Of course they're useful. You don't need to retrain LLMs so that they have internalized knowledge for them to be useful. The point of the LLMs is that they're able to go outside their training data and look up API descriptions etc and then perform work.
Even the existing open weight models we have today are enough for people to get good usage out of purely local AI on regular graphics cards for many years to come.
New foundation models can then be trained by companies who sell local datacenter deployment subscriptions. When you don't need to satisfy VCs and a quarterly market you can take much longer (=cheaper) to train them.