Research

When should AI inference stay on the device?

A practical framework for deciding whether a product should run AI inference locally or in the cloud.

On-device inference is not automatically better than cloud inference. This research note looks at the product constraints that make one architecture more sensible than the other.

The decision is multi-dimensional

  • Latency and offline requirements
  • Privacy expectations
  • Model size and device capability
  • Update frequency
  • Server and inference cost