“Dwarkesh Patel suggests that based on an economic equilibrium model, the total volume of tokens generated by a frontier model like GPT-5 during its operational lifetime should be approximately equal to the volume of tokens it was trained on.”