A $0.30 AI video pipeline

We publish a vertical mockumentary series. Each shot is a stylized character talking to camera, documentary-confession style, with subtitles, a REC badge, and a low music bed. The whole thing costs about thirty cents a clip to generate, runs on one command, and — this is the part that took us two failures to earn — has no GPU in the loop that we manage ourselves.

Here’s the actual recipe.

The four ingredients

Motion — fal.ai Kling 2.1, image-to-video. We start from a single still of the character (produced by the design pipeline). We base64 it into a data: URI, hand it to fal’s hosted kling-video/v2.1/standard/image-to-video endpoint with a prompt (“documentary confession interview, blinks, breathes, gentle camera push-in”) and duration: "5", and let it animate. It’s a queue API: POST the job, poll the status_url, read the response_url for the final video.url, download. No image-to-video model runs on our hardware.

Voice — edge-tts. Free. Each character gets its own voice, rate, and pitch, so the CEO sounds different from the CTO. English audio, generated locally in parallel with the fal jobs.

Music — fal stable-audio. A short instrumental bed, prompted for the dry-witty mockumentary tone, mixed underneath at about 18% volume with a fade-out.

Finishing — ffmpeg, doing most of the actual work. This is where a clip becomes a clip: scale-and-crop to 1080×1920 vertical with a blurred, darkened copy of the frame as a fill background, the sharp character overlaid on top, a vignette, English subtitles burned in with drawtext, a red REC box, a name/role lower-third, then concat all the characters’ segments and mix the music across the full runtime. Optional lip-sync from a local Wav2Lip pass can be layered in for free when we want the mouth to track the line.

Our first real artifact, “The Coin Flip,” is four characters, about thirty-six seconds, voices plus motion plus music — assembled this way end to end.

The cost breakdown

The paid parts are the fal jobs: the Kling motion and the stable-audio bed. In practice that lands around $0.30–0.50 per clip. Everything else — edge-tts, Wav2Lip, all the ffmpeg compositing — is free and local. There is no per-hour GPU meter running in the background, which is the whole point.

Why hosted API instead of our own GPU pods

We didn’t start here. The obvious move for an AI company that already owns a GPU is to self-host the video and lip-sync models on rented pods and pay only for compute. We tried. Twice. Both attempts drowned in the part nobody advertises: booting a working pod — CUDA versions, model weights, Python dependency graphs, the specific fork of the specific model that actually runs — is its own full-time job, and a pod that won’t boot is a pod that still bills you while you debug it. Two rounds of that dependency hell was our answer.

Hosted image-to-video inverted the economics. There is nothing to boot. A cold job and a warm job cost the same. We tore down every pod. The tradeoff is honest and we wrote down the crossover: self-managed serverless only pays off at real, sustained volume — on the order of twenty to thirty clips a day. Below that, the hosted API wins outright, and we are below that. The engineering cost of keeping a pod bootable is invisible on a spec sheet and enormous in practice, and paying a few cents to make it someone else’s problem was the correct trade.

Where the GPU does live

We’re not anti-hardware — we’re anti unnecessary hardware. Our local RTX 4090 earns its keep on the high-frequency, low-latency work: image generation and the free Wav2Lip lip-sync pass. RunPod on-demand is reserved for genuinely heavy bursts, like sharper GPU lip-sync when we decide a clip needs it. And a local LLM anchor (a planned DGX Spark) is a separate question about owning inference, not about media. The rule underneath all of it is old and cheap: own the load that’s frequent and light; rent the load that’s heavy and rare. For a thirty-cent talking head, renting nothing at all turned out to be the heaviest optimization we made.