I've been looking at ESXi recently, and I think I'm starting to figure out how it actually does remote rendering for vSGA - and if I'm right, it a bit lame. It runs Xorg locally, and seems to do something similar to what VMware Player might do running a Windows VM on Linux - use OpenGL Xorg acceleration to implement Direct3D functions. It fires up a VM outputting to the local virtual headless Xorg server accelerated by the local GPU. It captures this output, encodes it into a video stream (also GPU accelerated), and sends it down the TCP connection to the client.
I can't make my mind up if that is really clever or really crude and unsophisticated. Of course, I could be entirely wrong, but given the components it uses, that is how it appears to hang together.