AI Energy Crisis Deepens: New Breakthrough Cuts Power Use By 100 Times

Artificial intelligence is consuming a large amount of electricity, as shown by the International Energy Agency's latest report. In 2024 alone, AI systems and data centres in the United States used around 415 terawatt hours of power, which is more than 10% of the country's total electricity. This level of energy use is expected to double by 2030, leading to major concerns about long-term sustainability.

To address this challenge, a team from a School of Engineering, led by Matthias Scheutz, has created a new proof-of-concept system that is much more efficient.

Their results, which will be presented at the International Conference on Robotics and Automation and published in its proceedings, show that energy use could be reduced by up to 100 times, while also increasing accuracy.

Hybrid Method

The team's method is called neuro-symbolic AI. It blends neural networks, which learn from large sets of data, with symbolic reasoning, which uses rules and logical steps. This approach is similar to how humans solve problems, by breaking things down into smaller parts rather than just recognizing patterns.

Robotics Focus

Unlike widely known systems like ChatGPT, this research is centred on robotics. These systems, known as visual-language-action (VLA) models, combine vision, language, and physical movement. They take in camera input and instructions, then translate them into actions such as moving robotic arms or wheels.

Traditional VLA models rely heavily on trial-and-error learning. For example, when stacking blocks, a robot must understand the scene and try different placements.

This often leads to errors as shadows can change the appearance of shapes, or objects may not be placed correctly, causing the system to fail. This pattern is also seen in language models, which can produce incorrect or unrealistic outputs.

Also read: AI Could Beat All Human Experts Within a Year, Researchers Warn After New Test Findings

Better Reasoning

Symbolic reasoning improves this process by adding rules about concepts like shape and balance. According to Scheutz, this reduces unnecessary trial and error, allowing systems to find solutions faster and with greater reliability. It also significantly reduces the time required for training.

Test Results

The system was tested using the Tower of Hanoi puzzle. The hybrid model achieved a 95% success rate, compared to only 34% for standard systems.

Even when faced with a more complex version of the puzzle it had not seen before, it succeeded 78% of the time, while conventional models failed entirely. Training took just 34 minutes, whereas traditional methods required more than a day.

Energy Savings

Energy efficiency improved greatly. The new model used only 1% of the energy for training and 5% during operation compared to standard systems. Scheutz explained how even basic AI tasks can be inefficient, noting that even a simple online search can consume a large amount of energy.

As AI continues to grow, the demand for computing power is putting more pressure on infrastructure, with some data centres using as much electricity as small cities.

Researchers argue that current AI methods may not be sustainable. Neuro-symbolic AI, which combines learning with structured reasoning, could provide a more reliable and energy-efficient future.

Also read: AI-Generated X-Rays Are So Real, Even Radiologists Can't Tell the Difference

MENAFN06042026007385015968ID1110945694