The human brain runs on just 20 watts. Replicating its capabilities with today’s AI hardware could demand thousands of NVIDIA H100 GPUs, each consuming 700 watts — pushing power use into the megawatt range. That contrast is the starting point for Cambridge researchers who say their new chip material could slash AI energy consumption by more than 70%.

The chip switches at just 10⁻¹¹ amps — a million times lower than conventional designs.  
Credit: University of Cambridge


The breakthrough comes from a nanoscale device called a memristor. Unlike conventional computer chips, which separate memory and processing units and waste energy shuttling data between them, a memristor can store and process information in the same place. That’s how synapses in the brain work, and it’s why neuromorphic computing — hardware designed to mimic the brain’s architecture — has become a serious candidate for tackling AI’s spiraling energy demands.

Most memristors today rely on forming and rupturing microscopic conductive filaments inside oxide materials. The problem is that these filaments behave unpredictably, leading to inconsistent performance and higher power requirements. The Cambridge team took a different route. By engineering hafnium oxide with strontium and titanium, and using a two-stage growth process, they created tiny p-n junctions — the same kind of electronic gates found in conventional semiconductors. Instead of breaking and reforming filaments, the device changes resistance by adjusting the height of an energy barrier at the junction. The result is smoother, more controllable switching.

Dr. Babak Bakhit, who led the study, explained why this matters: “Filamentary devices suffer from random behavior. But because our devices switch at the interface, they show outstanding uniformity from cycle to cycle and from device to device.” That uniformity is critical for scaling neuromorphic hardware beyond the lab.

The energy savings are striking. Switching currents measured as low as 10⁻¹¹ amps — about a million times lower than some conventional oxide-based memristors. The switching energy falls into the femtojoule-to-picojoule range, rivaling or surpassing the most efficient neuromorphic devices demonstrated so far. In practical terms, this could mean AI systems that consume only a fraction of the electricity required today.

Equally important is the chip’s analog behavior. Traditional digital systems operate in binary states: on or off. Biological synapses don’t. Their connection strengths shift gradually, allowing for nuanced learning. The Cambridge memristors demonstrated hundreds of distinct, stable conductance levels, enabling brain-like analog computing. They also reproduced spike-timing-dependent plasticity — a biological learning mechanism where the strength of neural connections changes based on the timing of signals. In other words, the hardware itself begins to behave less like static memory and more like adaptive brain tissue.

Still, challenges remain. The current fabrication process requires temperatures around 700 °C, far higher than standard semiconductor tolerances. Making the technology compatible with existing chip manufacturing is the next hurdle. Bakhit acknowledged the difficulty: “This is currently the main challenge in our device fabrication process. But we’re now working on ways to bring the temperature down to make it more compatible with standard industry processes.”

The research, published in Science Advances, is still at the experimental stage. But the implications are clear. If manufacturing barriers can be overcome, neuromorphic chips based on this design could reshape the energy profile of AI. Instead of megawatts, we could be talking about systems that run closer to the brain’s 20-watt efficiency. That’s not just a technical improvement — it’s a shift in what kinds of AI applications become feasible, from mobile devices to large-scale data centers.

The story here isn’t about mimicking the brain perfectly. It’s about borrowing its most efficient tricks. Cambridge’s memristor doesn’t think, but it learns in ways that silicon hasn’t before. And if it can be brought into mainstream production, the future of AI hardware may look less like racks of GPUs and more like a network of chips that whisper, rather than roar, their computations.