This is why they enforce the use of client apps like Claude Code/Codex (although OpenAI is a little more lenient). Trying to create a network effect.
Having a better model is the only real moat, without that inference is a commodity
The benefits are real and our willingness to pay is real, but the valuations only support one and right now even the free cheap models might win. Thusly the collapse could still happen.
They’re all stuck in a cycle of spending huge amounts of money on training just to stand still (in business terms).
In the long run, it can’t continue because it doesn’t make any sense