Haven't Apple (A-series, M-series) and others been copackaging memory for a while? (SiP or MCM?)
Then there is the HBM which is an expensive process to mount on the chip.
Putting RAM inside the logic die is crazy expensive - just look at how much footprint L3 cache takes up on a modern chip. It'll always make more sense to put it in a separate die, made using RAM-specific processes. With Zen the memory controller and GPU has (once again) moved to a separate die, and with X3D we're now even seeing separate cache dies.
There's of course HBM if you want to consider that "integrated", but the large cost bump, additional manufacturing risks, and loss of flexibility don't make that an attractive proposition unless you really need the extra speed. Regular DDR dies? You gain basically nothing by co-packaging those.
Optimal: piecemeal. As in: X sockets on a board, each socket takes a chip that adds CPU core(s), GPU cores, and some amount of main RAM. More chips, more powerful computer.
Keyword: socket. Read: replaceable. Bad chip out, good chip in. Slow chip out, faster chip in. Soldered onto board = reliable, but a defect could junk the whole machine.