In the cloud you can often dodge the tradeoff
In cloud inference, engineers frequently get to sidestep hard tradeoffs simply by spending more: a bigger GPU instance buys lower latency without touching accuracy, and if power costs more, that mostly shows up as a bigger bill someone else notices later, not a decision the model architecture has to make. Edge deployment removes that escape hatch. Because memory, compute, and battery are all fixed by the device in the user's hand, improving one of latency, accuracy, or power on an edge deployment usually means giving up ground on one of the other two, and the engineer has to choose explicitly rather than just paying for more headroom.
A larger, more accurate model takes longer to run and draws more power per inference. A model compressed hard enough to run instantly and sip battery will typically lose some accuracy along the way. Running inference less frequently, say checking a sensor every second instead of continuously, saves power but adds latency to how quickly the system reacts to a real change. There is no universally 'best' point on this triangle, only the point that's right for a given product.