The 30% and the 52% do not fit together. Cloudflare says the static X25519 first guess was suboptimal for roughly 30% of the origin connections it has since measured [4], and that HelloRetryRequests ran at roughly 52% before the rollout and 3.7% during it [6]. A first guess wrong on three origins in ten does not produce a retry on half of handshakes, and the post states both figures without connecting them [17]. So the 48.3 point fall [13] is not a clean measurement of what any one key agreement algorithm costs. It measures what replacing a fixed guess with a per-origin one did to Cloudflare's origin population, whatever that population's client behaviour was beforehand.
Underneath, this is a cache with a cryptographic payload. Cloudflare is the TLS client on the origin hop [10], its ClientHello carries the hostname, the supported algorithm list and one or more keyshares [18], and after probing it sends the keyshare the origin said it prefers [5]. That needs per-origin state with a key and an expiry, and two situations defeat it: one hostname in front of a fleet whose members do not all run the same TLS stack, and an origin that changes its preference after the probe was taken. Cloudflare's post likens the retry to missing a shortcut in Mario Kart [20], which is generous, because you miss it again on every new connection [1].
For the latency figure to transfer to your own client, you need repeat connections to the same peer and somewhere to keep the answer. Cloudflare frames the work as covering 45 billion daily connections [11]; if every one of those were a new TLS 1.3 origin handshake, 48.3 points is about 21.7 billion handshakes a day that stopped paying the extra round trip [14]. Not all of them are new handshakes, so treat that as a ceiling. The p90 number travels better: a HelloRetryRequest costs exactly one round trip [2], so cutting more than 150 ms at p90 [7] means the retrying connections at p90 were crossing long distances [15]. Put the origin one metro from the edge that reaches it and the same retry buys you very little back.
On post-quantum, the retry delta is the wrong evidence. The narrower case is the one that holds: the TLS 1.3 happy path is a single round trip [2], Cloudflare now leads with the hybrid X25519MLKEM768 wherever the origin can speak it [5], and hundreds of thousands of domains have post-quantum origin connections that nobody configured [8]. Before this, getting that meant a manual choice at one end or the other, either enabled at Cloudflare or insisted on by the origin [9]. Note what the published wins measure: round trips and milliseconds [6][7]. ClientHello size is a separate cost line, and these numbers do not speak to it.
Cloudflare's stated reason for automating is that its 2029 target cannot depend on millions of site operators each becoming expert cryptographers [16][12]. That is the defensible argument in the post. The probe moved post-quantum from a setting somebody has to find into the output of a measurement, and operators who already forced it at the origin now hold an override instead of the only switch [9].
