If he's reimplemented the OpenGL 1 API, then he must be retransmitting geometry every frame, because the API is not sufficiently expressive to allow for retained state of that nature. He describes glBegin and glEnd as accumulating an array to batch out, but that batch gets re-accumulated and transmitted every single time the glBegin/End block is executed, i.e. every single frame.

Question from ignorance: Couldn't you internally buffer the geometries and, say, send them to the GPU every 5 frames or so?