The 10240 core system is a 20 node cluster with 512 cores on each node (Columbia cluster at NASA).
These big SGI systems prove little about the general scalability of Linux as they are all running HPC (high performance computing) workloads - read in some data, crunch floating point numbers for several hours then printf("42\n") at the end of the run. I.e. not stuff that will stress the scalability of the operating system.
Take one of those systems and have it run a workload where all the cpus are trying to create/write/read/close on a bunch of small files, and you'll soon see how poorly they scale for general purpose workloads.