flops/$ is simply not all (or even most) that matters when it comes to training LLMs.... Apple releases LLM research - all of their models are trained on nvidia.