AI & Agents
28 Videos for About $1 of GPU Time. The Hook Is Still the Expensive Part.
I rendered 28 short videos overnight on rented RunPod serverless GPUs for roughly 3 cents a reel. Production cost collapsed. My Instagram average watch time stayed at 3.8 seconds.
RunPod serverlessAI video costcost per reel
I rendered 28 reels overnight, and when I checked the bill it was about $1. My cap for the batch was $4. The cost per reel worked out to roughly 3 cents.
My Instagram average watch time at that point was 3.8 seconds. So I had made the part that costs money nearly free, and the part that decides whether anyone watches was exactly as hard as before.
This is the account of the pipeline, the numbers I measured, the three things that broke, and why I think the cheap half of video is the less interesting half.
How much does it cost to render an AI video on RunPod serverless?
Here is what I measured on my own run, using RunPod serverless endpoints that scale to zero when idle:
| Stage | Hardware | Billing rate | Time per reel |
|---|---|---|---|
| Voice | GPU (24 GB class) | $0.69/hr | about 2 min |
| Render | 16 vCPU CPU workers | $0.67/hr | about 40 s |
Add those up and a reel costs about $0.03. Twenty-eight reels is a little under a dollar, and I call it “about $1” because I did not itemise every small charge. The first batch finished in around 25 minutes of wall-clock time, limited by the account’s cap of 5 workers.
One detail helped the maths: serverless did not bill me for pulling the container image. If it had, my first runs, which were mostly debugging, would have cost more than the final batch.
What does the pipeline look like?
I will keep this at the architecture level, since the repo is private and I am not offering it to anyone.
- Voice runs on a GPU. The speech and word-timing work is the heavy model step.
- Render runs on CPU. Compositing the video in a headless browser does not need a GPU, so paying GPU rates for it would be waste.
- A storage bucket carries the files. The job payload limit is around 10 MB, so audio and video go through presigned bucket URLs, and the job only passes references.
- Everything is scaled to zero. Nothing bills while I am not rendering.
- Footage is stock. Clips come from a free stock library, picked per scene by ID.
I also added a light “edited finish” at the end: slight drift, a wipe-in, a colour grade, a vignette and a little grain. The reason was that plain stock footage looked obviously dropped in, and a bit of finish helped it look intentional.
What went wrong?
Three things, each with a clear fix.
1. The render crashed because Python counted the wrong CPUs. The call that asks how many cores exist returns the host machine’s count, not what the container is allowed to use. My renderer opened 30-plus browser tabs in parallel on a container with a small CPU quota, and the render died. The fix was to read the container’s actual CPU limit from its cgroup settings.
2. Job results expire. RunPod keeps a finished job’s result for roughly 30 minutes. A result that has not been saved by then is gone, so now I save the QA output immediately after the voice stage, before doing anything else.
3. My home network dropped connections to the bucket. Connections to the bucket timed out intermittently. I added retries and a fetch-only mode so I could resume pulling files without redoing the work.
There was also a console quirk worth knowing. The “latest generation” GPU tab silently excluded some older cards I would have been happy with, so cheaper options were hiding in plain sight. I only spotted this because the price did not match what I expected.
Did the second pass cost more?
A little. After listening to the batch I reworded a few lines and swapped footage on several reels, and I re-ran only the affected pieces. The retouches cost about $0.30 in total. Re-rendering selectively is where having separate voice and render stages paid off.
Why is the hook the expensive part?
Because production cost fell and attention did not.
On my own Instagram account, the posts that did well in the first two weeks were news-jacking and “how normal people use AI” content. The ones that did badly were static images, memes and PM-insider reels about tickets, PRDs and dashboards. The average watch time was 3.8 seconds, which I read as a hook problem.
A hook cannot be rendered cheaper. It needs a decision about what the first second promises, and that decision costs thinking. My working guess, and it is only a guess, is that the 3.8 seconds will not move because I produce more reels. It will move if I spend my effort on the first line of each script.
I have a longer account of making a month of video across three machines in one night if you want the earlier version of this story.
A checklist if you try this
- Split GPU work from CPU work and bill each at its own rate.
- Pass files through a bucket. Pass only references through the job.
- Read container limits from cgroups, not from the CPU-count call.
- Save results the moment a job returns.
- Retry network calls and keep a resume mode.
- Cap your spend before the first batch.
- Look at your watch time before you scale your volume.
I will find out within a couple of weeks whether volume or hooks move the number. I would put money on hooks, but I have not yet earned that bet.