Basically the way modern video codecs work is that they only fully encode every Nth frame, where N is some smallish integer - 15 or 30 is common. From one keyframe to the next, they essentially encode a diff from the prior frame. The idea is that most of the regions of the frame have very little change over short timescales.

It's a tradeoff. More keyframes means faster seeking (when you seek in a video, it has to start from a keyframe too, for the same reasons), but larger data rates for the same image quality.