Harnessing AI’s Potential to Transform Energy Systems and Mitigate Demand Challenges

The rapid expansion of AI-driven computing centers is leading to an unprecedented increase in electricity demand, which poses significant challenges for power grids and may hinder climate initiatives. However, the same artificial intelligence technologies hold the potential to transform energy systems, accelerating the shift towards sustainable power sources. William H. Green, director of the MIT Energy Initiative, emphasized the potential for monumental changes in the economy during MITEI’s Spring Symposium titled ‘AI and Energy: Peril and Promise.’ This event gathered experts from various sectors to discuss solutions to the challenges surrounding electric supply and achieving clean energy goals while maximizing the advantages of AI. The symposium underscored AI’s substantial energy consumption. After years of stable electricity demand in the U.S., computing centers now account for around 4% of national electricity use. Projections indicate that this could escalate to 12-15% by 2030, largely due to AI applications. Vijay Gadepally from MIT’s Lincoln Laboratory highlighted the alarming rate of energy consumption related to AI. He stated, ‘The power required for sustaining some of these large models is doubling almost every three months.’ For instance, a single conversation with ChatGPT consumes as much electricity as charging a smartphone, while generating an image requires cooling that equates to about a bottle of water. New facilities demanding 50 to 100 megawatts are emerging globally, fueled by both casual users and institutional research reliant on advanced language models like ChatGPT and Gemini. Gadepally pointed to OpenAI’s CEO Sam Altman’s testimony, noting that ‘the cost of intelligence will converge with the cost of energy.’ Evelyn Wang, MIT’s vice president for energy and climate, remarked on the dual nature of AI’s energy demands. ‘While these demands pose significant challenges, they also present an opportunity to leverage AI’s computational power for climate solutions.’ She noted that innovations in AI and data centers—such as efficiency improvements and clean power technologies—could benefit multiple sectors beyond computing. During the symposium, various strategies were discussed to tackle the AI-energy challenge. Some panelists suggested that although AI might initially increase emissions, its optimization capabilities could significantly reduce emissions post-2030 through enhanced power systems and accelerated clean technology advancements. Research by Emre Gençer revealed that regional differences in clean electricity costs can influence computing center operations. For example, areas in central U.S. benefit from lower costs due to abundant solar and wind resources. However, achieving zero-emission power would necessitate extensive battery deployment—five to ten times more than moderate carbon scenarios—which could elevate costs significantly. ‘To achieve zero emissions with reliable power, we must explore technologies beyond renewables and batteries,’ Gençer explained, suggesting that long-duration storage technologies and small modular reactors are essential complements. Kathryn Biegel from Constellation Energy noted a renewed interest in nuclear power due to the demands of data centers. Her company is restarting the reactor at the former Three Mile Island site to address this growing need for reliable, carbon-free electricity. Artificial intelligence is poised to enhance power systems significantly. Priya Donti from MIT showcased how AI could optimize power grid operations by incorporating physics-based constraints into neural networks, potentially solving complex power flow issues at unprecedented speeds. Moreover, AI has already demonstrated its ability to lower carbon emissions. Antonia Gawel from Google shared examples, such as Google Maps’ fuel-efficient routing feature that has led to over 2.9 million metric tons of greenhouse gas reductions—equivalent to removing 650,000 fuel-powered cars from the roads for a year. The potential of AI in accelerating materials discovery for energy applications was highlighted by Rafael Gómez-Bombarelli. He explained that AI-supervised models can efficiently connect structure to property, facilitating the development of essential materials for both computing and energy efficiency. Throughout the symposium, participants emphasized the importance of balancing rapid AI deployment with environmental considerations. While training AI models garners much attention, Dustin Demetriou from IBM pointed out that ’80 percent of the environmental footprint is estimated to be due to inferencing,’ indicating a need for efficiency across all AI applications. Emma Strubell from Carnegie Mellon University warned about Jevons’ paradox, where efficiency gains can lead to increased overall resource consumption rather than reductions. She advocated for treating electricity in data centers as a limited resource that requires careful allocation across diverse applications. Several presenters explored innovative methods for integrating renewable energy sources with existing grid infrastructure, proposing hybrid solutions that combine clean installations with natural gas plants already connected to the grid. Such strategies could significantly enhance clean capacity across the U.S. while minimizing impacts on reliability. The symposium affirmed MIT’s pivotal role in addressing the AI-electricity challenge. Green announced a new MITEI program focused on computing centers and energy solutions, aiming to tackle complex problems from power sources to algorithms that provide customer value. Participants expressed their priorities for MIT’s research efforts, with real-time polling identifying ‘data center and grid integration issues’ as a top concern, followed closely by ‘AI for advanced material discovery in energy.’ Most attendees viewed AI’s potential in relation to power as a ‘promise’ rather than a ‘peril,’ although many remained uncertain about its ultimate impact. When discussing power supply priorities for computing facilities, half cited carbon intensity as their primary concern, followed by reliability and cost.