Only because of brute force over efficient approaches.

Again, look up Deepseek's FP8/multi GPU training paper, and some of the code they published. They used a microscopic fraction of what OpenAI or X AI are using.

And models like SDXL or Flux are not that expensive to train.

It doesn’t have to be this way, but they can get away with it because being rich covers up internal dysfunction/isolation/whatever. Chinese trainers, and other GPU constrained ones, are forced to be thrifty.