Prompt caching: 10x cheaper LLM tokens, but how?

I'm captivated by the idea that semantic meanings can be mapped in high-dimensional space. Working to get a better handle on both the theory, and technical applications.