diff --git a/website/app/page.ts b/website/app/page.ts
index c7510707f..016ba304d 100644
--- a/website/app/page.ts
+++ b/website/app/page.ts
@@ -246,6 +246,92 @@ function sourceWindow(title: string, sample: string) {
   `;
 }
 
+/**
+ * The landing page's intro video, self-hosted rather than embedded.
+ *
+ * The player is the browser's own, so it works with scripting disabled. The
+ * YouTube iframe this replaced could not, because its player needs JS inside
+ * the frame, so the section used to hide itself from a JS-off reader rather
+ * than show a broken embed.
+ *
+ * REPLACING THE VIDEO, for whoever does it next.
+ *
+ * The bytes live in the Cloudflare R2 bucket `webjs-videos`, served from
+ * videos.webjs.dev. The key carries no version, so a new cut overwrites
+ * `intro.mp4` in place and nothing in this file changes:
+ *
+ *   env -u CLOUDFLARE_API_TOKEN npx wrangler r2 object put \
+ *     webjs-videos/intro.mp4 --file <new-cut.mp4> \
+ *     --content-type video/mp4 \
+ *     --cache-control "public, max-age=86400, s-maxage=31536000" \
+ *     --remote
+ *
+ * Then purge that URL at the edge (Caching, then Purge Cache, then the custom
+ * single-file purge), or the old bytes keep serving for the `s-maxage` year.
+ *
+ * Three things about this are easy to get wrong.
+ *
+ * 1. R2 object metadata cannot be edited after upload. `Cache-Control` is set
+ *    at write time or not at all, so every upload has to pass it again. There
+ *    is no bucket-level setting, and the dashboard uploader has no field for
+ *    it, which is why an upload made there serves no header of its own.
+ *
+ * 2. The `env -u` is load bearing. A `CLOUDFLARE_API_TOKEN` in the environment
+ *    overrides wrangler's OAuth credentials, and that token carries no R2
+ *    permission, so the upload fails with an authentication error that names
+ *    nothing. Unsetting it for the one call falls back to the OAuth login.
+ *
+ * 3. A purge clears the edge and never a browser. That is why `max-age` is a
+ *    day rather than a year, and why the header is deliberately not
+ *    `immutable`: anyone holding the old file needs a way to pick the new one
+ *    up, and 24 hours is that way. `s-maxage` keeps the edge copy long lived
+ *    regardless, since the edge can be purged and a browser cannot.
+ *
+ * A versioned key (`intro-2026-08-21.mp4`) is the other valid shape. Take it
+ * and the tradeoff inverts: put `immutable` back, raise `max-age` to a year,
+ * and update the src below on every cut.
+ *
+ * The thumbnail at `intro-thumbnail.webp` is uploaded the same way, with
+ * `--content-type image/webp`. It is 1920 wide because the box is capped at
+ * max-w-3xl (about 718 CSS px), so that already covers a 2x display and a
+ * wider encode buys nothing a viewport can show.
+ */
+const INTRO_VIDEO = html`
+    <section class="intro-video pb-16">
+      <div class="max-w-3xl mx-auto px-6">
+        <!-- preload="none" fetches nothing but the thumbnail until someone
+             presses play. It is deliberately not "metadata", which is the
+             usual choice and what rubyonrails.org uses. Their file is 658
+             kbps and this one is 2315, and at 400 kbps the metadata read
+             starved the poster: the video had pulled 6.2 MB by the time the
+             thumbnail painted, 25.8 seconds in. The cost is that the controls
+             cannot show the duration until the first play.
+
+             playsinline stops iOS Safari taking the video fullscreen. The
+             width and height state the intrinsic size, which
+             this layout does not lean on (the wrapper is aspect-video and the
+             element is w-full h-full, so CSS fixes the ratio), but which keeps
+             the shape right if the stylesheet ever fails. The src is absolute and
+             cross-origin, so it must not be wrapped in asset(), which resolves
+             a public/ path and would mangle it. -->
+        <div class="aspect-video overflow-hidden border border-border-strong shadow-[var(--shadow)] bg-black">
+          <video
+            class="intro-video-player w-full h-full"
+            src="https://videos.webjs.dev/intro.mp4"
+            poster="https://videos.webjs.dev/intro-thumbnail.webp"
+            aria-label="WebJs introduction video"
+            width="1920"
+            height="1080"
+            preload="none"
+            playsinline
+            controls
+          ></video>
+        </div>
+      </div>
+    </section>
+`;
+
+
 export default function LandingPage() {
   return html`
     <style>
@@ -470,45 +556,7 @@ export default function LandingPage() {
       </div>
     </section>
 
