To try to come up with an example, let's say we set as our goal to completely automate a process X, which consists of 10 subprocesses. Let's say we fairly quickly automated steps 1-9, but the 10th is tricky.
But we now realize the 10th step was only really necessary for certain edge cases, which we now realize we are fine not handling. So we "if" them away and now have a process that is 100% automated, even though it is different from what we originally wanted to achieve.
Perhaps a bit of the magic and allure disappears by pulling back the curtain: it sounds like an instance of analyzing and breaking down the problem into smaller ones; solving those pieces as you go along; further breaking them down as necessary; and tossing aside the nuts that are too tough to crack.
You can already buy cheap but powerful old servers. But newer hardware tends to be more power efficient. So, depending on time horizon you consider, it might be cheaper to buy newer hardware.
Assuming that GPUs power efficiency will increase, the same will be true about them.
Looking for (maybe hidden) stocks in systems also changes how one sees the world.