Time will tell, but right now we're not solving the problem of running LLMs by increasing efficiency, we're solving it by massive, unprecedented investments in compute power and just power. Companies definitely weren't building nuclear power stations to power their spell checkers or even 3D renderers. LLMs are unprecedented in this way.

True, but the usefulness of local models is actually getting better. I hope that the current unprecedented madness is a factor of the potential of cloud models, and not a dismissal of the possibility of local models. It's the biggest swing we've seen (with the possible exception of cloud computing vs local virtualisation) but that may be due to recognition of the previous market behaviour, and a desperate need to not miss out on the current boom.