Railroads had some of the same issue encountered by the dotcom bubble, and in particular dark fiber: Overcapacity.
Once railroads where demonstrated to be (extremely) profitable, people kept building them even if the existing demand was already met. The exact same thing happened with dark fiber (looking at you Enron).
AI is a bit different in the sense that the capacity can be used for research, and more importantly is not bound to a specific connection (between locations), and therefore a lot more fungible.
Not sure about canals, but while I think there are issue with current AI investments, I would not conflate them with the railroad bubble on this basis.
AI is a bit different in that unlike the rail tracks and the dark fiber the data centers arent actually going to be finished. It's much more of a stranded asset - there will be piles of chips and 1/4 finished construction projects, not usable assets.
No doubt somebody will use the assets for something but it's still wildly more wasteful than either railroads or dark fiber.