Really hard to do training hardware as the software stack is quite complex and needs to be general and flexible. Inference on the other hand is easier, which is

Really hard to do training hardware as the software stack is quite complex and needs to be general and flexible. Inference on the other hand is easier, which is what a lot of new hardware will be. I doubt if it makes sense for Tesla to build hardware for datacenter inference. Xai on the other hand could benefit from it.