It's an obvious cost optimization. Make the consumer directly cover the cost of inference and idle inference hardware.