My best guess is that for scaled RL most of the top Chinese AI labs are starting to use a lot of Huawei for inference and Nvidia for training (maybe not for weird architectures). As agent swarms, even more scaled post-training, etc becomes the norm, this will accelerate their domestic industry.