Just because video was not mentioned at WWDC does NOT mean video is permanently off the table.
Apple did more than merely stay quiet about external GPUs at WWDC. John Ternus, Apple's senior vice president of Hardware Engineering, said this in an interview with John Gruber:
Gruber: Are there technical barriers to having expandable graphic through PCI that would be only used for compute as opposed to video? Or is that just a design choice?
Ternus: I think, I mean, fundamentally, we've built our architecture around this shared memory model and that optimization. And so it's not entirely clear to me how you'd bring in another GPU and do so in a way that is optimized for our systems. It just hasn't been, it hasn't been a direction that we wanted to pursue.
Translation: "Never say never, but I don't see us doing this."
Multiple out of context 'words' ( 'video' , 'compute' , 'graphics' , 'GPU' ) all supposedly pointing at the same 'card' really doesn't help much.
The full context of Gruber's question is this:
"Obviously Industry Wide the theme of the year is AI and AI training , and that whole area it seems that all of the compute takes place on graphics card. And .... users looking to do their own AI training"
[ at 22:40 ]
Here Gruber is basically equating 'AI compute and 'graphics card' ( He is heavily entangling 'AI' to 'graphics' to the extent he is using them later as essentially as interchangeable words for the same topic. ) . GPU cards and AI training really not that necessarily entangled. Don't need a GPU card to do AI training ( or inference). And 'graphics' is dubious since vast majority of these "in the cloud" data center cards have any video out what so ever (So direct graphics output isn't the focus). Same issue for the contextual quoted comment where 'video' and 'graphics processor' are being used as completely interchangeable terms when that is not really true.
So when Gruber is asking about " are there technical barriers to expandable graphics through PCI that would only be used for compute , as opposed to video"... that is essentially in his context " expandable graphics through PCI that would be only used for AI compute .... "
First, Gruber is explicitly leaving video out of the loop all together. So Apple's response really isn't directly at 'display GPU' , but it will have very similar issues.
So we get the Apple responses of essentially "go buy your AI compute in the cloud" along with this "our optimizations are not aimed there" response. Which shouldn't be much of a surprise. Apple probably isn't going to be a yapping dog chasing after the large(est) language model problems.
Apple has a wide enough number of markets they are trying to penetrate with well funded competitors. ( AMD/Intel/Qualcomm in PCs space ) . Qualcomm in mobile and celluar modem space . Nvidia/AMD/Intel in consumer GPU space. Nvidia/AMD in non-AI-focused Pro card space. Qualcomm and others ( VR headset space). They don't need to chase another , 'red hot' area that is drawing gobs of both VC money and funding from entrenched players.
Apple's 'optimizations' probably has as much to do with their software as it does the hardware. Tossing CUDA and anything Khronos compute ( OpenCL , SYCL ) out the window basically completely detaches them for all of Nvidia , AMDs , and/or Intel efforts in compute GPU cards also. Apple is pretty much hostile to the software stack those cards are primarily aimed at. Apple has 'optimized' to Metal only. Which helps them find deeper synergies inside their own products behind their 'moat' , but at the same time
worse portably outside the 'moat'. That optimization isn't 'free' ( it is dual edged).
Apple ejecting three 8-pin AUX power connectors out the window from the Mac Pro a 'heavy compute' card is not an option they are heavily considering. Probably a mistake long term, but in the mist of doubling down on marshaling the transition it is understandable. ( it is a mistake to forcibly drumbeat the vast majority of folks trying to do something portable off your system. Especially, when it largely doesn't have a GUI interface anyway. )
As others have mentioned, if they want to add GPU power, they would most likely do this by instead having modular AS CPU and GPU SoC components, such that you could add extra AS GPU chips to customize a the Mac Pro for GPU-heavy needs.
Apple's comment doesn't surface anything modular in the terms of multiple, modular SoCs at all there in the response. People outside of Apple keep trying to 'invent' that, but the 'optimizations' he refers to and the pragmatic 'unified, uniform' == 'shared' memory approach they have taken doesn't really lead much to modularity of two SoCs present in the same system aligning with the optimizations that they did. Trying to couple multiple large GPU core clusters together does have very real NUMA issues to get around. Even more so if have video out assigned to the same set of GPU clusters and associated display controllers hanging on same internal bus and memory systems.
Could the whole SoC subassembly ( SoC plus hyper coupled other components) be on a swappable card. Maybe. I don't think that will get folks what they want though ( a generic , long term, longitudinal slot/socket. ). Or at easily affordable prices.
The optimizations that Apple has done would push more so to perhaps having multiple macOS instances run inside the same system. [ that could perhaps cluster instances without tons of external container overhead on the 2nd or 3rd system instance. but not a application transparent unified instance. ]
But this would be too expensive to do for the Mac Pro only, so it would need to be something they would want to do across the Mac line. I've no idea what the chance of this would be.
Once the a M-series has a decent PCI-e interface ( > x16 PCI-e v4 in bandwidth) making a "Mac on a Card" would not be either hyper expensive , nor blocked from interacting with other Macs that only have TBv4 ports on them. If it is a 'remote' computer that access over virtual Ethernet-over-PCI-e then it is just another Mac on the network. TBv4 would be al lot slower cluster interlink network than x16 PCI-e v4 , but faster than most 1-10GbE implementations. (relatively cheaper too since no additional wires or Ethernet switches needed).
Pretty hurdle there is trying to get Industrial design to sign off on not fully enclosing it in a container of their own design and possibly using an AUX power cable to power it. But it is completely possible to put a whole computer onto a card with Apple SoCs. (at least small to midized ones. Monster large ones probably not. ).