I doubt the bad performance is due to evaluating expressions itself. If I type N[Log[2]] into Mathematica, it evaluates in less than a millisecond. It's probably because Wolfram Alpha is using natural language process to try to process my query and then finally deciding that by N[Log[2]], I mean N[Log[2]]. And it's probably not because of that, but because their grid scheduler isn't optimized for sub-second latency.