> I can't see these earnings being sustainable for nvidia over the long run. We've barely scratched video and volumetric video training, even if text stuff hit

> I can't see these earnings being sustainable for nvidia over the long run. We've barely scratched video and volumetric video training, even if text stuff hits a data wall I think demand will stay high for quite some time due to that. The move to on-device inference actually takes more training, you need to train it longer to make it still perform well at edge parameter counts (with lots more training, Llama 3 8B is competitive with 70B). Nvidia is already more specialized towards training and has lots of competitors in inference. Training tends to have steps that don't just involve matmul; Jim Keller said Tesla Dojo ended up having problems from being too rigid, where GPUs have been able to adapt to new training regimes better, though Google's TPUs have seemed to not fall into that as much, maybe just iterated much faster?