Abstract
Generative AI is a technology that has experienced a rate of adoption even greater than that of the internet. While companies like OpenAI and Anthropic focus on model development, garnnering massive amounts of attention from the press, the physical infrastructure of the data centers that host and train these models is equally critical. My thesis portfolio examines the growing environmental, social, and economic costs associated with this rapid expansion. As generative AI expands, these sprawling facilities have transitioned from simple storage rooms into multi-billion-dollar complexes that require massive amounts of electricity and water to operate. There also exists a technical struggle between balancing operational efficiency and cost. This technical challenge is inherently sociotechnical, as the drive for efficiency influences the location of these centers and their resource demands place immense pressure on local power grids and water supplies. My research investigates this issue from two perspectives; a technical exploration of storage efficiency and a sociotechnical analysis of the inequalities inherent in data center siting decisions. By addressing the resource intensity of AI through both optimized infrastructure and ethical siting, this work aims to provide a roadmap for a more sustainable digital future where technological growth doesn’t come at the expense of social equity or environmental health.
The technical research focuses on the input/output (I/O) bottleneck that often leaves expensive processors idle during large-scale model training. Storing massive quantities of training data solely on high-performance solid-state drives is both cost-prohibitive and energy-intensive, while slower archival storage can create massive bottlenecks in training. Traditional data management policies, such as Least Recently Used (LRU), fail in AI contexts because they cannot accommodate the sporadic patterns of Large Language Model (LLM) training routines, leading to massive cache misses. To address this, I proposed a hybrid tiered storage system and a custom algorithm to intelligently manage data placement based on access frequency. The methodology combined a review of existing memory hierarchies and industrial literature with the development of a simulation to model the tiering architecture. By analyzing specific I/O hotspots and inactivity periods, I aim to validate a system designed to balance energy consumption with training performance.
My STS research paper examines the political and economic factors that shape where this computational infrastructure is built. While these facilities are often described as part of a "weightless" cloud, they are in fact high-value real estate investments that consume vast amounts of local electricity and water. Local and state governments compete to attract these facilities, emphasizing promises of new jobs and economic growth. Yet, many data centers provide few long term jobs, as they essentially run automatically once they are built. Utilizing an environmental justice framework I analyzed how these decisions reflect existing socioeconomic inequalities. My research involved mapping existing U.S. data centers against indicators like income levels, water availability, and regulatory stringency to determine if siting patterns correlate with economic vulnerability. The results suggest that the burdens of AI infrastructure, such as water stress and energy costs, are often disproportionately borne by resource-scarce communities, while affluent regions reap the technological and financial gains.
In conclusion, this portfolio demonstrates that solving the environmental crisis posed by generative AI requires more than just technical fixes. While my technical research provides a blueprint for more energy efficient storage management, the STS analysis highlights that even the most efficient technology can contribute to environmental injustice if its physical footprint is not ethically managed. I believe these projects could be fruitful in bridging the gap between engineering optimization and social responsibility. Ultimately, technological advancement must be balanced with a fundamental reevaluation of our data consumption habits and the political and economic processes that govern our infrastructure.
