First, I doubt your kernels statistics actually account correctly at that level of precision. (If it does, that alone would waste a lot of CPU cycles!)

Second, have you checked the memory footprint of NTPD ? "machine resources" is more than CPU cycles.

And don't forget: That number was one quarters purchase of servers, running for half a year. All the servers bought in 2010, 2011, 2012, 2013 and the other half of the year were not included.

Things add up.