"At one point, Jason says "one GPU core per instance." But actually we're told that the technology scales so efficiently and cost effectively that you can allow anywhere from a minimum of 10 users per GPU, up to 100 depending on the application"
10 users per CPU? 100? This sounds like wishful thinking and vast hyperbole to me. GPU's don't really like so much context switching, but more than that, a modern game tends to saturate system performance, not just RAM, but VRAM, GPU cycles, and CPU. The idea that you're going to virtualize a couple of instances of Crysis or Call of Duty onto a single GPU seems preposterous.
http://techcrunch.com/2009/06/16/videos-otoy-in-action-you-h...
OnLive had similar problems overhyping the scalability and economics of their service.