AI-generated video vs video streaming
Generating an AI video clip costs between 0.14 Wh and over 400 Wh depending on the model, a gap close to 3,000x. The only comparable independent measurement, published by Hugging Face researchers in September 2025, tests seven open models on the same GPU. This site uses their median value, 25.3 Wh, as an order of magnitude for a clip from a common open model. Here's why the gap is so wide, and what the figure this page used to show was actually worth.
A clip from a common open model (25.3 Wh, the median of seven measured models) equals about 20 minutes of streaming. The heaviest model in the panel (415 Wh) equals over 5 hours. The lightest (0.14 Wh) barely registers. Video stays the heaviest item in consumer AI, but a single number doesn't make sense here: model size and settings change everything.
Figures updated on September 23, 2026
Why does the range go from 0.14 Wh to over 400 Wh?
Because "generating a video" covers very different tasks. Julien Delavande, Regis Pierrard and Sasha Luccioni, researchers at Hugging Face, measured seven open video-generation models on the same Nvidia H100 GPU, using CodeCarbon, and published their results on September 23, 2025 (arXiv 2509.19222). Under each model's default settings: AnimateDiff uses 0.14 Wh per clip, LTX-Video 3.7 Wh, CogVideoX-2B 9.7 Wh, CogVideoX-5B 25.3 Wh, Mochi-1 52 Wh, WAN2.1 1.3B 90.5 Wh, and WAN2.1 14B, the largest model in the panel, 415 Wh. Their analytical model shows why: cost grows quadratically with resolution and frame count, and linearly with the number of denoising steps. Doubling a clip's length doesn't double its cost, it can quadruple it.
What was the figure this page used to show (940 Wh) actually worth?
It didn't trace back to a measurement. This page used to cite an order of magnitude of 940 Wh for a 5-second clip, with a source pointing to "How Hungry is AI?" (arXiv 2505.09598). That paper only measures language-model inference, not video generation: it says nothing on the topic. The figure couldn't be traced to a measurement that covers it. It's been removed and replaced by the median of the seven models above, 25.3 Wh, with its full source. The site's rule doesn't change: a published figure must link to a measurement a reader can check, not to an estimate carried over unchecked.
Is an hour of streaming a fair comparison?
It entirely depends on the model. Video streaming uses about 77 Wh per hour (IEA, network included). A clip from a common model (25.3 Wh) equals about 20 minutes of streaming. The heaviest measured model (WAN2.1 14B, 415 Wh) equals over 5 hours, a whole series. The lightest (AnimateDiff, 0.14 Wh) barely registers, a few seconds of streaming at most. The useful comparison isn't "AI video versus streaming" in general, it's "which model, for which use".
Related equivalents
Sources
- Delavande, Pierrard & Luccioni : « Video Killed the Energy Budget: Characterizing the Latency and Power Regimes of Open Text-to-Video Models » (arXiv 2509.19222, 23 sept. 2025, Hugging Face) : de 0,14 Wh (AnimateDiff) à 415 Wh (WAN2.1-T2V-14B) par clip, réglages par défaut de chaque modèle, mesuré par CodeCarbon sur GPU H100
- IEA : empreinte du streaming vidéo
Frequently asked
How much does an AI-generated video use?+
Between 0.14 Wh and over 400 Wh per clip depending on the model, measured across seven open models by Hugging Face researchers in September 2025 (CodeCarbon, H100 GPU, arXiv 2509.19222). This site uses the median, 25.3 Wh, as an order of magnitude for a common model.
Why such a gap between models?+
Cost grows quadratically with resolution and clip length, and linearly with the number of denoising steps. A light model like AnimateDiff (few steps, low resolution) and a model like WAN2.1 14B (high resolution, more steps) aren't doing the same task: they only share the same verb, "generate a video".
Is it worse than watching streaming?+
For a common model, a generated clip equals about 20 minutes of streaming (77 Wh/hour, IEA). For the heaviest model measured, the equivalent is over 5 hours. The answer depends entirely on the model, not on a general rule.
A number doesn't say whether to generate the video. A test clip re-run six times to pick the right one doesn't carry the same footprint as a single take used for a real campaign. That call belongs to a different job: the studio GHIS helps frame an AI use case before it gets deployed.