Are you speaking of chroma subsampling, or is there a property of the discrete cosine transform that makes it more effective on luma rather than chroma?

Probably chroma subsampling - storing color at lower resolution than luminance to take advantage of the aforementioned sensitivity difference. Since it’s stored at 1/4 resolution it can alone almost halve the file size.

Saying it’s the insight that led to JPEG seems wrong though, as DCT + quantization was (don’t quote me on this) the main technical breakthrough?