I worry about Big-O almost every day. I run an analytics startup and scaling to huge datasets is important. Sure, I could throw hardware at the problem, but even that has its problems and bottlenecks and would also inflate cost. Instead, I carefully handpick the algorithms used (where possible) to have reasonable time/space complexities for the tasks I want to perform. Not everything needs to be carefully tuned, of course, but for the bits that do, Big-O is a useful tool.
Before this, I've used Big-O when writing some hobby game engine code to make sure that my algorithms scaled nicely as number of game objects increased.
I've also used Big-O to compare algorithms in back-end web development. Usually, I didn't care, but there were a few cases where it mattered (one specific one I remember was sending out notification emails based on a complex criteria and the operations I was measuring with O(n) was number of database queries).