Reach out if you are interested in any way.
The process you describe would work to produce hardware that's best for some specific type of ML models, while the whole point of this article is that the currently "best" ML models are there not because they are objectively best, but rather because they are what's best within the bounds of the hardware architectures we have (i.e. we're not looking at best ML algorithms but rather at ML algorithms that "won the hardware lottery"), and it seems plausible that there might be better ML methods if there was different hardware that enabled them - but the process you describe wouldn't possibly find that different hardware, it would instead tune the hardware even more towards the algorithm path we're already stuck on.
GPU development seems driven by more general computational principles that might be summed up glibly as "We're hitting fundamental physics limits for single-core processors, what is the maximum amount of data we can move per second through the maximum number of cores?" Perhaps there is a way to extend this with a view toward the challenges of current ML model training and inference, but I imagine trying to approach it as a black-box optimization problem could be quite difficult.