Cloudflare says it is now fastest in 74% of top networks

Cloudflare is now the fastest provider in 74% of the world's 1,000 largest networks, up from 60% in April, the company said in its Birthday Week performance update. It has also changed how it collects the measurements behind that claim.

Fourteen points and 150 networks

The increase works out as Cloudflare becoming the fastest provider on an additional 150 of those 1,000 networks, and as 38 more countries where it now ranks first. The network list is ranked by APNIC's estimate of user population, and performance is judged by connection time — how long a device takes to complete a TCP handshake when requesting content — then scored as a trimean of the 25th, 50th and 75th percentiles, a method chosen to blunt the effect of unusual outliers while still reflecting what most users actually get.

A second source of real-world data

The more consequential change is the sampling. Since 2021 Cloudflare has drawn its figures from its own branded error pages, which can run a small background request to several providers — Cloudflare, Amazon CloudFront, Google, Fastly and Akamai — and time each connection from the visitor's browser. This year it has added Challenge Pages, the full-page checks that verify a visitor before they are allowed through to a site. When a Web Application Firewall rule triggers one, a non-interactive measurement runs quietly in the background while the visitor is on the page, which widens the sample to everyday browsing without asking anyone to do anything extra.

Thin margins, and why method matters

Cloudflare says the Challenge Page measurements are running on only a small fraction of eligible free Challenge Pages for now, and that the sampling rate will only rise if the extra data proves useful and does not add noticeable delay. That restraint matters, because the races are decided in milliseconds. In many networks the leading providers are separated by one or two — Cloudflare offers the example of itself at 50 ms against Fastly at 51 — which is a gap that ordinary day-to-day variation can erase. More measurements make the ranking less sensitive to a quiet week, which is a fair argument and also the obvious self-interest of a company that publishes its own scoreboard.

Our opinion

The headline is the 74%, but the honest story is the second data source, because a provider that measures itself sets its own exam paper. Error pages were always a strange vantage point: they only exist when something has already gone wrong, and the users who see them are not a random sample of the internet. Borrowing the Challenge Page as a measuring instrument is clever precisely because those pages are already scattered across ordinary websites and ordinary networks, so the sample stops being an artefact of failure. It also reframes what the number means. A fourteen-point jump in five months is the kind of figure that invites suspicion of a methodology change, and Cloudflare deserves credit for saying plainly that the improvement coincides with one. The remaining question is the one no self-published benchmark can answer: what the same network list looks like when somebody else holds the stopwatch.