Someone explain how you'd create a vector embedding using homomorphically encrypted data, without decrypting it. Seems like a catch 22. You don't get to know the semantic meaning, but need the semantic meaning to position it in high dimensional space. I guess the point I'm making is that sure, you can sell compute for FHE, but you quickly run up against a hard limit on any value added SaaS you can provide the customer. This feels like a solution that's being shoehorned in because cloud providers really really really want to have a customer use their data center, when in truth the best solution would be a secure facility for the customer so that applications can actually understand the data they're working with.

Most of modern machine learning is effectively linear algebra. We can achieve semantic search over encrypted vectors if the encryption relies on similar principles.