New brain-like chip could slash AI energy use by 70%

Artificial intelligence is facing an energy crisis. As models grow larger, the electricity required to power data centers is reaching unsustainable levels. Current computing architectures suffer from the ‘Von Neumann bottleneck,’ where the constant movement of data between the processor and memory consumes the majority of a system’s power. To solve this, researchers at the University of Cambridge and the University of Southern California (USC) have developed brain-inspired ‘memristors.’ These nanoelectronic devices integrate memory and processing into a single unit, mimicking the human brain’s efficiency. This breakthrough could reduce AI energy consumption by 70% and allow electronics to function in extreme environments previously thought impossible.

Modern computers waste significant energy moving data back and forth across a motherboard. This physical separation between the ‘brain’ that thinks and the ‘cabinet’ that stores information is the primary cause of heat and power loss in AI systems. Researchers at the University of Cambridge, led by Dr. Babak Bakhit, have engineered a memristor that eliminates this gap. By integrating processing and storage functions in a single nanoelectronic unit, the device replicates the synaptic behavior of human neurons. This ‘neuromorphic’ approach allows the hardware to handle complex AI tasks locally without the high energy overhead of traditional architectures. The Cambridge team’s three-year research project resulted in a device that could reduce energy use by up to 70%. This shift is critical as global energy grids struggle to support the massive scaling of AI infrastructure. By making hardware inherently more efficient, the industry can continue to advance without the proportional increase in carbon emissions or power demand that currently plagues large-scale model training and deployment.

Sophisticated material stack

The technical success of the Cambridge memristor relies on a sophisticated material stack. Traditional oxide-based memristors often rely on filamentary switching, which can be unstable and difficult to control at scale. To overcome this, the Cambridge team utilized a modified hafnium oxide thin film infused with strontium and titanium. This composition creates p-n junctions at the layer interfaces, allowing for precise control over electrical resistance. This design ensures more stable and uniform switching compared to legacy devices. Furthermore, the switching current required for this new chip is roughly a million times lower than that of conventional oxide-based memristors. This massive reduction in current is the primary driver behind the 70% energy savings. By utilizing hafnium oxide—a material already common in semiconductor manufacturing—the researchers have ensured that their discovery is grounded in materials that the industry understands. The inclusion of strontium and titanium provides the necessary stability for the device to maintain its state, effectively acting as non-volatile memory that does not require constant power to retain information.

Strategic autonomy

The development of energy-efficient, high-performance hardware is a matter of strategic autonomy for Europe and the UK. As AI becomes central to national security and economic competitiveness, reducing reliance on energy-intensive legacy architectures is a priority.