As I type this, I'm waiting for a process that could have a parallelism of about 9,000 threads but it's running on a mere 12-core machine (compiling a large C++ project). Bring on the "20 or 100 CPUs" for my next computer please. Though, as mentioned elsewhere, this is also likely to hit memory bandwidth limits for affordable systems, not to mention total memory usage limits given that more than a few of those 9,000 compiles peak at >1GB RSS.