Another axis would be power consumption. Pretty sure if you want LLM training in 1980 you'd also need fusion power in 1980. And it would be very, very hard to hide the cooling apparatus of your datacenter.

All that aside, it would also likely mean that the secret chip fabs eclipse the volume of the non-secret chip fabs by orders of magnitude... covertly? Or are we imagining chips that are 30yr ahead of their time? Either way it's just absurdities upon absurdities..