Energy-Efficient AI Networks and Environmental Impact

This white paper discusses the increasing power consumption and environmental impact of data center infrastructure driven by the demands of artificial intelligence (AI). It outlines how AI training networks, particularly those utilizing large language models, contribute significantly to energy usage, with data centers accounting for about 1% of global energy consumption. The paper examines the network power consumption associated with large-scale AI training implementations, specifically focusing on the transition from OpenAI's GPT-4 to GPT-5. It details the anticipated increase in computational requirements and power consumption, estimating that GPT-5 will require over 100,000 GPUs and consume approximately 122 MW of power. The authors propose energy-efficient optical innovations, such as energy-efficient optical (EEO) interfaces and optical circuit switches (OCSs), which could potentially reduce network power consumption by up to 80%. These innovations aim to enhance scalability and sustainability in AI training networks, addressing the critical need for energy efficiency in the face of growing computational demands.