The data center safety net disappears
When you deploy a model to the cloud, you're operating inside a set of assumptions so comfortable you stop noticing them. Need more memory? Attach a bigger instance. Need lower latency? Add another GPU or spin up a second region. Need more throughput? Autoscale. The model itself can stay large and unoptimized because the infrastructure around it is elastic and someone else is paying to keep it running. MLOps, as usually taught, is largely about managing that elasticity: rolling out versions safely, monitoring drift, scaling horizontally under load.
On a phone, a smartwatch, or a microcontroller embedded in a thermostat, none of that elasticity exists. The device has a fixed amount of RAM, a fixed amount of storage, and a processor that was chosen years ago for cost and battery life, not for running neural networks. There is no second instance to spin up. If the model doesn't fit, it doesn't run, full stop. This is why edge deployment is not just 'cloud deployment but smaller,' it is a different engineering discipline with its own hard constraints that show up before the model ever makes a prediction.