The current popular Ai models are only what works best on the current hardware.

We've already designed tons of different models that are outlined in many older Ai books, that can be used as compute scales (as Ai companies make more money to spend on more compute). Even the current models weren't invented recently, they're just now applicable because the hardware is there.

There's been a few algorithmic optimizations along the way a larger portion of the scale has been hardware.

2nd order companies are taking out 1st order companies by improving things, but that still keeps the ball moving forward.