Actually that's not true. When PCIe 3.0 was introduced, it took quite some time before the (still marginal) performance improvements materialized. We're talking about a few percent at most in games and benchmarks (many are just neutral) between PCIe 2.0 and 3.0, simply because it's not a bottleneck. For not overly data-intensive GPGPU applications (e.g. video things) I reckon that's the case as well. Yes, sure, there are applications you can run on a GPU where PCIe is a bottleneck. PCIe bandwidth benchmarks, for example.
Don't expect PCIe 4.0 consumer-ish (i.e. what this is about) GPUs soon (2019 earliest). It's wholly unclear whether you'll see any performance improvements at all from that — just look at the bus utilization in games and benchmarks, you'll be hard pressed to see more than 10-20 % (which is precisely why running a GPU on a x8 interface (or 2.0) is not an issue).
I was saying that the Blackmagic eGPU in an upgradeable form factor would mean that you get a 200-300 $ (or something like that) GPU enclosure with some electronics in it (Thunderbolt hub, power supply etc.). Then you stick a current GPU into it. This would make sense because GPU replacement cycles are much shorter than the interconnect upgrade cycle. That's why I said that most applications are still fine today with the over ten year old PCIe 2.0.
So just to reiterate: PCIe bus speed is simply not a limiting factor for the overwhelming majority of GPU applications.
And you’re right—PCIe 3.0 cards out today have plenty of performance for today’s applications. That why you would get a PCIe 3.0 card today. But it doesn’t matter if it’s fixed because you aren’t going to upgrade it.