[sanderjd]

But that's current hardware. What about future hardware? What about hardware optimized for inference? What about hardware optimized to run a particular model?