The AI party is still going — but the drinks got expensive
The hyperscaler AI data center boom is running into a very unsexy problem: the bill. Capex is expected to top $700 billion in FY2026, and the cost stack keeps climbing — GPUs aren’t cheap, NAND and DRAM are acting like they know they’re the new celebrities, and electricity bills are doing their best to become a line item with main-character energy.
ROI: the word everyone suddenly loves
For a while, the logic was simple: spend big, build fast, and let AI demand do the rest. But now investors are asking the awkward question at the back of the room: what if the returns take longer than expected? That’s putting pressure on hyperscaler margins and forcing some companies to delay or cancel data center projects rather than keep throwing money at shiny metal boxes.
Why this matters to investors
This isn’t just a nerdy infrastructure story. If the economics of AI deployment get shakier, that can ripple through:
- chipmakers tied to GPU demand
- memory suppliers like NAND and DRAM vendors
- power and cooling infrastructure names
- cloud hyperscalers absorbing the upfront spend
Big picture
The AI buildout isn’t dead — not even close. But the market is moving from “growth at any price” to “show me the payback,” and that’s when the easy money starts getting pickier.