There are three existing solutions I'm aware of:
1) The way PCs typically do it, by routing the video back to the motherboard physically by using a cable. Admittedly not very Apple-like. But Apple could have a custom PCIe card that routed the video back to the mobo over PCIe or an internal cable.
2) Plug the monitor into the computer's TB port and the PCIe graphics card is headless. Cheap onboard video handles bootup. After bootup, software allows the powerful GPU to do the rendering, but uses the integrated GPU for the physical output. This is already proven
for example if you want to use an eGPU on a laptop, but want to use the laptop's built-in display instead of a monitor connected to the eGPU. I don't know how this is accomplished--perhaps it's just mirroring the framebuffer from one GPU to the other.
3) The onboard card is MXM, not some custom format, so that it can easily changed with other MXM boards. This has been done for example
when an Xserve 3,1 was upgraded to a GTX980M. Users have the option of either choosing an MXM GPU upgrade, or PCIe GPU ugprade, (or both), with the caveat that when using PCIe you'd have to connect your monitor via DP instead of Thunderbolt.
In fact, if the nMP went to MXM you could at least upgrade the GPUs once in a while instead of it being a system all locked together in time. Granted, this isn't PCIe, but it's an improvement from where we are now.