TLDR: A new study by Hugging Face researchers indicates that the energy consumption of AI video generation scales non-linearly, with power usage quadrupling when video duration doubles. This raises significant concerns about the environmental impact and carbon footprint of generative AI, particularly for text-to-video systems.
New research from the open-source AI platform Hugging Face has unveiled a concerning trend in the energy consumption of generative AI, specifically in the realm of video creation. Published on September 26, 2025, the study, titled ‘Video Killed the Energy Budget,’ demonstrates that the energy demands of text-to-video generators increase exponentially rather than in direct proportion to content length. This non-linear scaling means that doubling a video’s duration can lead to a quadrupling of its associated energy consumption.
The implications of these findings are substantial. For instance, generating a mere six-second AI video clip requires nearly four times the energy of a three-second clip. To put this into perspective, producing a five-second video with AI demands an amount of energy comparable to running a standard microwave for over an hour.
Researchers emphasized the ‘structural inefficiency of current video diffusion pipelines and the urgent need for efficiency-oriented design.’ This study emerges amidst growing warnings from experts about the environmental consequences of generative AI technologies, which are often deployed without a complete understanding of their ecological footprint. A recent analysis by MIT Technology Review highlighted this knowledge gap, stating that ‘the common understanding of AI’s energy consumption is full of holes.’
The energy intensity of AI is becoming a critical global issue. According to one recent study, AI-related activities now account for 20 percent of the total power demand from all global datacenters. Industry leaders are also voicing concerns; former Google CEO Eric Schmidt recently posited that AI’s natural limit is not silicon chips but electricity, predicting that the U.S. might require the equivalent of 92 nuclear power plants to meet its AI ambitions. Major tech companies like Microsoft and OpenAI are already securing power through nuclear energy deals and investments in fusion startups.
While public discourse often focuses on the dangers of deepfakes and misinformation, experts caution that the environmental costs of generative video could trigger a crisis of its own. The Hugging Face team found that GPU usage constitutes over 80 percent of the energy consumption in every model tested, with larger systems consuming up to 3,000 times more power than lighter ones. This escalating demand threatens to undermine global climate commitments, as evidenced by Google’s 2024 report, which showed a 13 percent rise in carbon emissions largely attributed to AI.
Also Read:
- Nvidia Democratizes Realistic Facial Animation with Open-Source Audio2Face AI
- Deloitte Study: Consumer Trust in AI Innovators Crucial Amidst Surging Generative AI Adoption and Privacy Concerns
Sasha Luccioni of Hugging Face underscored the severity, stating, ‘Video diffusion is far more costly than text or image generation… highlighting the need for hardware-aware optimizations and sustainable model design.’ The report suggests potential strategies to mitigate this impact, including diffusion caching, pruning inefficient training data, and quantization. However, it remains uncertain whether these measures will be sufficient to make a meaningful impact on the overall electricity consumption of current AI systems.
