Most of the AI conversation is about software: which model is smartest, which lab is ahead, whose chatbot writes the best email. That framing is comfortable because software feels limitless — copy it a million times for free, scale it overnight, and the best one wins everything.
It’s also the wrong frame for understanding what’s actually happening — and where the money is moving.
The binding constraint is physical
A frontier model is trained and served on machines that need three deeply physical things: electricity, advanced chips, and the memory and packaging that tie them together. None of those scale like software. You can’t copy a power substation. You can’t download more high-bandwidth memory. You can’t will a new fab into existence this quarter. Each takes years, capital, and rare expertise.
So while the software narrative races ahead, the real build-out moves at the speed of the slowest physical part. That slowest part — the binding constraint — is what sets the pace for everyone. And it moves.
Why this matters for who wins
When a part of the chain is scarce and concentrated in a few suppliers, those suppliers get pricing power. Not because they have the best story, but because the laws of supply and demand are on their side for a while. The interesting question is never “which AI is smartest?” It’s “what is the system short of right now, and who owns that?”
That question has a different answer every few quarters. A year ago it was raw compute. Then memory. Then packaging. Increasingly it’s power and the grid. Each shift quietly reshuffles who benefits — usually before the headlines catch up.
How we think about it
THE ENTITY is built on this premise: model the AI build-out as what it physically is — a supply chain of connected, constrained layers — and watch where the pressure sits. Not to tell you what to buy (we don’t, and we won’t), but to make the constraint visible, traceable to real companies, and early enough to matter. Treat AI as physics, and a lot of the noise falls away.
See it live on the map
Plans from $19/mo — cancel anytime. Two minutes to your first read.