Thunderbolt is PCI-E + DisplayPort, which is most likely how the video feed is being sent back.
My first thought was that the video signal was carried over the DP link as well, but apparently not.... my reading from this is that it's carried over PCI-E:
A decision you'll need to make is if you want to plug a monitor into your video card or just use your laptop's monitor. Each has it's own pros and cons. You'll get faster performance with an external monitor, but you'll lose the convenience of not needing a giant monitor. This becomes relevant as people make better eGPU cases where your eGPU will be portable. Why bring a monitor to your friend's place when your laptop already has one?
It's actually kind of cool that you even get this choice. The way it works is by the NVidia Optimus drivers taking the video frame memory from the video card, piping it back over the Thunderbolt bridge to the Intel HD 5000 memory and overwriting Intel's memory so that you see the eGPU's output on the Intel LCD. Cool! If you're curious, this is the exact tech that's used when laptops have an NVidia internal discreet graphics chip.
Which makes perfect sense, of course, and eliminates my confusion as to how the whole thing was being accomplished. To the MBA's CPU/chipset, the eGPU is just another PCI-E attached device. The eGPU can communicate bi-directionally over PCI-E to anything else on the PCI-E bus, exactly as if it were an internal GPU. That's the whole point, of course.
My main point of confusion was that I didn't think about the software and driver magic that allows machines to have switchable dual GPU's... I was thinking that the DisplayPort aspect of Thunderbolt had to be involved somehow, but apparently it isn't. The Intel integrated GPU is directly connected to the LCD, and the discrete GPU sends its info into the display buffer of the Intel GPU. That's..... really cool that it can do that.
To those talking about USB being able to do this.... I really don't think so. Think about all the communication that goes on between the CPU and GPU; you want that link to be as direct as possible. A USB layer is just an intermediary that would add lag. I'm sure it's possible, but it would probably give pretty poor performance. Remember, there's two aspects of a speedy connection: bandwidth and latency. USB 3.0 might have plenty of bandwidth, but I would bet that its latency is orders of magnitude slower than PCI-E. This doesn't make much of a difference in mass storage or network communications (where the lag is inherently greater) but CPU and GPU have been very tightly coupled all the way back to the AGP days.