Edge deployment means the model sits on a device at the site of the work: a controller on a production line, a unit in a delivery vehicle, a box in a clinic. Inference happens there, on local hardware, and only summaries or exceptions travel back to a central system. The alternative is sending every observation to a data centre and waiting for an answer to come back.
Three constraints usually force the decision. Latency, when a control loop has milliseconds rather than seconds to respond. Connectivity, when the site has intermittent or expensive bandwidth and the process cannot stop when the link does. And data residency, when video, patient records or process telemetry are not permitted to leave the premises at all. Any one of these rules out a cloud-only design regardless of what the cloud version costs.
The trade is model size against physical reality. Edge hardware has a hard memory ceiling, updates have to be pushed across many devices rather than deployed once, and a fleet of a thousand units is a fleet of a thousand things that can drift out of sync. Organisations that run edge well treat it as one placement among three, alongside cloud and on-premise, assigned per workload.