Depending on the company, these costs can vary quite dramatically ...

In the past some CPUs, GPUs and other components have been designed (more or less) from the ground up for a company and the development of these components can often cost hundreds of millions or billions of dollars without including technology licensing fees.

Nintendo has been known to outright buy the design of a particular component to lower per-unit manufacturing costs and to give them the option to take it to another manufacturer to produce the component; with the XBox, Microsoft didn't do this and when nVidia refused to continue developing the GPU Microsoft had to stop production of the XBox.

With something like controllers you might assume the R&D costs would be low, but for each controller there are (probably) multiple tech-demos created to prove its value; and there are probably a lot of "failed" controllers produced for each successful one. I could be wrong but I wouldn't be surprised if Nintendo had multiple HCI teams working constantly on projects, with each of these teams having a budget of (something like) $5 Million per year, to be able to produce a system and its accessories; often with a lot of "failed" concepts because they're ahead of their time, too expensive, or not accessable enough. Over a 6 year period these costs add up dramatically.

I think a lot of it comes down to prototypes, the design, and the best cost efficient strategies to use. Should it have certain plugins, why or why not etc. Take Xbox 360 for example. It didn't have good ventilating so early in it's life it had some issues. Now it is fixed but that type of issue should be found in research and development.

I forgot to mention the "hidden" side of things ...

Operating systems, drivers, libraries, licensed middleware that is included in the SDK, network services, and whatnot all cost a lot of money but few people consider those costs.

Soleron pretty much described it. But as something that's in the market about 5 years after the development starts, one of the earliest things to do is project the tech available at the time the production starts - and the cost of it.

Sometimes projections can be off by a wide margin; for example Project Reality was supposed to have it's CPU clocked at 500MHz, but when it came out, CPU was running only at 93,75MHz. Or PS3, it's launch was eventually delayed a year because Cell and BD-drive weren't ready (and the product ended up having one SPE in it's CPU functioning as a spare part).