If everything is cached in memory then I/O is out of the picture... unless you count context-switching/paging between cache levels as I/O (and in that case anything that intensively uses the CPU is also an I/O operation).
In any case, when you're as large as Facebook, a 50% reduction in CPU-cost can have a significant impact as compared to a 50% reduction in CPU-cost on a site with 1000 hits/month. If even the minimal amount of CPU-overhead (when compared to DB-overhead or disk-overhead) can be reduced you can cut the number of machines you need to distribute that load, reducing costs.
In any case, when you're as large as Facebook, a 50% reduction in CPU-cost can have a significant impact as compared to a 50% reduction in CPU-cost on a site with 1000 hits/month. If even the minimal amount of CPU-overhead (when compared to DB-overhead or disk-overhead) can be reduced you can cut the number of machines you need to distribute that load, reducing costs.