Practical advice for hardware startups and product teams deciding where AI processing should happen, how to split the workload, and what that choice means for cost, speed, and user trust.