Block-funded research investigates ways to synchronize energy consumption with renewable energy availability
Dr. Zhang’s work is cosponsored by Block and the Scott Institute for Energy Innovation.
The Block Center for Technology and Society hosted a lunchtime discussion exploring the intersection of artificial intelligence, energy systems, and sustainability. The event featured Peter Zhang, who presented emerging research on how AI-driven technologies can be designed to better align with renewable energy generation and support a more resilient energy future.
Data centers in the United States consume approximately 280 megawatt-hours (MWh) of electricity each year across 5,400 facilities. That is roughly the same amount of energy used by an entire mid-sized country, such as the UK. With data centers projected to reach 7 to 10 percent of total US electricity use by 2028 and water consumption continuing to grow (Google alone used 8 billion gallons last year for cooling), the need for more resilient and sustainable energy systems is becoming increasingly urgent.
Peter Zhang’s project “Supply Chain Solutions for the Missing Renewable Energy and Nocturnal AI” explores how to schedule AI workloads to better use solar and wind power. The research is grounded in the concept of the “duck curve” - a pattern showing how solar energy availability creates large fluctuations in the electric grid. The duck curve illustrates a dip in electricity demand during the midday peak of solar production, followed by a steep ramp-up in demand in the early evenings when solar output drops and people return home to use electricity. This creates a challenge for grid operators, who must quickly ramp conventional power plants, often fossil-fueled, to meet demand while avoiding overgeneration and curtailment of solar energy midday.
Zhang’s research seeks to address this by using linear programming to optimize where and when data center workloads run. By shifting flexible AI workloads to times and locations where renewable energy is abundant but underutilized (such as at night when human demand is lower) the AI operations can reduce costs and emissions. The model accounts for factors such as energy prices, carbon emissions, migration costs, and latency and stimulates scenarios with three data centers (in New York, Chicago, and Los Angeles). Early findings show that shifting workloads could reduce both costs and emissions by taking advantage of cheaper electricity and available renewable energy. Extensions of this work include game theoretical models that analyze policy designs that could incentivize demand responsive energy use.
The potential impact of Zhang’s work is significant. If AI systems can operate on nocturnal schedules, they could help balance grid demand rather than compete with human-driven electricity use. This approach could make AI operations more climate-friendly and grid-resilient.