Investors love AI, as long as you’re a cloud host

Amazon posted a blowout quarter on Thursday, and Wall Street responded the way it usually does when the numbers look this good: it sent the stock up nearly 10% in after-hours trading. Net sales climbed 20%, and AWS alone brought in $42 billion, up 37% year over year. That's the kind of growth that makes eye-watering spending look justified rather than reckless.

And Amazon is spending eye-watering amounts. Property and equipment costs — the bucket that includes GPUs, natural gas turbines, and raw land — hit $173 billion for the fiscal year ended June 30, up from $107.65 billion the year before. The company also bumped its 2026 capex forecast from $200 billion to $220 billion, and it's now burning into cash reserves to do it. Amazon ended the quarter with $7.6 billion less cash than a year earlier, its first stretch of negative free cash flow this year.

Normally that kind of expense growth spooks investors. But AWS's revenue trajectory offers cover, even if the raw math doesn't yet balance the ledger. Given how long it takes a data center to go from groundbreaking to generating revenue, seeing demand climb in lockstep with supply is exactly what shareholders want to see. CEO Andy Jassy also pointed to quieter bets — the Trainium chip and the Arm-based Graviton processor — that don't show up in capex figures but could fatten AWS's margins over time. His view, stated plainly on the earnings call, is that Amazon doesn't need to own a frontier model to win, because he doesn't think one model will dominate everything.

Microsoft and Google told a similar story this week, with strong cloud numbers lifting their shares too. Meta didn't get the same treatment — its stock fell 8% after investors zeroed in on heavy spending with no comparable cloud-revenue offset. The pattern is becoming obvious: markets currently trust the infrastructure layer of AI far more than they trust the labs and startups actually building the models.

But that trust rests on a shaky assumption. Amazon's hosting revenue is, in a lot of cases, simply someone else's AI bill — with Anthropic, it's literally circulating the same dollars back and forth. If the economics don't hold up for the labs footing those bills, the cloud hosts won't be shielded just because they sit one layer removed. Everyone in this stack, no matter how differentiated, is ultimately betting on the same unanswered question about whether AI demand justifies the money being poured into it.