I think you are vastly underestimating the benefits of optimizing specifically for the existing workloads at the scale of AWS.
POWER from this perspective, does way too much. It's too much chip for AWS's needs, and a big waste of money.
That's what optimization means! You get the smallest/ simplest/cheapest thing that does exactly what you want, and ignore the big bad "do everything" chips that go for megabucks.
Then you can devote chip transistor count toward the instructions you want, and ditch all the dead weight that comes with a traditional chip like power, that tries to make a balanced general purpose cpu.
This results in a chip that isn't good for anything but the specific need, but it's super cost effective and higher performing.