It's fairly simple. Sell GPU compute + extra margins as only some GPUs can load the models + extra margins based on how much better closed source models are from open source ones + hopefully reduced cost due to batching