fiber buildout of the 90's. The amount of fiber they laid turned out to be massive overkill because of gains in throughput per fiber
But web traffic eventually grew to use up most that excess. It just happened after the investors got screwed due to bankruptcies (and picked up by other co's during fire-sales).
Same with the railroad boom of the late 1800's: overbuilt, bubble popped, investors got bleeped, but the rails eventually ended up being used as the population of the US gradually increased.
Thus, AI will likely follow a similar pattern. Right now big-tech is fighting over AI market-share, subsidizing their projects. Such corporate subsidies can't last forever, so it's a financial game of chicken. When investors get a clue and most AI projects are cancelled or sold to the winners, then the excess capacity may gradually get used for actual profitable products.
One problem though is that the motherboards may be obsolete by then. But at least the rack & wiring is already built.
(Some of my own stocks ended up partly tied up in this, so I have to ride along...)