Premise: A.I. used by developers to write code is in the final stages of being good enough to replace humans, after which the expense to improve the models hits a diminishing margin of return. The market will have a handful of essentially equal competitors and it becomes a race to the bottom for pricing. The breakthroughs after that is the A.I. surpassing humans at system architecture, system administration and inventing more efficient programming languages that humans don't understand.
Result: The compute used to train programming models is reallocated to train the next occupation that A.I. replaces. This becomes a snow ball effect of building on top of existing functionality where each new occupation is replaced more quickly than the ones before it by super agents.