A P-frame ("predicted frame") is a video frame that's reconstructed by looking only at frames that came before it. The encoder describes most of it as motion vectors pointing into earlier frames, plus small corrections for anything the prediction didn't get exactly right. The "P" stands for "predicted" — predicted from the past, not from the future.
P-frames are the workhorse of video compression. In a typical streaming GOP of 60 frames, one might be a full i-frame, six or seven might be P-frames, and the rest are even-smaller b-frames. A P-frame is usually 5–10× smaller than an I-frame on natural-motion content, because most of the picture is unchanged from previously decoded frames and only needs a brief "moved this much over" description.
Where P-frames shine compared to B-frames: they introduce no encoding delay. A B-frame requires the encoder to peek at future frames, which means buffering them and waiting before producing output. A P-frame only looks backward, so it can be encoded and shipped the instant its source pixels arrive. That's why ultra-low-latency systems — WebRTC video calls, cloud gaming, live sports betting feeds — sometimes use only I and P frames ("IPPP" coding), giving up some compression efficiency in exchange for sub-100 ms glass-to-glass latency.
