Because PC’s don’t work that way. The access time for a PCI card is limited by the PCIe standard built into the motherboard and the card’s hardware. It’s not something the GPU gets to control.
Things like this are why modern, high-performance SSDs use M.2 sockets instead of SATA: they can access data faster than SATA can send it. So they use a different connection to avoid having to be limited by SATA access speeds.
So to make this work, you’d have to come up with a new motherboard connector so that other GPUs can more efficiently access its data. And that connection will have to be incredibly high-bandwidth (at least on the order of memory access times, if not faster) to make this work. So not only would you have to develop a special card socket, you’d have to do substantial motherboard work; it wouldn’t work for existing motherboards.
Oh, and the target GPU that’s going to read this data also has to have a higher performance interface to it. So it wouldn’t even work with existing GPUs; you’d have to buy a new one.
So, you could by a raytracing accelerator card + a new motherboard that has this special socket + a new GPU that can work with it. Or… just by a GPU that can do raytracing. Because those already exist.
The particular confluence of effects that made 3D accelerator cards viable in the mid-90s simply no longer exists today.