> Unfortunately, we pay by the token, so I don't see the incentive for providers to spend time and money doing this for us.

Providing a better service, for one. Plenty of providers do offer caching, both input and output tokens, and usually give you a cheaper price for it too. Example from two of them: https://platform.claude.com/docs/en/build-with-claude/prompt... & https://api-docs.deepseek.com/guides/kv_cache