Pretty much what Apple did with its M-series processors. Apple eliminated the bus and used main RAM as VRAM so there is no need to move data from RAM to VRAM because it is already there. The big thing they eliminated was not the GPU card but the bus the card plugs into. Buses are slow and take up space and need a bus driver and receiver chips.
The big win is that when you have two kinds of memory, you are forced to move data between them. Making it just one pool of RAM fixes that. So lets see if Intel can do that. The huge problem with PCs was the stove piping, where every part was biuld intependantly and talked over some interface. moduarity is OK but the interfaces where choke points. For example why place the SSD on a card on a bnaus when you can build it right ito the main crossbar switch on the system chip? I think we are headed this way.
Currently, there is a huge shortage of Apple Minis because people are buying them for AI servers. Turns out they were too cost efecitve are sold out. The new intel/nvidia chip might help
For robots, even with this, we still have the problem of software. Hardware is easy but someone has to program it if you want a robot to replace your housekeeper.