HTTP/3 has been enabled on every plan since March. The headline improvement across all traffic is 4.1% on median page load, which is the kind of number that sounds like a rounding error — and averaged across all users, it is.
Segmented by connection quality, it stops being a rounding error.
| Connection | Median LCP change | p95 LCP change |
|---|---|---|
| Fibre, low loss | −1.2% | −2.8% |
| Cable / VDSL | −3.4% | −7.1% |
| 4G, good signal | −9.8% | −19.4% |
| 4G, congested cell | −17.2% | −34.6% |
| Satellite | −11.6% | −28.9% |
Why the shape looks like that
Two mechanisms do most of the work, and both only matter when the network is imperfect. First, QUIC's per-stream loss recovery means one lost packet stalls one resource rather than the whole connection — head-of-line blocking was TCP's problem, not HTTP/2's. Second, the combined transport and TLS handshake saves a full round trip, which is worth 40 ms on a congested mobile cell and 2 ms on fibre.
The operational surprises
- UDP amplification protection needed retuning. Our DDoS rules were tuned for a world where legitimate UDP to port 443 did not exist.
- Some corporate firewalls silently drop QUIC. Alt-Svc means clients fall back to HTTP/2 correctly, but it made a handful of debugging sessions confusing.
- Connection migration is genuinely good. Walking out of wifi range no longer resets an in-flight upload, which we did not expect to be as noticeable as it is.
It is on by default, there is nothing to configure, and you can turn it off per-site if you have a reason. We have not yet met the reason.