-    <section class="intro-video pb-16">
-      <div class="max-w-3xl mx-auto px-6">
-        <!-- The frame is hidden until it fires load, over a black box.
-             A cross-origin iframe paints its OWN canvas, and YouTube's embed
-             sets its black background on body with no color-scheme declared,
-             so the black only lands once that stylesheet applies. Until then
-             some engines paint an opaque white canvas, which no background on
-             the iframe ELEMENT can cover, since the element background sits
-             behind that canvas. Hiding the frame sidesteps the whole question:
-             what it paints early is simply not on screen, and the box under it
-             is already the black the player settles on, so the reveal is
-             invisible. onload is a plain HTML attribute rather than a template
-             hole, because an @event drops at SSR and this page never hydrates. -->
-        <div class="aspect-video overflow-hidden border border-border-strong shadow-[var(--shadow)] bg-black">
-          <iframe
-            class="intro-video-frame w-full h-full invisible"
-            onload="this.classList.remove('invisible')"
-            src="https://www.youtube-nocookie.com/embed/XghCghezod4?rel=0"
-            title="WebJs introduction video"
-            loading="lazy"
-            allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share"
-            referrerpolicy="strict-origin-when-cross-origin"
-            allowfullscreen
-          ></iframe>
-        </div>
-        <!-- With JS off the whole section goes away. Revealing the frame
-             instead would show YouTube's own noscript error ("An error
-             occurred. Unable to execute JavaScript."), because their player
-             needs JS INSIDE the frame and nothing this page does can supply
-             it. An empty space reads better than a broken player.
-
-             The rule still applies from in here: a style element takes effect
-             wherever it sits, including inside the subtree it hides. And it
-             is safe in a PAGE, which never hydrates, so the client never
-             rebuilds this noscript body as live nodes. The same construct
-             inside a COMPONENT would apply on every JS-enabled visit. -->
-        <noscript><style>.intro-video { display: none }</style></noscript>
-      </div>
-    </section>
+    ${INTRO_VIDEO}
 
     <section class="py-16">
       <div class="max-w-6xl mx-auto px-6">
diff --git a/website/test/ssr/intro-video-ssr.test.ts b/website/test/ssr/intro-video-ssr.test.ts
index eac856803..092619270 100644
--- a/website/test/ssr/intro-video-ssr.test.ts
+++ b/website/test/ssr/intro-video-ssr.test.ts
@@ -1,14 +1,16 @@
 /**
- * The landing page's intro video is hidden until its frame fires load.
+ * The landing page's intro video is self-hosted, not embedded.
  *
- * A cross-origin iframe paints its own canvas before the embedded stylesheet
- * applies, and on some engines that canvas is opaque white, which no
- * background on the iframe element can cover. Hiding the frame until load
- * removes the question: what it paints early is off screen, and the box under
- * it is already black.
+ * It replaced a YouTube iframe whose every property was a workaround for
+ * being cross-origin: a frame hidden until load (an iframe paints its own
+ * canvas, opaque white on some engines, before the embedded stylesheet
+ * lands), a plain onload attribute to reveal it (this page never hydrates, so
+ * an event hole would be dropped at SSR), and a noscript rule that deleted
+ * the whole section for a JS-off reader (their player needs JS inside the
+ * frame, so revealing it would have shown their own error).
  *
- * Each assertion here is a piece that silently breaks the whole thing if it
- * goes missing, which is why they are pinned rather than left to review.
+ * A native player needs none of that, and the assertions below pin the pieces
+ * that silently break it if they go missing.
  */
 import test from 'node:test';
 import assert from 'node:assert/strict';
@@ -17,26 +19,48 @@ import LandingPage from '#app/page.ts';
 
 const render = () => renderToString(LandingPage());
 
