>A human might notice 45ms in a shooter, but not when just surfing the web.

You are forgetting round-trip request latency, handshakes, and dependent resources (don't know what to fetch until receiving the first thing, then the second level of things, etc., this can stack to multiple seconds).

High latency is very bad for browsing. The main things it is fine for are streaming audio/video (non-interactive), large downloads, and backing up/uploading files.

You’re not wrong, but not relevant either. If your RTT to the server is 15ms, then connecting to that server with TLS1.2 will take 4×15ms = 60ms. On Starlink that will currently take 4×45ms = 180ms. Yes, 180ms is three times 60ms, but to a human that is irrelevant. Humans can’t detect time intervals of less than 100ms, and can only barely detect intervals less than 200ms. In a direct comparison between the two, humans will rate the 60ms connection as instantaneous and the slower 180ms connection as nearly instantaneous. This makes them “close” as the article described, especially when compared to old–school satellite links where the RTT can be insane.

And the handshake time doesn’t compound; there is no need to open a new connection to the server for every resource needed by a webpage. The web browser just keeps reusing the same connection until it runs out of resources to fetch from the server.

> And the handshake time doesn’t compound; there is no need to open a new connection to the server for every resource needed by a webpage. The web browser just keeps reusing the same connection until it runs out of resources to fetch from the server.

Third party analytics and other resources often means it compounds even for handshaking, but dependent resources can compound too even on the same server even if they don't each need a handshake.