What it really costs to render a faceless video on a cloud GPU

The setup

We used Modal, a serverless GPU platform, because it bills by the second and needs no servers to manage. The GPU was an NVIDIA L4 with 24 GB of memory, listed at $0.000222 per second when we tested. The code was TallyReel's unchanged video pipeline from Studio, the same one that runs on people's own computers, with AI pictures drawn by FLUX.2 klein on the GPU.

We fixed three test videos so every run did the same work:

• An English story with two recurring characters drawn from cast sheets, AI pictures, and the Natural voice (VoxCPM2).
• An Urdu story with AI pictures and the Natural voice.
• An English facts video with AI pictures and the Fast voice (Kokoro).

Time and GPU cost per video

The cold start is the first video after the machine wakes up: it has to load the models before it can work. After that, a story Short settled at about 3 minutes and about four cents.

GPU cost is the time on the card multiplied by the L4's per-second price. It leaves out CPU, memory and idle time, which is why the real bill was higher (see below). Measured 5 October 2026.

Where the time goes

Pictures are cheap on a GPU. Each AI picture took about 3 seconds on the L4, against about 50 seconds on an M1 Pro Mac. For a six-scene video that is the difference between under half a minute and five minutes.

The voice was the slow step. The Natural voice needed about 4 seconds of GPU work for every second of speech, so it dominated the story runs. We then found that the voice model was compiling itself at the start of every job, a step that pays off on long runs but not on a 30 second Short. Turning that off took the voice step from 307 seconds to 99 seconds, and a 30 second Natural voice Short dropped to about 2.5 minutes.

The Fast voice is why the facts video is the cheapest row: it reads in a fraction of the time, at the cost of sounding less natural.

The real bill was about 1.5 times the estimate

GPU seconds are not the whole bill. A container also uses CPU and memory, which are billed separately, and there are seconds where the GPU sits idle while models load or files upload. When we compared our GPU-only estimates with what Modal actually charged, the bill came to about 1.5 times the estimate.

To get to a price we could sell at, we added a 30 percent allowance on top of the real bill for everything else a video needs: redos when someone regenerates a scene, the AI that writes the script, content moderation, and storage for the finished file. That gives the figure in the title: about $0.08 per Short, all in.

Long videos cost more than their length suggests

We also ran a 5 minute documentary in 16:9 at 1080p, with AI pictures and the Natural voice, splitting the voice and scene work across several GPUs at once to keep the wait down. The two runs took about 17 and 18 minutes from start to finish, with GPU-time estimates of $0.51 and $0.72.

The real bills for 5 minute videos came to $1.50 to $1.60, about twice our estimate. Running many cards in parallel shortens the wait but adds startup and idle time on each card, and that overhead is billed. A 5 minute video is about ten Shorts of content, but its real bill was about as much as twenty Shorts at our all-in figure. If you price long videos per minute from Short-video numbers, you will lose money.

A faster card is not always cheaper

On 12 October 2026 we ran the English story from a cold start on both an L4 and an L40S, a faster and more expensive card. The L40S finished in 2 min 18 s against 4 min 27 s on the L4, but its GPU cost was $0.075 against $0.059. For a job that is not urgent, the slower card is the cheaper one. The L40S is worth it only when waiting is the bigger cost.

What we learned to watch for

• Always check for leftover machines after an interrupted run. Once, an interrupted test kept a GPU running in the background.
• Bake the model weights into the container image. Downloading them at the start of every job adds minutes and money to each cold start.
• Skip compile steps that only pay off on long jobs.
• Never redeploy the renderer while videos are rendering. A deploy can stop a job halfway, and you pay for the half that is thrown away.
• Re-price from the real bill, not from GPU seconds. Our GPU-only estimates were low by 1.5 times for Shorts and 2 times for long videos.

How this shapes TallyReel

These numbers are why TallyReel can offer cloud video making at all, and why the plans charge by length rather than per video: a minute of a long video costs more to make than a minute of Shorts. If you have a capable computer, Studio does the same work locally, and then the GPU cost is zero.

Models and links

• Modal pricing
• NVIDIA L4 Tensor Core GPU