What I'm thinking of isn't so much 'optimization' as basic design decisions. For a lot of IoT products that integrate machine learning/computer vision features,

What I'm thinking of isn't so much 'optimization' as basic design decisions. For a lot of IoT products that integrate machine learning/computer vision features, you've got a choice of implementing that stuff in the cloud or on device. You need a bit more compute to do the work on device, but ongoing costs are far lower and a lot more predictable. As a side effect, you also enable fully independent operation.