Intel's early integrated GPUs were practically free: their chipsets had a minimum size necessary to accommodate all the various IO PHYs around the perimeter, and they had enough free space in the interior to add a minimal GPU without much increase in total die size. After proving to be a lucrative way to kill off the competition's low-end GPUs, iGPUs started moving into the CPU package and later on-die to offer better performance.

Nowadays, we're running against barriers from the end of Dennard scaling, which prevent chips from operating at 100% utilization and all but forces us to start including more specialized units that will only be used for some tasks.