If they manage to make memory more efficient, they’ll just increase the context size and/or model size.

We just haven’t reached the diminishing return of gen AI capabilities yet.

Models will get more useful if you have higher context size or higher param size. Then people will just use the models even more, leading to even more memory demand.