WebP vs. AVIF at the edge: a Cape Town store weighs CPU against bytes

WebP vs. AVIF at the edge: a Cape Town store weighs CPU against bytes
A furniture catalog served 3,400 product photos. They flipped on edge AVIF, expecting the headline 50% size win over WebP. It came — average image dropped from 142KB to 71KB.
But their PoP did on-the-fly transcoding, and AVIF encode is brutally CPU-heavy. First-request latency for an uncached size jumped to 1,400ms over slow Cape Town mobile.
They split the policy: AVIF for hero and listing images that get cached hard and reused thousands of times; WebP for user-uploaded review photos with low reuse, where encode cost never amortizes.
Median image bytes fell 38% overall, and uncached-image latency dropped back to 420ms.
AVIF wins on high-reuse assets. WebP wins when each variant is encoded once and rarely served again.
The number that mattered: 420ms for a fresh image, down from 1,400ms.