Generating a 5-second AI video consumes energy equivalent to an hour of microwave operation – study

New research from the Hugging Face team has revealed significant energy efficiency issues with generative AI systems that create videos from text queries, Futurism reports. The researchers found that increasing the length of a video results in a disproportionate increase in energy consumption: a 6-second clip uses four times more energy than a 3-second clip.

The authors of the work emphasize that such dependence indicates a structural inefficiency of modern video generation algorithms. For comparison, the creation of one image with a size of 1024×1024 pixels is equivalent to 5 seconds of operation of a microwave oven, while the generation of one 5-second video corresponds to more than an hour of its continuous operation. As the duration of the videos increases, the costs increase even faster, which means significant hardware and environmental consequences.

The researchers suggest several ways to reduce the load: using caching, reusing already generated materials, and so-called "pruning" - removing ineffective examples from training data sets. At the same time, the scientists admit that these measures may not be enough to significantly reduce power consumption.

Other studies suggest that AI systems already account for about 20% of global data center energy consumption. Big tech companies are investing tens of billions of dollars in infrastructure expansion, making it harder to meet climate goals. For example, Google’s 2024 report found that the company’s carbon emissions increased by 13% compared to the previous year, largely due to the implementation of generative AI solutions.

At the same time, Google introduced the video generator Veo 3, whose users created more than 40 million videos in just 7 weeks. The company does not disclose the environmental impact of this product, but researchers suggest that it could be quite significant.