Researchers at UMass Amherst and partner institutions are exploring how artificial intelligence itself could help solve the massive energy demands that AI infrastructure creates, by designing systems that adapt to renewable energy availability and grid conditions. Rather than simply consuming more power, AI data centers could become flexible resources that support the electrical grid by shifting computational work to moments when clean energy is abundant and power is cheaper.

Why Is AI's Energy Hunger Becoming a Critical Problem?

The scale of AI infrastructure has grown dramatically. "The scale of the demand and the complexity of AI infrastructure is an order of magnitude greater than it was for previous generations of digital infrastructure," explained Mohammad Hajiesmaili, associate professor in UMass Amherst's Manning College of Information and Computer Sciences, who was recently named one of Popular Science's "Brilliant 10" for his work to reduce computing's carbon footprint. This explosive growth means that meeting energy demands while keeping electricity affordable for everyone else requires rethinking how computing infrastructure is designed and operated.

To address this challenge, UMass Amherst computer science and engineering faculty, led by Prashant Shenoy, received a $12 million, five-year National Science Foundation Expeditions grant in 2024 to develop the field of computational decarbonization. The research team, working with collaborators at the University of Chicago, UCLA, MIT, Carnegie Mellon University, and the University of Wisconsin, is applying computational approaches to automate decarbonization across the electrical grid, the built environment, transportation, and computing itself.

How Can AI Systems Be Designed to Support Clean Energy?

The key insight from this research is that AI infrastructure has built-in flexibility that traditional power consumers lack. "AI infrastructure has a flexibility that, if utilized, can be perfectly paired with the recent expansion of clean energy," noted Hajiesmaili. Unlike a factory or office building that needs consistent power, many AI workloads can tolerate carefully managed pauses, delays, or small reductions in service quality. This means data centers could reduce demand during periods when the grid is stressed and shift work to times or places where power is cleaner, cheaper, or more available.

One researcher pioneering this approach is Ramesh Sitaraman, distinguished professor of computer science at UMass Amherst, who has spent the past 15 years focusing on reducing carbon emissions from large computing systems. Rather than simply cutting overall energy use, Sitaraman developed a tool called CarbonCast that predicts the carbon intensity of a given power source over the next 96 hours, similar to a weather forecast. With this information, he aims to develop algorithms that move computational workload from locations with fewer renewable energy sources to those with more renewable capacity, reducing overall carbon emissions.

Prashant Shenoy, distinguished professor of computer science, has been experimenting with time-shifting energy usage through algorithms that account for grid conditions such as cost, supply-demand balance, and availability of green energy to optimize the timing of AI workloads. He won an ACM e-Energy 2024 Test of Time Award for this experimental research. To test these approaches in real-world conditions, Shenoy and David Irwin, professor in UMass's Riccio College of Engineering, built a micro data center powered by solar with battery support at the Massachusetts Green High Performance Computing Center (MGHPCC) site. "Our solar and battery-powered micro data center has enabled us to experimentally test a range of carbon-aware algorithms under real-world conditions," said Irwin.

Ways to Make Data Centers Part of the Energy Solution

• Workload Shifting: Moving computational tasks to times when renewable energy is abundant and grid demand is lower, allowing data centers to consume power when it's cleanest and cheapest.
• Micro Edge Infrastructure: Building small collections of servers that can be easily powered by renewable sources like solar and wind, rather than relying on massive centralized data centers that require constant baseload power.
• Grid-Aware Algorithms: Developing software that monitors real-time grid conditions and automatically adjusts computing tasks to support grid stability, turning data centers into active participants in energy management rather than passive consumers.
• Satellite-Based Computing: Exploring deployment of data centers on Low Earth Orbit satellites, which have abundant access to solar power and require no water for cooling, though this approach requires solving challenges related to the satellites' constant motion and lifecycle emissions.

The challenge with renewable energy is that it's intermittent; solar and wind power depend on weather and time of day. But AI infrastructure's flexibility could be the missing piece. "One of the biggest issues with clean energy is that it's intermittent; it depends on the availability of sun and wind, for example. But AI infrastructure can be designed to adapt to signals from the grid," explained Hajiesmaili.

Sitaraman is also exploring more unconventional approaches to sustainability. He is studying ways to redesign internet infrastructure using "micro edges," small collections of servers that can be powered by renewable sources. While such small data centers may be less reliable overall than current large, heavily fortified data centers, Sitaraman is developing clever algorithms to ensure that even if 10 to 20 percent of micro edges are not functional at a given time, the overall system would still work reliably.

Looking further ahead, Sitaraman is exploring the possibility of deploying data centers on Low Earth Orbit satellites, which orbit Earth about once an hour. This approach offers advantages including abundant access to solar power and zero water consumption for cooling. However, it poses special challenges; the satellites are not stationary, requiring smart algorithms to move data efficiently between servers in constant motion. Additionally, the satellite's launch, repair, and replacement lifecycle causes carbon emissions that must be accounted for and reduced for this approach to be truly sustainable.

In spring 2026, UMass awarded a Strategic Partnerships to Advance Research and Creative activity grant to a group of researchers led by Hajiesmaili, with the goal of establishing a large-scale research and workforce development initiative in sustainable computing and AI. "Both projects have a strong focus on AI and data centers, where the growth of computing is increasingly tied to questions of energy, infrastructure, and long-term societal impact," said Shenoy. "These projects will develop the research foundations, testbeds, and workforce needed to help future AI systems scale in ways that are reliable, efficient, and sustainable".

The research community is also organizing around these challenges. In June 2026, the Association for Computing Machinery's Special Interest Group on Energy Systems and Informatics, or ACM SIGEnergy, hosted its first annual ACM Sustainability Week in Banff, Canada, to convene researchers working to apply computational techniques to optimize energy and carbon emissions from different perspectives. The group was founded by Prashant Shenoy and is currently led by David Irwin.

Meeting AI's growing energy demands while continuing to provide electricity to the rest of society will require reimagining the design of computing and energy infrastructure to be scalable, reliable, and sustainable. The good news is that AI's inherent flexibility, combined with the expansion of renewable energy capacity, creates an opportunity to align these two trends. If managed properly, data centers could shift from being a source of pressure on the electrical grid to becoming part of the energy affordability solution.