Image
Could the Human Brain Hold the Key to Energy-Efficient AI?
Artificial intelligence is fueling breakthroughs in medicine, research, and education. It is also consuming staggering amounts of electricity.
As companies race to build ever-larger AI models, data centers are drawing unprecedented levels of power, straining electrical grids and raising concerns about environmental impacts and consumer energy costs.
“Every AI model trained or inferred at scale pushes the grid closer to its breaking point,” warned Jason Williamson, assistant professor of data science at the University of Virginia, in a May 6 presentation at the DSC Next conference in Amsterdam. “This is not a theoretical risk. It is an active infrastructure emergency.”
Some researchers believe the solution may lie in an unlikely place: the human brain.
Neuromorphic AI is a computing approach that mimics the way the human brain works. This technology could significantly decrease AI’s energy footprint and alleviate some of the infrastructure challenges, according to Williamson, who is the CEO of the AI product lab Mythworx outside of his University role.
The Price of Progress
In Europe, Williamson said the strain that AI is putting on the infrastructure is creating a particularly impossible equation. He pointed to four main constraints: limited available land, existing grid challenges, mandatory sustainability reporting, and drought-stressed regions unable to provide water for data center cooling.
While the U.S. has more available resources for data center construction and expansion, much of the new infrastructure costs are being passed down to consumers. In 2025 alone, the price of residential electricity jumped over 7%, and in some states, as much as 20%, according to a Consumer Reports article. Inflation, fuel costs, grid upgrades, and other factors are contributing to these cost increases as well, with the AI boom adding pressure to infrastructure that is already strained in many areas.
“The energy and environmental costs are rapidly growing for AI, but neuromorphic approaches provide a way to significantly improve matters in the future,” said UVA data science professor David Danks.
Researchers at UVA are already exploring how to get there. In 2023, the University joined the BrainChip University AI Accelerator Program, giving students and researchers in the Charles L. Brown Department of Electrical and Computer Engineering (ECE) access to neuromorphic computing technology and tools.
The Edge of Discovery
Faiyaz Elahi Mullick, an ECE doctoral candidate, is experimenting with edge computing systems, which conserve massive amounts of energy by processing data on small devices like smartphones instead of sending data to large cloud servers, which are typically housed in data centers.
His research centers around computing systems that mimic the mammalian brain’s ability to remember.
As it stands, many AI systems struggle with a problem known as catastrophic forgetting. Sometimes, when you train an AI system to recognize dogs, for example, and then teach it to recognize a plant, it will overwrite its memory and forget what it learned about dogs. Humans and other mammals, on the other hand, can learn new things throughout their lives without automatically erasing old memories.
Researchers are exploring whether a biological mechanism known as adaptive synaptogenesis, which helps the brain form new connections while conserving energy, could help AI systems do the same.
“The brain is enormously efficient,” said ECE professor Avik Ghosh. “Faiyaz is building a mathematical model and already seeing some gains over conventional AI using adaptive synaptogenesis and trying to see how far this can be pushed.”
Until recently, the ability to learn new information without overwriting old knowledge was largely limited to biological organisms. New research in this field could change that, steering AI in a new direction.
The Hurdles Ahead
However exciting the field may be, many researchers agree that neuromorphic models are not likely to completely replace LLMs anytime soon. Mullick says that some of the tasks that LLMs perform are not inherently easy for neuromorphic systems.
He said that software is not the only challenge. “Contemporary AI hardware is another massive limitation. There is not only a need for a new software paradigm, but also hardware paradigms.”
While neuromorphic systems may not replace today's LLMs anytime soon, researchers believe they point toward a future where AI becomes less dependent on sprawling data centers and more efficient, like the biological systems that inspired them.
