For many reasons, including:

- We already have the expertise to do so, so the cost to us is much smaller

- We can ensure a much higher level of service by making sure every component along the line is up to par (you'd cry if you heard what some colo providers do)

- We can fix problems in-line with our priorities, since we're the ones that fix them (like keeping published sites up, even if the editor is down, for example)

- We can set up and manage geographically redundant datacenters

- And finally, we're able to do so extremely cheaply

It's definitely something to consider when you reach a certain scale, and especially if you have the know-how to get it done. We've calculated that for every server we purchase, we'd pay that cost every 2-3 months on AWS, for perspective.