major percentage of these losses can be attributed to the outrageous expenses of training language models. Those cost seems to be developmental and not operational, after they reach AGI I assume it will go down