> Also by the way, caching does not make LLM inference linear. It's still quadratic, but the constant in front of the quadratic term becomes a lot smaller.

Touché. Still, to a reasonable approximation, caching makes the dominant term linear, or equiv, linearly scales the expensive bits.