Comment by _ntka

It's all about cost and performance. If you can train a foundation model 2x faster with JAX on the same hardware, you are effectively slashing your training costs by 2x, which is significant for a multi-million dollar training run.