Ask any AI researcher what keeps them up at night and chances are the answer involves power strips. Not safety switches or extension cords, but the sheer amount of electricity swallowing data centers around the world. The International Energy Agency estimates US AI systems and data centers consumed about 415 terawatt hours in 2024. That is more than 10 percent of everything the nation produced that year. By 2030 the agency expects that number to double.
Two separate research teams may have found ways to reverse that trajectory. One approach promises to cut energy use by 100 times. Another claims a 70 percent reduction. Both have appeared in peer reviewed publications within the last two months.
The Problem With Modern AI Energy Hunger
Current AI systems rely on traditional computer chips that constantly move data between memory and processing units. That back-and-forth transfer requires enormous amounts of electricity. The more powerful the model, the worse the problem becomes.
For robots that interact with the physical world, this creates specific challenges. These systems, known as visual-language-action or VLA models, take camera input and language instructions and translate them into real world movements. They need to identify objects, interpret scenes, and plan actions, all while running on hardware that was never designed for the task.
A robot asked to stack blocks into a tower illustrates the issue perfectly. A conventional VLA system must scan the setting, identify each block, determine its shape, and work out how to place pieces on top of one another. Shadows may cause it to misread a block shape. It might place a piece incorrectly, toppling the whole stack. The system learns through trial and error, and trial and error costs energy.
A Different Way to Think About Thinking
Researchers at Tufts University took a step back and asked what would happen if AI systems could break problems into steps and categories, the way humans do. Their approach, called neuro-symbolic AI, combines standard neural networks with symbolic reasoning.
The concept works like this. Instead of relying purely on patterns from data, the system applies rules and abstract concepts. It understands that blocks have shapes and that stacking a wide base with a narrow top creates an unstable structure. It does not need to try and fail thousands of times to learn this. It already knows.
Like large language models, neuro-symbolic VLAs still learn from large training sets. But they can apply rules that limit the amount of trial and error during learning and get to a solution much faster. Not only does the system complete the task faster, the time spent training it drops significantly.
Matthias Scheutz, Karol Family Applied Technology Professor who led the research, presented findings at the International Conference of Robotics and Automation in Vienna. The team tested their system using the Tower of Hanoi puzzle, a classic problem that requires careful planning. The neuro-symbolic VLA achieved a 95 percent success rate. Standard VLAs managed just 34 percent. For more complex versions of the puzzle that the robot had never seen during training, the neuro-symbolic system succeeded where the standard approach failed completely.
Looking at the Numbers
The Tufts work is not the only breakthrough pointing toward more efficient AI. Researchers at the University of Cambridge published separate findings in Science Advances describing a new type of nanoelectronic device that mimics how the brain processes information.
The Cambridge team, led by Dr Babak Bakhit from the Department of Materials Science and Metallurgy, developed a modified version of hafnium oxide that functions as a highly stable, low-energy memristor. Unlike conventional memory chips that keep processing and memory in separate locations, this device combines both in one place, similar to the brain.
The design avoids the unpredictability that plagues older memristor designs. By adding strontium and titanium and using a two-step growth process, the researchers created small electronic gates at the interfaces between layers. Tests showed the devices operate at switching currents roughly a million times lower than some conventional oxide-based memristors. They remained stable through tens of thousands of switching cycles and retained their programmed states for about a day. The team estimates the approach could cut AI energy use by as much as 70 percent.
Why This Matters for the Industry
The numbers add up quickly. According to the International Energy Agency, global data center electricity consumption will grow to 945 terawatt hours per year by 2030, up from 415 terawatt hours in 2024. That is roughly equivalent to Japan's current total power consumption. Server electricity consumption for AI workloads is projected to grow by 30 percent annually.
Even modest improvements at scale translate to massive savings. The Tufts approach claims 100 times less energy consumption for certain robotics tasks. The Cambridge work suggests 70 percent reductions for hardware-level operations. Combined with existing efforts to improve data center cooling and power delivery, the industry has multiple paths toward a more sustainable future.
Scheutz noted that a single AI generated summary at the top of a Google search result consumes up to 100 times more energy than producing the list of website links beneath it. That single data point illustrates the scale of the problem and the potential of solutions currently in development.
What Comes Next
Both research teams are moving beyond laboratory proof of concepts. The Tufts work is headed toward wider testing in robotic systems. The Cambridge team is exploring commercial manufacturing pathways for their memristor design. Neither approach will solve the energy problem overnight. But both offer pathways toward systems that do more with far less.
The challenge of sustainable AI has finally started attracting serious resources. When 10 percent of a nation's electricity output goes to a single industry, the math tends to concentrate minds quickly.