My guess is a nascent AI will live inside a giant supercomputer cluster somewhere, and will probably run on electricity. Which means that if its consciousness was interrupted by a power loss for even a microsecond, just like your regular computer will crash, an AI so deprived will die. It's possible an AI might be more robust than a biological mind, but I wouldn't count on it; deprive neurons of oxygen for but a few minutes and they're doomed. A superpowerful AI would presumably be a much more complicated thing than a computer OS, and my guess is it would react just as well to a power interruption as any desktop would if the power were disrupted.
Literally the only way an AI could destroy or seriously harm humanity is if it were hooked up to the world's power grid, or directly controlled nuclear reactors or nuclear missile launch codes. Yeah if we do something that stupid (looking at you Terminator 3) then it could destroy us. But heck, even a regular computer system that was so connected could wipe us out if it crashed or malfunctioned or was hacked.
So morale of the story? Let's deep-six the Internet of All Things while we still can. AI guys, knock yourself out and build the next Skynet - just make sure to surround it with an airgap. To the DoD, PG&E, et al, I say don't even think about hooking everything up to a single network!