I don't think this changes anything. Deployed networks typically use only inference from pretrained weights, and those weights are what get transferred for mode

I don't think this changes anything. Deployed networks typically use only inference from pretrained weights, and those weights are what get transferred for model "updates". You can have all your devices using weight array W0 for a neural net architecture, spend a million compute hours training that net on cutting edge systems to produce a much better weight array W1, then upgrade all the deployed devices by sending them W1 which will be the same size as W0.