Renting GPUs and moving to open weight models took a $1.2M monthly AI bill down to roughly $ 100K.

That is the core of Ben Cera's new episode. Four months of frontier model spend while Polsia went from 500 users to 5，000 paying customers， then about two months of configuration work on rented GPUs， and June came in at around $ 100K. He also takes you inside the data center where the racks sit.

The detail I found most useful： the math only worked because most of his users were asking simple things of simple codebases. That workload never needed the most expensive model available. Episode 3 is worth the watch if you are building anything agentic！