-test('the intro frame ships hidden, over a black box', async () => {
+test('the intro video is a native player pointed at our own host', async () => {
   const out = await render();
-  assert.match(out, /class="intro-video-frame [^"]*\binvisible\b/, 'the frame must start hidden');
-  assert.match(out, /aspect-video[^"]*\bbg-black\b/, 'the box under it must be black');
+  assert.match(out, /<video[^>]*\ssrc="https:\/\/videos\.webjs\.dev\/intro\.mp4"/);
+  assert.match(out, /<video[^>]*\scontrols/, 'the controls are the whole no-JS story');
 });
 
-test('the frame reveals itself with a plain onload attribute', async () => {
+test('the player keeps the accessible name the embed had', async () => {
   const out = await render();
-  // A plain HTML attribute, not an @event hole: this page never hydrates, so
-  // a template event binding would be dropped at SSR and the frame would stay
-  // hidden forever.
-  assert.match(out, /onload="this\.classList\.remove\('invisible'\)"/);
+  // The iframe this replaced carried title="WebJs introduction video". Without
+  // a replacement the player is an unnamed media element to a screen reader.
+  assert.match(out, /<video[^>]*\saria-label="WebJs introduction video"/);
 });
 
-test('a JS-off reader gets no embed at all', async () => {
+test('the first paint shows a poster rather than a black box', async () => {
   const out = await render();
-  // Without JS the load handler never runs AND YouTube's player cannot run
-  // inside the frame either, so revealing it would show their own noscript
-  // error rather than a video. Hide the whole section instead. The rule must
-  // live in noscript, which a browser with scripting on parses as inert text.
-  assert.match(out, /<noscript><style>\.intro-video \{ display: none \}<\/style><\/noscript>/);
-  assert.match(out, /<section class="intro-video /, 'the rule needs its hook on the section');
+  // Without this the box is empty until someone presses play, which is worse
+  // than the embed it replaced.
+  assert.match(out, /<video[^>]*\sposter="https:\/\/videos\.webjs\.dev\/intro-thumbnail\.webp"/);
+});
+
+test('the bytes stay off the wire until someone plays it', async () => {
+  const out = await render();
+  // The file is 85 MB at 2315 kbps. Under preload=metadata it starved the
+  // poster on a 400 kbps link: 6.2 MB pulled before the thumbnail painted at
+  // 25.8s. "none" is the deliberate deviation from what rubyonrails.org does.
+  assert.match(out, /<video[^>]*\spreload="none"/);
+  // iOS Safari otherwise takes the video fullscreen the moment it plays.
+  assert.match(out, /<video[^>]*\splaysinline/);
+});
+
+test('none of the cross-origin workarounds survive', async () => {
+  const out = await render();
+  assert.doesNotMatch(out, /<iframe/, 'no embed of any kind');
+  assert.doesNotMatch(out, /youtube-nocookie|youtube\.com\/embed/i, 'no embed url');
+  assert.doesNotMatch(out, /intro-video-frame/);
+  assert.doesNotMatch(out, /onload="this\.classList\.remove/);
+});
+
+test('a JS-off reader gets the video, not a deleted section', async () => {
+  const out = await render();
+  // The old markup hid the section outright. A native player works with
+  // scripting disabled, so the rule that removed it has to be gone.
+  assert.doesNotMatch(out, /\.intro-video \{ display: none \}/);
+  assert.match(out, /<section class="intro-video /, 'the section still renders');
 });