Use somebody else's engine. Game companies and their employees should be focused on building content: writing, graphics, sound effects, music, level design, etc. At the end of the day nobody who played Call of Duty said to themselves "That rendered much more smoothly than Medal of Honor".
What sells games is content. What's important is shooting a bunch of civilians in a Russian airport, not how many civilians the game engine could render--as long as it's comparable to the competition's civilian killing rendering engine.
I came to this realization while writing my own input pipeline for an Android proof of concept. After hours of tearing my hair out over a performance issue I finally found the relevant issue described on a blog. After a short exchange of pleasantries with the author I had my solution--but it was his solution. I was solving a problem that somebody else had already figured out, and all that effort wasn't going to sell one additional copy of our game.
