I work in the "AI" space, and I've struggled to understand how companies will recoup the substantial costs associated with training and running current large language models.