Generative AI, the marvel behind tools like OpenAI’s ChatGPT and image generators such as Midjourney and DALL-E, has long been celebrated for its creative potential. However, the open secret within the tech community is the astronomical amount of power these systems consume. A recent yet-to-be-peer-reviewed paper from a collaboration between AI developer Hugging Face and Carnegie Mellon University sheds light on the energy demands of various AI models, emphasizing the tangible carbon footprint associated with turning to AI over human artists.
The study discovered that the least efficient model, Stable Diffusion’s open source XL, consumed nearly as much power per image as it takes to fully charge a smartphone. Generating 1,000 images with this model results in carbon emissions equivalent to driving 4.1 miles in an average gasoline-powered passenger vehicle. Across all models tested, the average energy consumption for generating 1,000 images was 2.907 kWh, roughly equivalent to charging a phone’s battery to 24 percent per image. In contrast, text generation proved to be significantly less power-intensive, requiring only as much energy as three smartphone charges for 1,000 queries.
The extrapolation of these findings to a global scale raises concerns about the environmental impact of large-scale AI operations. Companies like OpenAI and Google are already grappling with soaring energy bills, with estimates suggesting that AI servers worldwide consume energy equivalent to what entire countries like Argentina use annually.
The environmental toll extends beyond direct energy consumption, as cooling these servers demands an astonishing amount of resources. Google’s 2023 Environmental Report revealed a 20 percent increase in water usage, reaching an astronomical 5.6 billion gallons to cool its AI servers.
As the world confronts the looming threat of climate catastrophe, the study serves as a stark reminder that, even on a per-image basis, the energy costs of utilizing generative AI tools are substantial. Questions linger about how these results compare to widely used AI image generators like Midjourney or OpenAI’s DALL-E, which were not included in the study. The broader challenge for the AI industry is clear: addressing and mitigating its carbon footprint to avoid contributing further to the global environmental crisis.