It’s really very simple:
He wants government regulation that anoints his company as part of an oligopoly.
The worst outcome for him is to have spent hundreds of billions training frontier models only for them to plateau and be replaced for most tasks by open-source models running on local devices. He has publicly talked about how risky this rate of capex spend is, and if they are wrong only slightly in their usage projections, they will go bankrupt.
The way to protect against this is to fear monger enough that the government steps in and says only company x y z get to run LLMs above a certain size for safety reasons.
This ensures his company won’t be undercut by some open-source Chinese model that can run on a Mac mini and satisfy 95% of consumer demand for AI.
This is very similar to Meta-led fear mongering about TikTok. They wanted their competition blocked from the marketplace.