>they already have developed their own CUDA-like support, that is not 100% compatible, but works well enough for their internal uses at Alibaba.

I wonder whether that's inference or training.

not sure about everyone else, but i can’t justify sending my data workflows to china.