There were already comparisons made on the point you mentioned in as early as 2011/2012 on techinferno (the DIY eGPU hub on the internet back then) and that is true. Have also seen the AW amp, sort of a rerouted PCIe than through a thunderbolt controller.
Well if you ran
eGPUs for laptops before back in the heydays (around sandybridge) before Intel started "punishing" small vendors who aren't paying royalties for the tech, the compression tech baked into the GPU driver and the card's data bandwidth reqs are key puzzle pieces too that can be worked on outside of specs minted into the interface. The successful implementations all worked on what the
"user can control" than something they cannot.
Overall still, compression tech and physical interface limits haven't reached the crossroads to put large data throughput into something as small as 4 PCIe lanes. Unless there's something that happens to the next revision of thunderbolt where you can run multiple of these controllers in tandem and weave multiple small roads into an expressway (and whenever needed a eGPU riser/dock supports connections to multiple TBs simultaneously), I don't think there's a shortcut getting there. Though working around thunderbolt usage is an option

like what Dell did.
But how many vendors will make their own PCIe based solution than taking something ready made and available?