Your example graphic is kind of outdated, but kinda answers the question at hand. Transistor count increases with the tech-ups (getting to a lower nm fab process) but once certain caps are hit, the increase in performance is no longer in-pair with the transistor count increase and that's the sign that a new architecture is needed. The story itself is quite long. An example of what I'm saying was the Athlon architecture performance jump or the i-core one. In both cases, new archtecture designed prevailed vs. increase in frequency (in 1st case) or number of cores and frequency (in second case). Your graph shows exactly a cap being hit, where the architecture becomes inefficient to keep-alive / maxed out.
So basically, more bullets in a gun are good, as long as the enemy gun does not out-range yours. When that happens, it's time that you re-design your gun.