The one truly hopeful aspect of a bubble-burst scenario is that extra capacity always finds a use-case, and in this case practically any other use-case would be both less harmful and more real.
IMO the most likely way to soak up the extra capacity is actually yet another iteration of AI rather than say, doing productive but boring work with any other techniques for curing disease or something. Still, a crash and a next iteration might be more likely to involve fresh ideas on architecture, or focus on smaller expert models that have less fake results and actually empower users. Right now I think there's a clear bias in research and execution. OFANG does want results, but also wants results that tech giants. Are subsymbolic techniques really the best techniques, or are they just the best at preserving the moat of big-data and big-compute?