Theoretically, yes a video card could implement USB4 (by having a USB 3.2 host controller, and a USB 2.0 host controller on board, implementing the routing fabr

Theoretically, yes a video card could implement USB4 (by having a USB 3.2 host controller, and a USB 2.0 host controller on board, implementing the routing fabric etc. But in so doing, you have limited the bandwidth available to the GPU (although they actually do still make framebuffer GPUs that don't have any calculation acceleration support). Also don't forget the need to support USB Power delivery (there is no requirement on being able to deliver power, but PD communication is required, and many USB4 devices will expect to be able to negotiate for some level of power delivery), so this card will probably need those extra power cables like normal accelerating GPUs use. But if you had such a card, and your motherboard has one two, then you have one machine that is actually two seperate USB4 hosts. This opens up a bunch of questions like does the host OS actually support this sort of configuration (especially if the Connection Manager is being implemented in software rather than hardware). Then you get weird stuff like being able to connect a Type C cable between the two hosts of the same system to get Host-To-Host Tunneling, but where both hosts are the same system. That is some wacky stuff. Even Host to Host tunneling is esoteric, basically allowing system that implements each host to pass each other arbitrary data over over the link. The example in the spec is exchanging IP packets, but I;m not sure that details of this usage are really standardized. It might also be possible to support add-on cards for USB4 by a routing chip manufacturer that makes use of the device to device communication features of of PCIe, to allow multiple USB4 routers to work together as a single virtual router. That sounds far more plausible, although doing that sort of thing would increase the contention on the PCIe bus.