Seven Mental Models to Understand the AI Compute Era

Most people watching the AI arms race are watching the wrong thing.

They track model releases. They count parameter counts. They read into benchmark scores. They watch Nvidia’s stock price as a proxy for everything else.

These are useful signals, but they are downstream of something more fundamental that almost nobody discusses in the right frame: the physical infrastructure — as explored in the economics of AI compute infrastructure — race happening beneath the model layer, and the structural logic governing who wins it.

Between Q1 2024 and Q4 2025, total tracked AI compute capacity grew 8.5× — from 2.5 million H100-equivalent units to 21.3 million. That expansion is not a story about chips. It is a story about power: who controls the substrate on which all AI capabilities run, who is building strategic independence, who is locked into a single supplier, and which organizations will still be relevant at the frontier in five years because they made the right structural bets in 2024 and 2025.