I think we'll see

- better hardware

- more efficient model runtime algorithms/code

- smarter/more efficient models (same capability with less parameters)

So ideally these will all come together and help.