The physical cost of the GPU or IP design in terms of materials and fabrication is going to be low - on that measure the profit margin is huge.
The real cost of a GPU is the development: skilled staff, Research and development, RTL hardware guys, testing, drivers, SDKs/APIs, marketing, paying license fees, patent disputes, etc. On that basis profit's a lot lower. Easiest way to get it is to read the financial report a of a major graphics company.
Many major OEMs buy in graphics chips/IP for a few reasons. As 3dcgi said, it's cheaper and they can cost reduce on large production runs. Custom silicon can give them features that make their product individual (and just as often allows them to spec an existing chip without the features they don't require). Everything comes down to cost and turnaround time I guess but there are two other big reasons...
1) if you develop your own chip/IP there's a large risk someone else will release something better - if you stay in house there's going to be a your of pressure to use the design even if it turns out to be uncompetitive. If you're sure you're ahead of the curve then that's a good thing and it stops anyone else using things but it's probably easier to shop around and pick whatever looks best when you need to buy.
2) they likely can't create competitive graphics chips quickly enough. There's a very finite amount of engineers, mathematicians, low level computer programmers with the skills to make them. The existing companies have a hard enough time fighting over the experienced hires. If a company wants to enter the field they either need to pick up all the suitable graduates/PhDs they can find, poach a few senior people to lead things and wait 5-10 years or buy an existing company or two to get the staff/designs/patents...
The cost of the graphics chip in your PC/phone/console/whatever is really the cost of not doing the above split over the number of chips produced.
