Because training a model costs millions, so each time you experiment with trying to create a new kind of model it costs millions.

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