The improvements in transformer implementation (e.g. "Flash Attention") have saved gobs of money on training and inference, I am guessing most likely more than the salary of those researchers.