Reproducing LDM's ~512x Compression Number From Scratch

An interviewer asks you to justify, with arithmetic rather than a cited fact, why latent diffusion is described as making video generation roughly 512x cheaper for a 5-second, 24fps, 720p clip compressed by a factor of 8 in each of the temporal and spatial dimensions.

1. Compute the total pixel-value count for the uncompressed clip and for its latent representation, showing your work.
2. Derive the compression ratio and explain, in one sentence, why this translates into roughly proportional training/generation cost savings.
3. A colleague asks: "if 8x-8x-8x gives ~512x, wouldn't a 16x compression factor per dimension give an even better ~4096x, so why not just compress more aggressively?" What's the tradeoff being ignored in that question?