AI token growth, model efficiency and the GPU demand outlook

A user, citing Goldman Sachs research, describe total token demand growing 18-fold from December 2025 to September 2026, compared with roughly 8–9-fold for frontier models on the same index. The post argues that token growth alone cannot explain AI infrastructure needs: if more tokens come from models requiring fewer GPU cycles, GPU demand could grow much more slowly than token consumption. It also argues that falling token prices raise the bar for sustaining infrastructure investment. The key variables, in this view, are token volume, pricing and computing required per token.