AI model distillation, while compressing inference value, drives increased total token consumption and expands economically viable applications, leading to higher demand for physical infrastructure like GPUs, HBM, networking, and electricity. Frontier models depreciate rapidly, while physical infrastructure assets like datacenters and fabs offer longer-duration economic moats due to high capital and time barriers.
Model efficiency gains accelerate AI adoption, forcing hyperscalers to increase capital expenditure on physical infrastructure like GPUs, HBM, and power.