> Right now, we don't know if AI will be able to solve bigger problems, or if they'll be so efficients that you no longer need a whole datacenter for running them.

For the latter: I assume that having a whole data centre will always be an advantage. I am saying that for a fixed target, like proving the Rieman hypothesis from scratch, the required hardware will shrink.

And, yes, the Rieman hypothesis hasn't been proven yet. So to take your fears into account, replace my example with something they've already done, like constructing a solution to the Navier-Stokes-problem.

We don't know how good models will get, but the pattern of open weight models keeping up on a relatively short delay has been holding pretty well. And even among the proprietary models there's healthy competition. The concentration of resources is pretty well counteracted by these factors.