A practical framework for deciding which parts of an AI inference pipeline to run on local hardware versus serverless cloud. Using a speech-to-English translation tool as a working example, the post introduces a four-axis decision model (privacy, cost shape, maintenance burden, capability access) to partition workloads. ASR runs locally on-device to keep raw audio off the network, while translation is offloaded to DigitalOcean's serverless inference endpoint via a standard OpenAI-compatible API call. Measured results show 99.98% payload reduction (2.65 MB audio processed locally, only 883 bytes sent over the wire) and a translation cost of ~$0.0006 per utterance. The key insight is that hybrid boundaries are often discovered through constraints rather than designed upfront, and the pattern generalizes to any pipeline with a sensitive/high-frequency stage alongside a heavy/occasional one.
