Yeah, Amazon Web Services also offers GPU server instances. Really, lots do.
But like I said, a modern Xeon with its integrated gpu is perfectly adequette for what's needed here. A discrete GPU shouldn't be needed under the assumption Infinity's proc gen will work to generate a single point, instead of needing to do patches. I did not assume that much.
Like the collision detection you end up doing on the gpu for 20 people would likely be done in the same amount of time for 500 or even thousands of people on the integrated GPU, including the overhead you have to call to the GPU and reading the results.
I do not see why an actual GPU would be needed in the way Flavien describes it. If he's able to make the algorithm where it can generate individual points instead of large patches, there wouldn't be those memory issues or the need for a powerful GPU generating large patches.
Not only that, but the integrated GPU can do both the generation of the height points, and the checking of collision, all in the same program very, very quickly.
If it is the case that a single point can be looked up, not necessitating big patches that use a lot of memory, I think something like a E3-1285 v4 would be all the graphical power needed, and would be better than a discrete GPU since there is shared memory to use. It has 48 execution units at 1150mhz each, around 850GLOPS total, and 124mb of eDRAM shared between the CPU and GPU, if needed.