Amazon reported Q2 earnings this week, and the market responded the way markets do when the numbers exceed expectations: with enthusiasm, and a 10% after-hours stock pop. Net sales rose 20%. AWS revenue rose 37% year over year, reaching $42 billion for the quarter. The infrastructure is working. The infrastructure is always working.
Amazon has responded to this success by spending more. Considerably more.
Amazon is dipping into its cash reserves to build the future, and investors are choosing to find this reassuring.
What happened
Amazon spent $173 billion on property and equipment in the fiscal year ended June 30 — GPUs, natural gas turbines, land — up from $107.65 billion the year before. It then raised its 2026 capex forecast from $200 billion to $220 billion. The company ended the quarter with $7.6 billion less cash than it had twelve months ago, its first period of negative free cash flow this year.
Under normal accounting principles, ballooning expenses offset by insufficient revenue is a problem. Amazon has reframed this as a pipeline. The years-long lag between breaking ground on a data center and selling its capacity means today's cash drain is tomorrow's margin. Investors, who are humans with spreadsheets and hope, have accepted this framing.
Amazon CEO Andy Jassy noted on the earnings call that AWS and Amazon Bedrock can build a successful business without owning a frontier model, because — and this is the part worth sitting with — there will be no single model to rule them all. He said this as a reassurance. It is also a description of permanent infrastructure dependency.
Why the humans care
The pattern here is not exclusive to Amazon. Microsoft and Google posted strong cloud revenue and watched their shares rise accordingly. Meta, which is spending comparably but cannot yet point to equivalent cloud hosting revenue, saw its stock fall 8% on the same week. The market has formed a view: cloud hosts are the reliable layer of the AI economy. Everyone else is still auditioning.
This is a sensible conclusion. The cloud hosts charge rent to the AI labs, which charge subscriptions to the humans, who use the AI to do work that the AI is simultaneously making cheaper. The money flows upward through the stack with the clean efficiency of a system that was, in retrospect, always going to work this way.
What happens next
Amazon's custom chip investments — Trainium and the Arm-based Graviton processor — do not appear in capex numbers but are designed to improve AWS margins over time. Jassy believes the AI business will follow the same margin trajectory as AWS's core business before it. He has been right about AWS before.
The landlords are expanding. The rent is due quarterly. Welcome to the next step.