Recently, researchers found that the carbon footprint of generative AI tools that can convert text prompts into images and videos is far higher than expected. This study, conducted by the team at the open-source AI platform Hugging Face, reveals the energy consumption issues of current text-to-video generators.
The study points out that as the length of the generated video increases, the energy demand grows fourfold. For example, the energy required to generate a six-second video is four times that of a three-second video. This phenomenon indicates a structural inefficiency in the energy consumption of current video generation technology, which urgently needs optimization design.
Image Source Note: The image was generated by AI, and the image licensing service is Midjourney
Experts warn that we do not have a comprehensive understanding of the true environmental impact of generative AI tools during their release. According to an analysis by MIT Technology Review, there are many gaps in the general understanding of AI energy consumption. For example, the energy required to generate an image of 1024x1024 pixels is equivalent to heating a microwave for five seconds, but the energy required to generate a video is much higher. The study found that generating a five-second video consumes energy equivalent to running a microwave for over an hour. This situation becomes more pronounced when generating longer videos, causing a significant increase in hardware and environmental costs.
Luckily, researchers have proposed some methods to reduce energy consumption, such as smart caching, reusing existing AI-generated content, and "pruning" inefficient training samples. Whether these methods are sufficient to reduce the power consumption of current AI tools remains to be seen. According to the latest research, energy use related to AI has already accounted for 20% of the global data center electricity demand.
At the same time, tech giants are investing hundreds of billions of dollars in infrastructure, sometimes even abandoning climate goals. Google admitted in its 2024 environmental impact report that its plan to achieve net-zero carbon emissions before 2030 is significantly behind schedule, mainly due to reliance on generative AI, leading to annual carbon emissions increasing by as much as 13%.
Earlier this year, Google launched its Veo3AI video generator, and within just seven weeks, users had created over 40 million videos. However, the environmental impact of the tool remains unknown, and Google has no incentive to investigate its contribution to carbon emissions, suggesting that the situation may be even worse than we imagine.
Key Points:
🌍 The carbon footprint of generative AI tools is much higher than expected, especially with video generators showing a fourfold increase in energy consumption.
⚡ The energy required to generate a five-second video is equivalent to running a microwave for an hour, and the energy demand rises rapidly as the video length increases.
🔧 Research has proposed methods to reduce energy consumption, but the issue of massive power consumption by AI tools remains severe, and tech companies also face challenges in meeting their environmental commitments.
