Not a direct answer to your question, but I interviewed a while back with this company called Fathom Radiant that's developing an optical computing system for AI training. It's further off than the next few years, but their ultimate goal is to use that technology to establish a computing center that outcompetes other options for training very large models, then leverage that position to support AI safety by allocating resources to groups that meet certain safety standards. It's an interesting approach and seems like a cool technology.