At work we have a number of NVIDIA QuadroPlex boxes that are cabled to a card that goes into the host machine.  The cables on these boxes are something approximating 2m in length.  The host card contains some kind of PCIe bridge chip and somewhere, either on the host card or in the external box backplane, there's a PCIe switch to allow the two GPUs in the box to connect to the single PCIe slot in the host machine.  I have no idea how special the hardware is there, but clearly the 2m long cable doesn't cause problems in that scenario.  They use VHDCI cables for the link from the host card to the external box.  I know that the USB-based PCIe riser cables are nothing compared to what NVIDIA did for the QuadroPlex, but the QuadroPlex might be considered the top-end version of what is feasible.  Besides the QuadroPlex boxes, there are various vendors that sell other variants of high-end external PCIe backplanes with on-board PCIe switches and the like.  The only problem is that all of these external boxes tend to cost $5k to $10k for an empty box of PCIe slots and power supply.

It would be interesting to see how well these low-cost PCIe risers would work if they used a higher quality cable more along the lines of what NVIDIA does for the QuadroPlex.

Since Christian brought up the HF frequencies:  One thing that's not so good about building the "milk crate" or "open air" machine that lots of people do is that with no computer case, there's no RF shielding, and you're likely going to emit all manner of RFI, polluting the radio spectrum with all kinds of terrible hash.  As an amateur radio hobbyist, I would beg that people take that into consideration and consider ways to do better than leaving the boards just laying on a shelf.  If you've got a pile of motherboards sitting on a shelf, some conductive mesh (e.g. for window screen) fashioned into a faraday cage around the systems would help null that out...