- Model capability is converging: leading open and closed models differ by only a handful of index points while costs are 5–7x apart, making premium models uneconomic for many uses.
- Effective token prices fell ~40% since March while usage rose ~50% since July’s spending peak; price competition pressures model-lab margins, but demand still fills infrastructure.
- Open models lower token prices without lowering compute needs: a ~2.8T-parameter model can occupy nearly an 8-GPU B200 node, preserving demand for cloud infrastructure.
- The author favors hyperscalers and NBIS as tollbooths, but the thesis depends on demand outrunning price deflation; stalled growth, capex, or labs’ own infrastructure could break it.
Investment pitches
filter:
84