That's not at all what DeepSeek showed.

It showed that a single training run of a particular model could be done with a massive amount of compute, but not so massive that only OpenAI/Google/etc. can do it. Definitely not you nor I, or a university, or a mid-tier company.

In any case, that's small potatoes here. OpenAI spends most of its money not on training, but on inference. Inference is still way too expensive.

Even if inference got cheaper that's not great for OpenAI. It means that other people will launch a similar chat experience for fewer dollars.

Fundamentally, OpenAI needs a moat. And it has nothing.