Cambridge Researchers Build a Brain-Like Chip That Could Cut AI Energy by 70%

A team at the University of Cambridge has engineered a nanoelectronic chip that mimics how biological neurons work, cutting the energy needed to run AI computations by up to 70%. The work was published in Science Advances in March 2026.

The material at the core of the device is modified hafnium oxide, doped with strontium and titanium to create a p-n junction at the interface between layers. Where a conventional chip shuttles data back and forth between separate processor and memory units, this device handles both in the same place, eliminating one of the main sources of wasted energy in current AI hardware.

The device is technically a memristor, a component that retains its resistance level after power is removed. But it behaves differently from earlier designs, which stored information by forming unstable conductive filaments. Here, resistance is adjusted by changing the height of the energy barrier at the material interface, giving the chip hundreds of stable, distinct conductance states.

The numbers from the lab are striking. Switching currents run roughly one million times lower than those in conventional oxide memristors, and the device sustains tens of thousands of reliable switching cycles. Lead researcher Dr. Babak Bakhit says the design mirrors spike-timing dependent plasticity, the process by which biological brains strengthen or weaken synaptic connections based on the timing of neuron signals. That is how learning works in biological systems, and it is what the chip is designed to replicate in silicon.

There is a manufacturing hurdle to clear. The deposition process currently requires temperatures around 700 degrees Celsius, above the tolerances of standard semiconductor fabrication lines. Dr. Bakhit's team is working to bring that threshold down before the design can transfer to commercial production.

An earlier piece here covered a different approach to the same problem: neuro-symbolic AI architectures that cut power consumption by a factor of 100 through software design rather than hardware. The Cambridge work operates at the chip level, and the two approaches are not mutually exclusive. If both continue to develop, the gap between current AI energy costs and what might eventually be possible is considerably wider than most people assume.