67.6% of these clips sit on a single free hosting project

By Alpha Lay · 2026-09-26

When you build an index it is easy to focus on content and ignore where the files actually live. After breaking down hosting providers, two things were worth writing up.

Where the data comes from

Each case records two source fields:

• src (source page): where the content was originally published
• mhost (video host): the domain the video file actually loads from

These frequently disagree — which is the point.

Finding 1: Sources are highly concentrated; so is hosting

First, source pages:

87.4% of cases come from X (Twitter), with 8 source sites covering everything.

Now where the video files actually live (de-duplicated scope):

One host carries two thirds of the unique cases. And that domain, h3-field-notes-production.up.railway.app, is a default Railway project domain — Railway's free and trial-tier projects get suspended, then deleted, after extended inactivity.

Finding 2: Official tutorial demos and community work are two different populations

This is the more surprising part. Those 44 records on external-cdn.morphic.com come from a morphic.com tutorial, "how to write MiniMax H3 prompts." Comparing them against the community channels (Railway + beatapi, 960 cases):

The clips chosen for official demos run half as long as community work, with 2.6 times the vertical rate.

The contrast makes sense: a tutorial demo is a sample cut to illustrate one sentence — it has to be short and legible. Published community work, by contrast, has to be complete and play on a landscape timeline.

But there is a practical consequence: if you build your sense of "what AI video looks like" from official tutorial demos, your expectation will be systematically too short and too vertical. What the community is actually making is 15-second landscape clips.

Finding 3: Native X hosting is only 13.4%

Worth noting: video.twimg.com — X's own video hosting — accounts for only 13.4%. That is, the video files for the vast majority of cases are not on the original publishing platform at all, but have been moved elsewhere by third parties.

The reason is simple enough: X's video URLs carry time-limited signatures and are unsuitable for long-term indexing. Reposting parties migrate files to their own storage to keep links stable. For an indexer that is helpful (links last longer) — at the cost of one more intermediary that can vanish.

What you can do with this

If you depend on any AI video case index: ask where its video files live and whether there is a backup. One free host carrying 67.6% is an objective fragility.

If you use official tutorials to calibrate your aesthetic sense: know that demos are cut for explanation — median 8.1s, 38.6% vertical — while actual work is 15.2s and 14.7% vertical. Planning your content around demo durations will leave you a third too short.

If you produce paid media: the 0–8s range accounts for only 4.7% of community cases (45 of 960), yet 34.1% of official tutorial demos. That means public reference for ultra-short ads lives mainly in official tutorials, not in community work.

Method and limitations

• Host counts change over time: video links expire and domains change. This is a snapshot as of 2026-09-19 (the index generation date).
• We did not re-test link reachability: we recorded hosting domains but did not HTTP-check every file to see if it still plays. So "625 cases hosted on Railway" means "625 records show Railway hosting" — not "all 625 currently play."
• A small-sample warning: morphic has only 44 cases. It represents "the demos in this one tutorial," not morphic's overall output.
• All figures use the 1,274 index records; host statistics use the de-duplicated-by-original-publication scope.

Reproduce it

Every case card shows its source page and video hosting domain: https://cases.aishifu.shop/