Short-Form Video Has a 43% Problem. Here's How the Math Actually Works.
Research finds that 43% of short-form video goes unwatched. Separating prefetch waste from re-delivery shows which costs platforms can recover.
In February 2025, researchers from Synamedia, Inria, Sky and Ozyegin University presented research (opens in a new tab) at the Mile-High Video conference. Roughly 62,000 professionally produced short-form videos, more than 200 million views, from a popular news app running a Reels-style swipe feed.
Average view rate across the dataset: 57% (that’s 62% for the vertical mobile-format videos and 50% for the TV-style clips running in the same feed). Meaning 43% of the runtime, on average, goes unwatched.
Now try reading that as a bill: in a prefetch-driven feed, a large share of that unwatched video was already delivered when the swipe happened. Shipped, billed, skipped. Only about one video in ten got watched to the end.
And that number isn’t hidden, the paper is public. It just gets ignored, because every conversation is about the demand side: installs, ROAS, coin packs, retention curves. What it actually costs to deliver the video almost never comes up. For a business running on gaming margins, it’s a blind spot.
Delivery is inside your CAC math, whether you price it or not
Omdia puts global microdrama revenue at $11 billion for 2025 and expects $14 billion by the end of 2026, with roughly $3 billion of it outside China and the US now the largest international market. In February, Omdia reported (opens in a new tab) that microdramas have overtaken the big streaming apps on US mobile engagement.
The revenue is real. The margins…let’s talk about it.
At MIP London this year (opens in a new tab), the GM of COL Group International described the cost structure in one line: ninety percent of the budget goes to marketing. Holywater’s co-CEO Anatolii Kasianov put install costs up to 20 to 30 dollars. So the operating model is the gaming model: buy traffic at dozens of dollars a head, monetize a percentage of it, and live or die on whether LTV clears CAC before the retention curve hits the floor.
Now place delivery inside that model. Every install you buy streams video from day one. Free users, trial users, users who never buy a coin: they all watch. Delivery is a variable cost attached to every user your ad spend brings in, payer or not.
Where the 43% goes
A swipe feed feels instant because video players are optimized for the behavior. While someone watches this clip, it’s already downloading the next few. Swipe, and the video is just there. No spinner, no wait. Prefetch is what makes the format work.
It’s also what makes it bleed.
Swipe past a clip and the download already happened. Your CDN shipped it, you got billed for it, nobody watched it. Multiply that by real feed behavior, people skimming, sampling, skipping most of what’s offered, and you land on the paper’s 57% view rate.
The researchers call the unconsumed segments what they are: bandwidth and storage waste, and point to measurement studies putting the cost to streaming companies in the millions per year.
To make it concrete. A user who scrolls past eight of ten prefetched clips just cost you egress on eight downloads nobody watched. And that cost grows with engagement. The more your feed gets used, the more it wastes.
Nobody messed up here. This is the price of the model. Seamless swiping runs on speculative delivery, and the speculation misses almost half the time.
Two wastes, which you need to understand
“Short-form video” sounds like one delivery problem. It’s two, because of how people watch.
Swipe feeds, the TikTok and Reels pattern, are sampling behavior. Viewers skim and skip. Completion is low, so the bleed happens before the watch: prefetched bytes nobody watches.
Serialized micro-drama is the opposite. Nobody samples episode 47 of a show they’re hooked on. Viewers binge to the end, and your whole audience shows up at once when an episode drops. Completion is high, so prefetch waste barely registers. The bleed happens during the watch: the same hit episode, shipped in full, separately, to every viewer, in the same hour.
For example, on a drop with 100,000 concurrent viewers your CDN delivers the same few thousand unique segments tens of millions of times, and bills you per GB of delivery. I published the worked arithmetic a few weeks ago so I won’t dive into it again here. In short: your bill is unique content times audience times rate, and on a synchronized drop the audience multiplier is almost pure repetition.
One wastes by guessing wrong. The other wastes by repeating itself.
What changes when delivery stops meaning egress
Full disclosure: this is the half I work on, so read accordingly. I’ll keep it short.
Re-delivery exists because every viewer pulls their own copy from the CDN edge. You don’t have to do it that way. The CDN can ship each segment to a handful of viewers, and their browsers pass it to the next viewers directly, peer to peer. No peer around? The request just goes to the CDN like normal. That’s what we build at Sudden, a small script you add to your site.
The part I like: the thing breaking your bill is the thing that makes this work. More concurrent viewers means more browsers already holding the segment, so offload climbs right when traffic spikes. Drop night stops being the expensive night.
Short-form gets a side benefit. Peers can serve prefetched bytes too. The player keeps guessing what gets watched next, and it keeps guessing wrong; but a wrong guess served from another viewer’s browser never touched your CDN. The prefetch waste doesn’t go away. You just stop paying egress a big share of it.
The honest scope
Some limits your engineers will find anyway, so here they are. Sudden runs in web browsers. Not native apps yet, not live. App-only platform with no web player: nothing to attach to today. But app-first isn’t app-only. Most micro-drama platforms run real web surfaces. And, peer-assisted delivery addresses re-delivery. Prefetch waste belongs to the player and the ABR community. Anyone claiming one mechanism fixes both is selling.
What I’d pull this week
Two numbers settle this, and both are sitting in dashboards you already have.
First: your average view rate. How much of what you deliver actually gets watched. Whatever’s missing from 100% is your prefetch waste.
Second: how much of your CDN bill your newest episodes eat in the two days after they drop. If a handful of fresh titles are taking a third or more of your egress in their first 48 hours, your traffic is synchronized; thousands of people pulling the same segments at the same time. That’s your re-delivery waste. If your egress is spread evenly across the catalog instead, you don’t have much of it.
Pull both numbers and you know which half of your delivery bill is which; and which of the fixes in this piece even apply to you.
It’s worth the hour. The platforms that will thrive in this next cycle won’t be the ones that bought the most installs; they’ll be the ones that fixed their unit economics first.
Full credit to Burak Kara, Gwendal Simon, Ayse B. Demir, Ali C. Begen and Florence Agboma for putting production numbers behind something operators mostly knew as a feeling. Their call for open short-form delivery protocols deserves more attention than it’s getting.
If you run a vertical or short-form platform and want to put your own numbers through this, my DMs are open (opens in a new tab).
Want to run the numbers on your own traffic?
Talk to our founders about your audience, delivery stack, and where peer-assisted delivery could help.
