Even the stereotypical "impractical" research algorithms (with efficient complexity but huge constants) are implemented. That's the magic of computer science, that you can breath life into abstract ideas. But if you cannot implement it at all, there are millions of other cool ideas that you can implement, so you may gravitate towards them. (another argument is that if you can't implement it, it's classed as pure maths. And if you can't prove it, or define it clearly as a conjecture, I'm not sure what it is or what you could do with it - perhaps futurism/science fiction, which has its place.)
The value of implementation is due to computer science ideas tending to be complex and abstract - but once you have an actual concrete working model that you can play with, they can become almost trivial to understand. And so, comparatively, more progress in research is made based on working code than on purely theoretical ideas. It's so much easier to find mistakes in an idea if you can build it. And once your theory is embodied in living, working code, it's tremendously easier to reason from that as a base; much as it is easier to see what is over the next hill after climbing it.
Hardware is closer to your idea of incentive: the inventor of the Wallace tree multiplier told me that it was a simple and obvious idea, but it wasted a lot of silicon; and no one entertained that kind of waste. It simply wasn't conceivable. Perhaps this is due to the incredible rate of improvement in silicon (Moore's Law), so that it's psychological difficult to appreciate (or even believe) that your iPhone will be 1000 times faster in 15 years; or what changes that even means for algorithms. Perhaps the most basic assumptions of the algorithm changes? When x10 quantitative improvement often translates into a qualitative improvement, what does a x1000 change mean? Too hard to predict.