> if you own the entire training to inference pipeline i'm not sure what this offers

336x faster than Python, and swapping backends in a production environment can be far from trivial