It’s the cost of compute hardware required to train a model of that size