Now you could argue whatever shared data store they're accessing needs to run things concurrently in order to maximize performance (and certainly many do), but even then we're usually looking at relatively simple shared memory patterns often with fairly coarse locking that scales well to a few cores, but not much beyond that (which often doesn't matter since you can saturate your network and/or memory bus before that). Maybe once in a while you'll see a lock-less concurrent data structure, but they're still fairly exotic.
What? Care to elaborate? (AFAIK, the only popular single-threaded webapp server is node.js, which is one of the slower web servers around: http://www.techempower.com/benchmarks/#section=data-r4)
...that scales well to a few cores, but not much beyond that (which often doesn't matter since you can saturate your network and/or memory bus before that)
You're describing limitations to current implementations that seriously hinder scaling. Developers are far from happy with this state of affairs, which, thankfully is not a necessity. You most certainly will not saturate the network if the data is distributed wisely. Here's a post I once wrote for highscalability.com that discusses the networking issue: http://highscalability.com/blog/2012/8/20/the-performance-of...
The memory bus is a bigger problem in general: if you do enough work on enough cores you will saturate the bus. But this, too, is simply the current state of affairs and not some insurmountable barrier to scaling.
Is there any evidence that this is true for most modern software? I can think of far more instances of software that use threads for simple things like networking or UI because it is easier and/or their library makes them than software that is so CPU-bound for single algorithms that they need to parallelize to maximize performance across CPUs.