AI-Driven Electricity Usage Forecasting Shows Industry is Far from Achieving Net-Zero Goals

As the artificial intelligence sector grows swiftly, concerns about the ecological effects of data centers are increasingly being discussed. New projections indicate that the industry may fall short of achieving net-zero emissions by 2030.

Fenki Yu and researchers from Cornell University in New York have evaluated the potential energy, water, and carbon consumption of current leading AI servers by 2030, under various growth scenarios and specific U.S. data center locations. Their analysis integrates anticipated chip production, server energy demands, and cooling efficiency, coupled with state power grid data. While not all AI enterprises have declared net-zero objectives, major tech firms involved in AI, like Google, Microsoft, and Meta, have set targets for 2030.

“The rapid expansion of AI computing is fundamentally altering everything,” says Yu. “We’re striving to understand the implications of this growth.”

The researchers estimate that establishing AI servers in the U.S. may require between 731 million to 1.125 billion cubic meters of additional water by 2030, along with greenhouse gas emissions ranging from 24 million to 44 million tons of carbon dioxide each year. These estimates hinge on the pace of AI demand growth, the actual number of advanced servers that can be produced, and the sites of new U.S. data centers.

To address these issues, the researchers modeled five scenarios based on varying growth rates and outlined potential measures to minimize the impact. “The top priority is location,” Yu explains. By situating data centers in Midwestern states with abundant water resources and a significant share of renewable energy in the power grid, the environmental fallout can be mitigated. The team also emphasizes that transitioning to decarbonized energy sources and enhancing efficiency in computing and cooling processes are essential strategies for minimizing environmental impact. Collectively, these three measures could potentially lower industry emissions by 73% and reduce water usage by 86%.

However, public resistance may disrupt these predictions, particularly regarding the environmental ramifications of establishing data centers. In Virginia, where 1/8 of the world’s data centers are located, residents have voiced opposition to upcoming construction plans, citing concerns over water resources and broader environmental impacts. Similar petitions against data centers have arisen in Pennsylvania, Texas, Arizona, California, and Oregon. As per Data Center Watch, a firm that monitors data center developments, local opposition is stalling approximately $64 billion worth of projects. Even where certain locations successfully deny data center projects, questions remain regarding their potential power and water consumption.

This new research is viewed cautiously by those analyzing and quantifying AI’s environmental effects. “The AI field evolves so quickly that making accurate future predictions is incredibly challenging,” says Sasha Luccioni from the AI company Hugging Face. “As mentioned by the authors, breakthroughs in the industry can radically alter computing and energy needs, reminiscent of DeepSeek’s innovative techniques that reduced reliance on brute-force calculations.”

Chris Priest from the University of Bristol in the UK concurs, highlighting the necessity for increased investment in renewable energy infrastructure and the importance of data center placement. “I believe their projections for water usage in direct cooling of AI data centers are rather pessimistic,” he remarks, suggesting that the model’s “best case” scenario aligns more closely with “business as usual” for contemporary data centers.

Luccioni believes the paper underscores a vital missing element in the AI ecosystem: “greater transparency.” She notes that this issue can be addressed by “mandating model developers to track and disclose their computing and energy consumption, share this information with users and policymakers, and commit to reducing overall environmental impacts, including emissions.”