In an AI compute short squeeze, everyone in the stack looks smart. That doesn’t tell you who earns durable long-term returns. You can be right on demand and still own the wrong businesses. Today we estimate that every incremental GW of data center capacity currently generates $40-50bn of revenue across the stack. Data centers are under-earning (~$2.2mn/MW) relative to the value they unlock at the token factory layer (~$40mn+/MW). Frontier models face real pricing pressure from open-source and the hyperscalers. A 40% token-price cut can push standalone inference near breakeven — while hyperscalers still make money on the hardware. The picture will look very different in 6–12 months. Energy capacity, tokens per watt, model-layer competition, and financing are all in motion. We built a dynamic tool to track how it plays out. This map will look different in 6-12 months. Four forces to watch: - Energy, the most acute bottleneck today - More tokens per watt from better hardware and software - Rising competition at the model layer - Financing all this capex So we open-sourced our model. Input your assumptions and watch how value shifts across the stack.