Since AI training is highly parallelizable one can look at it as a question of cost for training, that is given enough electric power, any nation state can trai

Since AI training is highly parallelizable one can look at it as a question of cost for training, that is given enough electric power, any nation state can train the model they want. And it would be nice to know how much cost reduction is due to fab process advance and how much is due to architecture and software advances. My suspicion is that the fab node is not strategic. It is a commercial issue to integrate more cores as fab advances, but training can be done over more chips and boards if integration isn't there, maybe for higher cost.