Cost is irrelevant. They don't overclock them because it invalidates their results, increases their electric bills, and makes bad hardware much harder to diagnose.
Demand for faster chips come from the development industries be it drafting, drawing, application development, or game design. Businesses are the largest purchaser of cutting edge technology. If your statement was remotely true, most processors wouldn't have a limited CPU multiplier. The only processors that are intentionally catering to overclockers is those with unlocked multipliers and, because changing the heatsink voids the warranty, they legally don't cater at all.
Moore's Law pretty much guarantees that faster processors will come out even if demand is minimal because it is nonsensical to release slow/older processors when new processors cost the same or less than older processors to manufacture.
We also can't forget AMD and Intel fighting for first place.
Users that overclock comprises of an extreme minority of the segment.
Intel/AMD marking their chips for a specific speed is the same strategy as FAA certifying aircraft for flight worthiness. These are known-to-be safe parameters and exceeding those parameters means heading into no man's land. When you are dealing with a scientific or experimental workload (I deal with both), hardware problems are the last thing you want to deal with. Being a sheep means being safe. Better safe than sorry (in my case, wasting hours diagnosing a software problem when it originated in the hardware).
To me, it is completely contradictory why people would want to overclock when contributing to a scientific project. Yes, you get more points but, what good is it if it is wrong? I wish I knew what BOINC used to catch and counter computational errors. I hope two sets of hardware process everything so if they don't match, it is performed a third time to figure out who is right.
