The Bottleneck Moved Downstairs
Four months ago the hard part of building AI was getting chips. Today it is getting power and memory, and the bill has quietly moved from equity to debt. Here is the whole stack, read from the bottom up.
Most AI coverage starts at the top, with a new model and a benchmark score. This starts at the bottom, with a substation. Between May and September 2026 almost every constraint that mattered came from below.
The answer to "can we plug in" became no
Texas stopped connecting new large customers. 474 gigawatts are waiting in the queue.
The bottleneck moved from packaging to memory
Memory prices roughly doubled in a quarter. Even Nvidia could not absorb all of it.
Capex stopped being a demand story
Four companies will spend $725 billion this year. Part of that rise is just memory costing more.
The frontier settled on one price
Three frontier models from two labs all charge $50 for a million words of output.
Capability climbed. Adoption did not.
L5-L6 Tooling and apps
Agents got much better at real work. The share of firms actually using AI barely moved.
The money moved from equity to debt
$489 billion of AI borrowing this year. Lenders now carry risk that shareholders used to.
Europe got later and enforceable at the same time
The deadlines slipped and the fines switched on, in the same month.
What we still don’t know
Open questions
The things we could not confirm, and what would settle each one.
Strata, the whole AI stack, explained simply. Every number carries a source and a confidence label.
Today · The Stack · Stories · Archive · Glossary · Break the Chain · Focus · About · RSS feed
Privacy · Terms · Corrections · Report an error
Also from Dheeco: Dheeco · AP Fact Check
Published by dheeco.com · © 2026 Dheeco