Researchers embracing the face of open source AI platforms have found that due to nonlinear energy scaling, the carbon footprint of the generated AI tools is significantly worse than previously estimated.
In a newly published paper, the researchers detailed how the energy demand from text to video increases exponentially, rather than directly proportional to the length of the content. This study established that when the duration of the generated video doubled, the associated energy consumption is quadrant. To illustrate this principle, this paper provides a concrete example. Creating a 6-second video clip with AI requires 4 times the energy that produces a 3-second clip. “These findings highlight both the structural inefficiencies of current video diffusion pipelines and the urgent need for efficiency-oriented design,” the researchers concluded in their paper.
This study emerges in warnings from experts that generative AI technologies are being deployed without a complete understanding of their environmental impact. A recent analysis by the MIT Technology Review supports this concern, saying, “A common understanding of AI energy consumption is full of holes.” When comparing different types of generation tools, gaps in understanding are important. Creating a single 1,024 x 1,024 pixel image using an AI generator will consume the equivalent of heating something in the microwave for 5 seconds, but the video requirements will be orders of magnitude larger.
A study of embracing faces found that generating only a five-second video clip requires energy comparable to running standard microwaves for over an hour. This disparity underlines the intensive nature of video generation. Nonlinear scaling means that longer video clips escalate at even faster power consumption. According to the paper, this trajectory means “rapidly increasing hardware and environmental costs” for users and developers of these technologies.
There are potential ways to mitigate these high energy needs. Researchers propose several strategies, including implementing intelligent caching systems and the practice of reusing content generated by existing AI to avoid redundant processing. Another proposed technique is “pruning.” This involves systematically identifying and removing inefficient examples from the large dataset used to train AI models. This process helps streamline the model and reduce the operational energy footprint during generational tasks.
However, it remains unclear whether these efficiency measures are sufficient to have a meaningful impact on the overall power consumption of current AI systems. The scale of the problem is already quite large. Data from a recent survey shows that AI-related activities currently account for 20% of the total electricity demand from all global data centres. In response to growing demand for AI, major technology companies are investing tens of millions of dollars in building new infrastructure.
Google's 2024 Environmental Impact Report revealed that the company is significantly behind its plan to achieve net zero carbon emissions by 2030. The report revealed a 13% increase in carbon emissions compared to the previous year. Earlier this year, Google released the VEO 3 AI video generator. The company later announced that users had used the tool to create more than 40 million videos within the first seven weeks of availability. Specific environmental charges for VEO 3 have not been revealed.
Featured Image Credits
