MIT News discusses the environmental impact of generative AI in a two-part series. This article examines the technology's high resource consumption, while the next will focus on strategies to minimize its carbon footprint. Generative AI promises various benefits, such as enhancing productivity and advancing research, but its rapid growth leads to significant environmental concerns that are hard to quantify or mitigate. Training models with billions of parameters, like OpenAI's GPT-4, requires immense computational power, resulting in high electricity consumption and increased carbon emissions. In addition to electricity, significant water resources are needed for cooling the training and deployment hardware, which can strain local water supplies and disrupt ecosystems. The surge in generative AI applications also raises the demand for advanced computing hardware, adding further environmental strains from its production and transport. Elsa A. Olivetti, a professor at MIT, notes that the environmental impact extends beyond electricity usage to broader systemic consequences. Data centers, critical for hosting generative AI models, significantly contribute to environmental issues. These centers require a tremendous amount of energy for their operations.

For example, a single data center can house approximately 50, 000 servers, and the rising demand for data centers has led to rapid construction, far outpacing sustainable practices. Predictions suggest that by 2026, global data center electricity consumption could reach 1, 050 terawatts. Training large AI models, such as GPT-3, has serious energy demands, with studies estimating high electricity use and substantial carbon emissions. Moreover, energy demands not only persist during training but also during real-time use, as evidenced by each interaction with generative AI consuming much more electricity than traditional web searches. Current trends suggest that future generative AI models will require even more energy, with shorter lifespans leading to frequent new model releases. This cycle wastes the energy expended on previous versions. Data centers’ water consumption also poses significant environmental challenges, as significant cooling water is required, impacting local biodiversity. The environmental costs associated with producing the computing hardware itself—from GPUs to CPUs—are substantial due to complex manufacturing processes and resource extraction methods. Despite the unsustainable trajectory, authors propose that responsible development of generative AI necessitates a thorough evaluation of its environmental and societal costs alongside its perceived benefits. Olivetti emphasizes the need for a contextual understanding to measure and comprehend these evolving trade-offs.