The comments of that article cite a couple things:

* The version of Caddy used for the test was missing several recent updates that significantly improve its performance. * Performance improvements for QUIC have always been about removing/recovering from dynamic network connections, and are about running connections at scale. Simple HTTP 1 requests in environments where there are far less shenanigans going on will always be fast. (lib)curl powered requests being single threaded requesting across the local network stack are not representative of the types of performance gains to expect in an HTTP/3 world.