I think you slightly misunderstood, and I wasn't clear enough—sorry! It's not a 30-70% speedup; it's 30-70% more cost-efficient. This is mainly due to non-NVIDIA chipsets (e.g., Google TPU) being cheaper, with some additional efficiency gains from JAX being more closely integrated with the XLA architecture.
No, we haven't run our JAX + XLA on NVIDIA chipsets yet. I'm not sure if NVIDIA has good XLA backend support.