If technology advances as rapidly as we'd expect from the last decades; especially if RAM advances to cheap 256GB, 512GB levels, then I think we can expect local AI to be a serious thing. Possibly even power training your own AIs at that level (500m, 1B, 4B, 8B).
There is a logical limit as to what these things can do and what datasets are available. Facebook NLLB translating AI machine is 600m for example, very lightweight. But it probably trained on massive datasets. It's important that all of this is readily available for the local maker/hacker to train for themselves.
I think it's reasonable to expect that European and other World Universities will make this information available, as it is of common interest.
The thing is: this limiting development of machines that impedes datacenter-centered AI from working (quick obsolescence) is what could make it work locally.