So, uhm, not to dunk on Photon, but I've just built a proof of concept #SQLite database containing all of the addresses in OpenStreetMap and created a full-text search index over them.
A query like "give me all OSM objects where any of the address fields starts with 'Frankf'" takes 36 ms and uses 11 MB of RAM.
That's a FTS5 trigraph search over 13.5 million unique strings and then joining them with nodes & ways.
The database takes 25 GB, compared to Photon's 230 GB.
I'm impressed.