A server often knows which CSS and JS the HTML will need before the template finishes. Waiting for 200 plus <link rel=preload> wastes the think-time. HTTP 103 Early Hints (RFC 8297) sends an informational response with Link headers (rel=preload, rel=preconnect) while the final response is still cooking. Supporting CDNs (Cloudflare, some Google stacks, etc.) can forward 103 to the browser so it starts fetching on an idle connection.
This replaces a trendy llms.txt content-discovery idea with a real HTTP performance mechanism.
The sequence
client GET /product/42
103 Early Hints
Link: </app.css>; rel=preload; as=style
(time passes, DB)
200 OK
content-type: text/html
...The browser must not assume the final status. 103 is a hint. If the final page would not have used app.css, you wasted bandwidth. Hints should match what HTML will definitely need: the main stylesheet, a critical font, a preconnect to the image CDN.
HTTP/2 and HTTP/3 can push; Early Hints is not push. Push was poorly adopted. 103 is "please request these" using ordinary caching.
Operational catches
Origin must be allowed to emit 103 before the body. Many app servers buffer; you need an explicit flush of headers (Node, nginx http2_push is different). CDNs may generate 103 from a cache of known assets for a URL pattern when origin is slow — that is the Cloudflare-style product. If origin is fast, 103 helps less.
Don't hint personalized URLs that leak. Don't hint 50 images. Measure LCP, not "we enabled Early Hints" as a checkbox.
Support
Browsers and hops vary. A client that ignores 103 still works. That is the right degradation.
Read RFC 8297 (it is short) and your CDN's Early Hints doc. Then compare a slow-TTFB HTML document with and without 103 for the main CSS. If LCP does not move, the bottleneck was not the stylesheet, or 103 never left the edge. The status code is only useful if a hop in the middle understands it. Confirm with chrome://net-export or CDN logs, not with hope.