A single 2-megapixel camera recording around the clock generates a volume of data that, at the back end, takes racks of servers to analyse. More and more projects are moving the judgement forward instead: the camera classifies people, vehicles and boundary crossings locally and sends back only the clips and structured data that matter.

The appeal of this edge architecture is easy to see. Bandwidth use falls sharply, especially where many high-definition streams converge. Load on central servers is shared. And if the network drops, the front end keeps analysing and storing locally, so nothing is lost.

The trade-offs are just as real. Front-end compute means more heat and higher power draw, so enclosure thermal design, chip selection and power headroom all need re-checking. Algorithm firmware has to stay upgradeable, or the hardware is still capable while the model is two years out of date. And when many devices are deployed together, batch firmware upgrades and version consistency become an operational task of their own.

To judge whether an edge solution is mature, do not ask how many algorithms it can demonstrate. Ask how stable it is after six months at full load, whether firmware upgrades go smoothly, and whether the device falls back to basic recording when something goes wrong.