The problem is how do you enforce coordination? This isn't like nuclear testing that can be detected around the world on seismographs or from satellites in space.
AI research happens in benign looking buildings that could be doing a hundred different non-AI workloads, much of it things that industry or governments would not want 3rd parties prying into. There is too much to potentially gain to trust coordination.
There is a lot of research out there with answers to your questions. GPU supply chains are extremely constrained, making it a surprisingly tractable problem at least for now.
How do you tell the difference between inference and training workloads from outside? The same GPUs can be used for both and the pacing suggested here relates to training and not inference as far as I can tell.