growth, and the backlog increasingly lean on the same handful of names — and those names lean on each other.

There is a technical term of art for how this gain is measured that is worth translating, because it governs how much trust the number earns. A stake in a private company is a Level 3 asset: it has no market price, so its value is estimated from models and the terms of recent funding rounds rather than observed in a trade. A Level 3 mark is a considered estimate, not a quote. When such a mark contributes 85% of a quarter's headline profit, the quality of that profit is only as good as the inputs to the estimate — and the inputs are private valuations in the most enthusiastically bid corner of the market. That is not a reason to disbelieve the number. It is a reason not to treat it as the same kind of object as a dollar of AWS operating income.

The number that has to be paid

Now the other direction. In the same release, Amazon raised its 2026 capital-expenditure plan to roughly $220 billion, up from about $200 billion, and was unusually specific about the reason: the cost of memory. The AI build-out has run into a physical shortage of high-bandwidth memory and DRAM, the chips that sit beside the processors in every AI server, and the price of that memory has climbed hard enough to move a number as large as Amazon's capital budget by tens of billions. The same squeeze is running through Micron, Samsung and SK Hynix on the way up; it reaches Amazon as a bigger bill for the same build.

Capex is the opposite of a mark-to-market gain in the one respect that matters here. It is cash, it is committed, and it cannot be revalued away in a soft quarter. The $220 billion leaves the building whether or not the Anthropic stake holds its price. So put the two numbers where the quarter put them, side by side: a $53.4 billion paper gain on the AI trade, and a roughly $20 billion increase in the cash cost of building for it — booked in the same filing, pointing in opposite directions. One is a number the market handed Amazon and can take back. The other is a bill Amazon has already signed.

What the backlog does and doesn't derisk

The bull answer to the capex worry is the backlog. AWS reported about $496 billion in commitments — contracted future revenue — up roughly two and a half times a year earlier, and the division grew 37% in the quarter, its fastest in eighteen. Chief executive Andy Jassy told the call that Amazon will not have enough capacity to meet all its AI demand in 2026, and that he expects the same in 2027, with the "lion's share" of 2027 capacity already reserved and some of 2028 committed. The AI and chips businesses each now run above a $25 billion annual rate; advertising reached $19.8 billion, up 26%. The demand, in other words, is not the thing in doubt.

On a trading desk you learn to read a backlog two ways: as demand you have de-risked, and as capacity you can no longer flex. Jassy is selling the first reading, and for AWS's revenue line he is probably right. But "we have already committed capacity we cannot build fast enough" is also an admission that the $220 billion is not discretionary spending that can be throttled if sentiment turns. You do not get to spend it only in the quarters the story is holding. The backlog that de-risks the revenue is the same fact that locks in the outlay.

And the backlog's quality is a concentration question of its own. A $496 billion figure reassures in proportion to how many independent customers stand behind it. A large share of frontier-scale cloud commitments sits with a small number of AI labs — several of which Amazon also invests in, and whose valuations feed the very mark that produced this quarter's profit. If that is the shape of it, then the backlog and the Anthropic gain are not two separate sources of comfort. They are closer to the same bet, counted twice: once as contracted demand, once as investment income.

The downside no one's pricing

Here is the exposure, stated plainly. Mark-to-market cuts both ways by construction. The rule that let Amazon book $53.4 billion when Anthropic's implied value rose will, with exactly the same indifference, require it to book a loss if that value falls. Private AI valuations have moved in one direction for two years, and the accounting has quietly made a chunk of Amazon's reported earnings a function of whether they keep doing so. A market that spent late July, in Fortune's phrase, "in revolt over AI spending" is not a market that has ruled out a repricing. In the quarter that delivers one, the same line that added 85% to this profit can subtract from the next. Nothing about the cloud business will have changed. The reported number will move anyway, and the headline will read as a miss.

So the honest way to hold this result is in two hands. In one, a real business: AWS growing at its fastest in four years, $27.5 billion of operating income, advertising compounding, a backlog that says the demand is genuine and durable. In the other, a headline profit most of which is a self-assessed price on an illiquid stake, sitting one repricing away from running in reverse — booked in the same quarter Amazon committed another $20 billion of hard cash to the build-out that price is meant to justify. The market read the $62.6 billion and bought the winner. The number worth watching is the $27.5 billion, because that is the one that will still be there if the AI trade stops marking itself up.