If the inference is profitable and training new models is actually an endless arms race that's actually the best outcome for nvidia specifically.