Most AI run somewhere else.
A camera or a sensor collects data, sends it to a server, and waits for an answer to come back.
That works until the connection is poor, the power budget is small, or the data cannot leave the site at all, which describes a great deal of defence, industrial and remote sensing work in India.
Running the model on the device instead solves those problems but creates a new one, which is that the hardware doing the computing has to be small enough and frugal enough to sit on a drone, a satellite payload or a factory sensor.
That is a chip design problem before it is a software problem.
