Interestingly enough, I suspect Facebook may be pretty influential in this shift. In April of this year, they publicly released their physical server design to the world under something called the Open Compute Project. You can read all of the ins and outs of the project here, but the crux of what they did was boil a standard-issue Intel-based server to only the most necessary components it needed to be one of their cluster nodes. And nothing was was considered sacred - everything from how the power supply was constructed down to ways the case could be modified to maximize airflow were considered. Cosmetics were irrelevant also, since the goal was to build a better server. The results of their exercise were no short of remarkable - efficiency of their datacenter power and cooling consumption went up an estimated 38 percent, while buildout costs went down by 24 percent.
Based on this success and the momentum Facebook is certainly building with the customized server model, I am excited to hopefully see a return to the "good ol' days" of hardware hacking becoming a mainstream part of building out infrastructure. I know this will date me, but I remember the days when processors, logic boards with specific bus specs, and other components were obsessed over and chosen like wine when a meal is ordered. Once things were assembled (I still proudly carry a scar from an incident with a soldering iron), jumpers were manipulated on the board to turn functions on or off with the hopes a LITTLE more performance could be gained to make DOS 3.0 boot just a tiny bit faster. See? I told you I was old...
It was the closest thing I could equate to what wrench heads did with cars to make them go faster as hot rods. Which was the best I could hope for, since I often forget what side of the car the gas pump goes in...I am THAT bad with cars.
Let's hope we get to go retro in the future and have some adventuresome options with how we build out the next set of servers we buy...vroom vroom!