PCIe is obviously a given since expandability is a feature that separates the Pro from Apple's other desktops. I believe it will continue support stock AMD cards purely for choice and software optimisation, but that Apple's marketing focus will be on its own own GPU.
How is "software optimization" when the AMD GPU drivers do not exist on macOS on Apple Silicon? You need to have something working to optimize it. "the drivers exists on the macOS intel kernel so therefore they exist on the Arm branch also" isn't necessarily true. macOS on Apple Silicon sprouted from iPhones and iPads which never had more than one GPU vendor. That really hasn't changed in two years ( WWDC 2020 - 2022 not a peep about adding depth to macOS on Apple Silicon on the GPU coverage front or substantive changes to DriverKit in the GPU stack space. )
It isn't just Apple's "Marketing focus" . It has been their technical focus also. There is nothing in DriverKit about Graphics stack specific extensions. There is no technical movement at all on 3rd party stuff inside the kernel going forward (kernel extensions have been deprecated and will disappear in some future version of macOS. Probably not soon but it is coming. Not some marketing slogan; a technology announcement. ).
There is deep fallacy in the notion that "PCIe == GPU Display output cards" . That is closer to being a "marketing focus" than a technological one. GPU Display output cards are a narrow subset of "PCIe slot cards".
When has Apple put a huge marketing focus on completely stock , general market GPU cards? Apple didn't do that in the Mac Pro 2006-2012 days. There was a subset of cards commissioned by Apple for mac boot ROMs and for sale at the Apple Store. But Apple was never trying to be Newegg/Microsoft and sell every generic GPU card possible? That never was their marketing strategy at all.
Macs on M-series have walked away from UEFI boot. So all the native boot support in generic GPU cards is gone. There was a quirky "happens to work" period that the Mac Pro 2019 walked into where happened to coincide with some incrementally better support of generic market cards working. Most GPUs that "happen to work" coverage came from other Macs using dGPUs as embedded GPUs ( iMac , MBP 15" , iMac Pro got drivers so happens to work in eGPU/Mac Pro also). By 2019 Nvidia support was dying ( so large chunk of generic off-the-shelf card market gone). The Pro Vega and W6x00x were as much Pro card coverage as generic desktop retail add-in card coverage driver development.
Apple doing a AMD MI210 wouldn't be too surprising. Apple chasing the generic Windows gamer card market would be. An AMD for 'extra' computational add-in TFLOPs. (if a bit easier perhaps with one or two Pro GPU packages that overlap pro/mainstream ) But for GUI display whatever Apple ships with the system is a GPU technical focus. And that will likely be Apple GPU. Apple's technical focus will highly likely be on getting more developers to write highly optimized Apple GPU code first and foremost.
A non-boot , non-GUI 'compute' GPU card would more easily get around the lack of UEFI and 'native' iPhone/iPad apps running on macOS (that expect only Apple GPU environments. )
Apple's run rate on Macs is around 20M per year. About to enter year 3 of M-series deployment. Meaning not all that long from now will have 60M M-series systems out there. That is a sizable user base not to be optimizing code for. Throw the M-series iPads on top and it is even larger. Those the systems "paying the freight" for the CPU/GPU/NPU core R&D; not the Mac Pro.
This begs the question: will the GPU be integrated into the SoC as a true scaleable architecture from Ultra, or will the machine have its own bespoke SoC? Gurman and Amethyst are leaning towards the former, yet I'm still not sure if this would be the most practical way to go, especially with regards to the number of efficiency cores and the size of the package with GPU cores added.
A major component of "practical way to go" is pragmatically going to be "economical way to go". Modularlity without economics of how much the bespoke/custom costs is not going to lead to practicality on Apple's balance sheet. Mac Pro is probably in the sub 100K/yr run rate product.
The Mac Pro doesn't sell in high enough volume to have 100% completely custom parts. If it did then Apple could just sell Threadripper and Radeon parts.
There probably is some customization ( subset of components added that won't appear in Mac laptop SoC. ), but "start over from scratch" isn't likely going to come to the Mac Pro. Mac Pro will likely get more of the cores that are target developed from the laptop & mobile in for CPU/GPU/NPU . Economics of the bespoke/custom RAM packages that Apple is using ... same thing economies of scale to pay for R&D are likely outside the bounds of what Mac Pro can pay for solely on its own.
Apple is likely selling 5-10x as many Mac Studios ( mid-upper end iMac 27") as they are Mac Pros. Cobble the Mac Pro + Mac Studio and perhaps a 'large screen' iMac Pro together and probably can stop slavishly using the MBP 16" Max dies as a building block. (e.g., at about x9 as many at about a 1M/year run rate). But Mac Pro still would be tightly coupled to embedded SoC systems for critical volume ( just not immediate directly to laptops). The monolithic Max is a rather chunky 'chiplet'. Scaling past two probably creates undesirable issues. Apple probably needs a better factored functionality chiplets for desktop. However, it would still be chiplets on a package as the core's foundational design was based on that base SoC package approach.
Probably could get a die building blocks that are more optimized to building "duos" and "quads" ( Ultra and Extreme class) that could happen to work for a desktop "Solo" for a 'Desktop Max' for the Studio. The laptop (and lower 'half' of desktop ) would just be purely monolithic chips ( at least until Apple weaves in cellular radios chiplet).