It's just the natural evolution of tech towards higher levels of abstraction. In the beginning most dev was on CUDA because the models had to be built and trained.

But since there are plenty of more advanced models now, the next level is getting built out as more developers start building applications that use the models (e.g. apps using GPT's API).

So where 5 years ago most AI dev was on CUDA, now most is on the LLMs that were built with CUDA to build applications.