I don't think it'll use MPX. That was a weird proprietary workaround Apple invented because Intel Xeon E5 CPUs did not support anything faster than PCIe 3.0; it added extra lanes / power connectors and an internal Thunderbolt bus so that you could pipe graphics out of the stock TB ports instead of a port on the GPU itself. Not sure why they needed that, but they did.
It's a technological dead end, and the 7,1 Mac Pro is such a small market that it's never going to be taken up by many third-party peripherals manufacturers, except for boutique pro solutions providers, probably.
Since Apple controls the SoC, there's no reason not to support full PCIe 5.0 (with 4.0 backwards compatibility) which will be 4x the speed of 3.0. As
@MisterAndrew says, it's up to Apple to support 3rd-party cards.
That said, it really depends on what Apple sees as the market for the Mac Pro. If it's a workstation aimed at high-performance niche markets like audio-video / music production, engineering, scientific computing, VFX, then all that matters is that the GPU has enough horsepower to do those tasks, not necessarily the brand.
Feature film VFX, arguably one of the big use cases, is rarely rendered on a single machine anyway, but sent to a render farm. But the individual VFX artist's workstation needs to be fast enough to render changes and preview animation, scenes, etc. So far, in tests we've seen, the M1 Max and Ultra are quite capable vs the Vega in the current 7,1.
What I can imagine in the 8,1 will be either a really monster SoC with more graphics cores, more high-performance CPU cores, and
much more unified system memory and cache, or they (less likely) might take the step of having a separate, socketed Apple Silicon GPU but that is tied closely to the CPU; that might be user-selectable / upgradeable.
Even less likely but still possible, might be Apple Silicon GPUs on standard PCIe 5.0 cards. These might be solely for offline rendering and not realtime previews, because being outside the SoC would mean roundtrip delays in sending and receiving data, which then have to be assembled, buffered, synced etc. But it would be a selling point for a rackmountable render-farm solution.
I can see the expansion slots being used for TB / USB-C expansion, additional Ethernet for RedNet/Dante or Avid integrated systems, 4k+ video capture cards, HDMI for previewing on various devices, PCIe SSDs, certain kinds of pro audio i/o (ADAT optical), etc. But maybe not for GPUs, indeed.
