Google is investing in coming up with smaller models aswell i.e "TensorFlow Lite". I think it is unjustified of her to condemn approaches at this stage, all research starts humble and almost always inefficient. People were doing Resnet-50 not so long ago and then processes improved. People came up with better configs with lesser layers YOLO for instance. So big models in and off itself isnt a problem. I dont see this as a good enough reason to resign.