Intel is the one that has been underestimating ARM. And it is coming back to haunt them, now.
We won't truly know the answer to this one, but one thing is for sure. When they were touting Moorestown, the predecessor to Medfield, they didn't have enough experience. Things change when you go from paper to real designs. With Moorestown the actual power usage might have been much higher than on paper.
Important thing is they seemed to have learned from that. Just like they learned ever increasing pipeline stages with Netburst and derivatives didn't work.
We can't ignore software either. Until recently(probably little before Google collaboration announcement), x86 Android versions they worked on lagged behind cutting edge ARM Android versions. And software optimizations are just as important as hardware advantages. They seem to have that too now.