Speculation Rules, Speed Brain, classic edge prefetch and video segment lookahead — the CDN prefetch landscape measured against what each layer can actually predict.
Winner depends on your workload.
Winner depends on: which layer you want doing the guessing. Browser-driven Speculation Rules (deployable on any CDN, automated by Cloudflare's Speed Brain and Akamai's Ion tooling) now own next-page prediction; classic edge-to-origin object prefetch remains an Akamai-style config art; segment lookahead belongs to video delivery. Nobody's crystal ball is free — every wrong guess is billed bandwidth.
Three different things called prefetch
The word covers three mechanisms that share nothing but optimism. Browser prefetch — now formalized in the Speculation Rules API — has the browser fetch the likely next page before the click, guided by JSON rules the site (or its CDN) supplies. Edge-to-origin prefetch has the CDN pull embedded objects or likely-next content into its own cache ahead of demand — a technique Akamai has configured for two decades, retrieving a page's images and scripts while the HTML is still streaming to the client. And segment lookahead, in video delivery, pre-positions the next chunks of a stream. A comparison only makes sense per layer.
| Provider | Prefetch story | Mechanism |
|---|---|---|
| Cloudflare | Speed Brain applies Speculation Rules prefetch automatically; Workers can inject custom rules | Browser-layer, edge-automated; prefetched responses served from cache |
| Akamai | Ion documents Speculation Rules deployment via Property Manager or EdgeWorkers; classic embedded-object prefetch in delivery configs | Both layers: browser-rule injection and edge-to-origin prefetch |
| Fastly / any programmable edge | No boxed product; VCL or Compute can inject speculation rules or trigger origin prefetches | Build-your-own at either layer |
| Amazon CloudFront | No native prefetch feature; rules can be served like any header/body content, Functions can inject them | Browser-layer only, self-assembled |
| Video specialists | Segment prefetch / origin lookahead features on media-focused platforms | Delivery-layer, format-aware |
Speculation Rules moved the game to the browser
The structural change of the last two years: next-navigation prediction migrated from CDN magic into a Chromium browser API. A page declares candidate URLs (or CSS selectors matching them) with an eagerness level, and the browser decides when to prefetch or even fully prerender — which is why the API can improve TTFB on any CDN. The CDN's remaining jobs are distribution and judgment: injecting the rules without origin changes (a response-header behavior in Akamai's Property Manager, a Worker or Speed Brain on Cloudflare, thirty lines of VCL on Fastly) and making sure prefetched requests are answered from edge cache rather than punched through to the origin. Note the ecosystem caveat both vendors document: only Chromium-based browsers act on the rules today; everyone else silently ignores them.
The old art is not dead
Edge-to-origin prefetch — the CDN fetching a page's sub-resources or predictable next objects into its own cache — still earns its keep in two places. First, origins that are slow or far: warming the edge while the HTML streams hides origin latency that no browser hint can. Second, sequential access patterns machines generate — paginated APIs, ranged downloads, and above all video, where the next segment's URL is arithmetic, not prediction. That determinism is why segment lookahead is standard equipment on media platforms while general next-page prefetch remains probabilistic gambling.
The bill for being wrong
Every prefetch that is never used is bandwidth you paid to deliver — at both edge-to-client and, if it missed cache, origin-to-edge rates — plus analytics noise (prefetched pageviews that never happened) and, for prerender, client CPU and memory spent on a guess. Tune eagerness conservatively, exclude personalized and side-effectful URLs, verify your analytics stack distinguishes speculative loads, and measure the hit rate of the speculation itself: a prefetch that converts under a third of the time is often a net cost. The connection-level economics sit on top of the protocol work we covered in the HTTP/3 matrix, and the cache-hit mechanics in cache hit ratio, explained. Facts verified against provider documentation, July 2026.
Wondering whether speculation would convert on your traffic — or just inflate your bill? The assessment measures your navigation graph first.