Honestly I think their approach is the long term most useful one. Being SOTA is probably very costly and difficult. Optimising all the smaller stuff that's real world useful for people seems like it would have better long term use.

On top of that I would imagine the research side of Google might disk over things from one modality being useful to another - something like an audio optimising or memory optimising for text to speech could maybe also be useful for translation or world models. Etc etc