Cache Transcoding is a Cloudflare prototype that compresses text files with Zstandard (zstd) before writing them to the cache disk, inside Pingora, the company’s proxy. In early tests, eligible files shrank to about one third of their original on-disk size. It was built by Aashi, an intern in Cloudflare’s 1.1.1.1 internship program, and the result is in a Cloudflare post dated September 1, 2026.
It is not a feature you switch on in a dashboard. The post describes only a prototype, with no release date. Even so, it shows where CDN costs are heading and gives you a good reason to review how your site delivers text.
What is Cache Transcoding and why did Cloudflare build it?
Cache Transcoding exists because memory and disk got more expensive. Cloudflare says RAM and disk prices climbed sharply over the past year, and Aashi’s question was: can we get more cache capacity out of the servers we already have? The answer was to compress what goes to disk instead of buying more disk.
According to the post, the mechanism has four steps:
- On a cache miss, the proxy encodes the response body with zstd before writing it.
- The cache metadata records that the stored form is the compressed one.
- Before the response leaves for the visitor, the body is decoded back to the original.
- In Tiered Cache, the compressed form travels between tiers. Decoding only happens on the last hop, the one that talks to the client.
Nothing changes for the visitor. For Cloudflare, the same server stores more objects and moves less data between data centers. The prototype uses zstd at level 3.
How much does zstd save in Cloudflare’s cache?
Zstd shrank eligible files to about one third of their original size on disk, on average, in the early tests. In a controlled test corpus, the compression ratio was 2.834x. Cloudflare measured the CPU cost on both sides:
| Measure | Result |
|---|---|
| Compression ratio (test corpus) | 2.834x |
| Encoding | 4.31 ns per byte, about 232 MB/s |
| Decoding | 1.56 ns per byte, about 641 MB/s |
| Test scale | more than 1 million requests, 10 cache servers |
| CPU increase | “minor”, in the range of “a few percent” |
Encoding runs once per cache fill. Decoding runs on every response served, which is why, according to the post, limiting the technique to popular content did not help.
One honest caveat from Cloudflare itself: the test corpus was deliberately compressible and “does not represent every text object on the Internet.” One third is an early test result, not a constant for the whole fleet.
Which files does Cache Transcoding apply to?
It applies to 200 OK responses of compressible text, with no Content-Encoding set and a known Content-Length of at least 4 KiB. HTML, CSS, JavaScript and JSON are the typical targets. Images, video and fonts are left out because they already arrive compressed.
The post’s list of exclusions is long and worth reading as a checklist of “what my site does that blocks this gain”:
- responses that already come compressed from the origin (text already served as gzip or Brotli is not recompressed);
- responses without a known length up front;
- requests with
Rangeand slice subrequests; - binary content.
The 4 KiB floor leaves out only about 1% of the bytes that would otherwise be eligible, so tiny files are not the problem.
Does this matter for my site’s traffic?
It matters less than the headline suggests, and Cloudflare’s own traffic sample shows why. According to the post, text was 67.3% of requests but only 22.3% of bytes. Media was 21.4% of requests and 63.3% of bytes.
A rough calculation, mine and not Cloudflare’s: if all that text shrank to a third, the saving would be on the order of 15% of stored bytes. That would already be a lot at CDN scale, but it is an optimistic ceiling, because it depends on a compressible corpus. And the heavy weight stays in images and video, which Cache Transcoding does not touch.
For a company website, the takeaway is different. In 2014, when I started in SEO, turning on gzip on the server was a fixed technical checklist item. In 2026 the item still holds, but the conversation moved up a level: the text you deliver, in what format and at what size, is now an infrastructure cost for whoever sits in the middle.
What can you do on your site on Monday?
Start by seeing what your server delivers today. With a terminal open, run curl -sI -H "Accept-Encoding: br, gzip, zstd" https://yourdomain.com/ and look at the Content-Encoding header in the response. Repeat with your main CSS and JavaScript files.
Then a short check:
- Confirm that HTML, CSS, JS and JSON go out with some
Content-Encoding(br or gzip). If they go out without one, that is your cheapest gain today, regardless of any prototype. - Measure the weight of the images and videos on your home page in the browser’s network panel. If media dominates, that is where the real savings are, as in the BEACON dataset, where the LCP bottleneck was not the download.
- If you use Cloudflare, follow their blog to see whether Cache Transcoding becomes a feature. Today there is no setting for it.
- Relate the result to Core Web Vitals: fewer bytes on delivery help loading, but the effect on ranking depends on much more than file size.
What else did Cloudflare’s interns ship?
Cache Transcoding is one of several intern projects cited in the 1.1.1.1 program’s one-year review, published on October 5, 2026. Cloudflare had announced a goal of hiring up to 1,111 interns in 2026, a number that winks at the 1.1.1.1 public DNS resolver. So far, that is 750 internships across 48 teams and nine offices, including Lisbon, according to the one-year review.
Two other pieces of work cited there relate to network infrastructure:
- BGP ORIGIN. Iliana Xygkou and Bryton Herdes measured how many networks rewrite BGP’s ORIGIN attribute to pull in traffic. On about 70% of IPv4 paths and 67% of IPv6 paths the value had been reset to IGP, and the authors argue for removing ORIGIN from route selection. It is in the full study.
- Post-quantum. Interns helped build an internal AI tool that finds cryptography in Cloudflare’s code, part of the company’s goal of post-quantum security by 2029.
Neither changes anything in your Monday operations. They are here as context for the kind of work that comes out of this program.
What does Cache Transcoding mean, in short, for anyone with a website?
Cache Transcoding is a cache compression prototype using zstd that, in Cloudflare’s early tests, cut eligible text files to about a third on disk, with a CPU cost described as a few percent. It has no release date, does not cover images or video, and was measured on a deliberately compressible corpus.
Your decision does not depend on it: check the Content-Encoding of HTML, CSS and JavaScript, find out how much of your page is media, and only then think about the CDN. If you need to decide where to cut weight first, the same logic applies to anyone blocking bots: knowing what really hits your server comes before any fine-tuning.