CUDA api?

You are assuming that this would actually allow hosting of larger games. I dispute that and I seriously doubt you have anything other than conjecture to show otherwise.

I can go get a machine with 64-cores right now. Can't do the same with GPU's nearly as easily.

In addition the bus bandwidth necessary to talk to those GPU's doesn't scale.

I'm familiar with software engineering but you still have to actually code up both paths. You are doubling the work. In practice the CUDA side has more restrictions that will cause the design to be limited so working around those will likely cost more than double.

This isn't just a worry, this is a massive problem with the whole idea. If you can't run the entire simulation on the GPU (which is basically impossible) then you can only shuffle parts of it back and forth. This makes the programming model crap and limits scalability. There is a reason you aren't seeing games do stuff like this except for fluff.

My meta question was kind of rhetorical. If using CUDA for this game is a good idea I'm not smart enough to figure out how.

Oh, by the way, did you know that the guys from intel did exactly that? Check it out.
Shortly put: valve helped intel to improve their opensource video driver, while intel helped valve to tune up their source engine. And as a result L4D2 started to run twice as fast on intel graphics on linux.

We've been close to Intel since the beginning of Uber so we are working with them a bit (although we haven't send them a build yet). We are currently testing the game a lot on their integrated graphics though as it's basically the minspec for GPU.

I would be happy for nvidia or ati to get involved. I'm fairly certain we'll be sending them game to the ATI lab for compat testing.

I think you misunderstood what I meant, I meant that it would be better if Uber spent their time improving the game without using CUDA which will benefit everyone, rather than using CUDA and having to spend that time maintaining the two code paths instead which would result in a less polished game and only have marginal benefits that can only be used by a handful of users. I totally agree with your opinions on the use of CUDA for PA, it's just not practical and would be unlikely to yield much in the way of performance improvement.