The companies that are illegally training on copyrighted data that can keep moving forward with an obfuscated dataset will hang in there. The ones who can't-or get sued into oblivion-will eventually just get acquired or give up. If "centralized" means anything in this arena, it's the generalized training data, yes.

Think of it like this: all companies want these "AI" platforms for is to make their own data more easily parsed and accessible, right? The ones that have engineering resources may be paying for OpenAI now, but once the tooling in the FOSS side is a bit more complete, don't you think these customers of OpenAI would rather just host their own and run off their own trained data? That's where things are already shifting.

This all follows a pattern that has happened time and time again. Last decade it was all the stupid "smart" assistant craze (those are all dead, btw), and now it's this stupid thing. Nothing new to see here.