Comment by bel8

Nvidia GPUs aren't even the absolute SOTA for LLM inference anymore. Some labs are moving to ASICs and Huawei already have their own custom chips running DeepSeek as we speak.

There's enough money and scale on the line that software affinity like CUDA is no longer the deciding factor and there's margin for custom stacks.

Even more so after the USA GPU exports ban which is proving to have backfired by speeding up China's tech growth.