Fabs are very expensive. You can see numbers ranging from 4 to 10 billion dollars each. Those are the fixed costs that you now need to spread over production. It looks like Nvidia does around 20 million chips a year. If a $4bn fab was useful for 4 years, then those fixed costs work out at $50 per chip!

If you make other people's chips on the same fab then the per chip costs go down, assuming there is spare capacity. The economies of scale really do help. Which is why there are a few foundries who do most of the chips, and those foundries can spread the costs over a large number of chips including later ones that don't need the newest process. For example Intel is known to do their CPUs on the current process, while doing the chipsets on the older processes.

In other words, work out what set of numbers would make sense, and you won't be able to find any. New processes keep getting more expensive, and competitors adopt them too, so you have to sustain your current fab as well as the work to do the next generation one. This requires a lot of available money.

Don't build a fab unless you are sure you can use almost all its capacity for the next several years.

If you've got a successful (i.e. working well) and fully utilized leading edge fab, it can be more profitable than printing bank notes.

However it's so insanely expensive to set up that type of chip manufacture the only way it is really viable now is to share the costs. Samsung, TSMC and Intel (and maybe GF) are the only groups left who can afford it (see http://www.eetimes.com/document.asp?doc_id=1327060 ) and you have to keep up the r+d every year to stay competitive...

Actually, AMD/ATI spun off their fab into a separate business. Makes sense, really. Lets them more easily fab their chips where it makes sense, while still giving them fallback for safety.

Another benefit to not running your own fabs is that you are a bit removed from the rat race to keep your tech up-to-date.

Getting a new, smaller process to run well is expensive and difficult, but gives important benefits. If you don't run your own fabs, you can't jump ahead of others.

But you are also safe from investing a billion or a few in an upgrade and it not working out properly, and you still can get basically the same tech as all the other fabless companies.

They have a very close relationship with their preferred foundries and design chips with a specific foundry and process in mind. The big benefit of going fabless is that you don't have to worry so much about keeping your fab busy enough to cover its expenses. Intel's basically the only company that always sells enough of its own chips to keep its fabs busy and profitable.