Old-style SSR had three waterfalls: fetch all data before sending any HTML, load all JavaScript before hydrating anything, hydrate everything before anything responds. <Suspense> boundaries break all three, because each boundary can be sent and hydrated on its own.
Streaming HTML. With renderToPipeableStream or renderToReadableStream, the shell, everything outside boundaries, is sent as soon as it renders, with each pending boundary's fallback in place. When a boundary's data arrives, React sends its HTML in a hidden <div> at the end of the stream, plus an inline <script> that moves it into place. Boundaries complete in any order. The swap needs no React on the client, so content appears before the bundle loads. Since React 19.2, reveals are batched: at most one every 300 ms, as on the client.
Selective hydration. hydrateRoot hydrates the shell first; each boundary hydrates separately once its HTML and its code (say, a lazy chunk) have arrived, in tree order, yielding between pieces so the page stays responsive. If the user clicks inside a boundary that isn't hydrated yet, React hydrates that boundary synchronously in the click's capture phase, then dispatches the click normally. Other boundaries wait.
In React 19.3.1 that works only when the boundary can hydrate. If its code hasn't loaded, React stops the event and raises the boundary's priority; the click is lost, not replayed.
Errors. A component that throws on the server makes its boundary's fallback final in the HTML; the client renders that boundary from scratch and reports the error to onRecoverableError. Only an error in the shell fails the response.
Partial pre-rendering (19.2) applies the same split at build time: prerender with an aborted signal returns the shell as prelude plus a postponed state; at request time resume streams only the boundaries that were left.