It appears to be a SSE channel for each event stream returned, not one channel for all subsequent updates. So a PHP backend for instance could batch all the updates from one request, would open a new SSE channel, send the updates over that, then close it. With an in-process server like swoole or most things not PHP, you could presumably reuse the channel across requests in whatever framework-specific way makes that happen. Would probably need sticky sessions in any scaled-out deployment.

This is just what I can glean from the docs, I've never actually used datastar myself.