The human brain performs extraordinarily complex tasks while consuming remarkably little energy. Today's artificial intelligence systems are far less efficient. Training and operating increasingly powerful models require vast data centers packed with energy-hungry hardware, raising concerns about whether the industry's rapid growth can be sustained.
That contrast has inspired researchers at the University of Cambridge to develop a new electronic device modeled on the way the brain handles information. Their work belongs to a field known as neuromorphic, or brain-inspired, computing. Systems built on this approach could theoretically consume over 70% less energy than conventional computers, although the Cambridge device has not yet been tested as part of a complete AI system.
The potential savings come from changing the relationship between memory and computation. In an ordinary computer, information is stored in one location and processed in another. Data must therefore travel back and forth whenever a task is performed. Although each transfer requires only a small amount of electricity, the total becomes enormous when an AI model is processing billions of pieces of information.
The brain follows a different architecture. Its networks of neurons and synapses can store information and process signals simultaneously, largely avoiding the constant movement of data. The Cambridge researchers hope to reproduce this advantage with a microscopic device called a memristor.
Unlike an ordinary memory component, a memristor can retain a record of the electrical current that has passed through it, even after the power is removed. Its resistance can also be adjusted gradually rather than being limited to simple "on" and "off" states. In this respect, it resembles a biological connection that becomes stronger or weaker as the brain learns.
Memristors have been studied for years, but many existing designs remain unreliable. They depend on tiny conductive paths that repeatedly form and break inside a material. Because this process occurs somewhat randomly, the same device may respond differently from one cycle to the next. Such variability makes it difficult to combine large numbers of memristors into a dependable computing system.
The Cambridge team adopted a more controlled design. Instead of relying on unpredictable internal paths, their device switches states through changes at an electronic boundary within the material. This allows it to respond more smoothly and consistently across repeated tests, addressing one of the main obstacles to the wider use of memristors.
Its energy requirements were also strikingly low. The smallest current recorded by the researchers was roughly a million times lower than that used by some earlier memristors. Moreover, the device could maintain hundreds of distinct states, rather than merely the two states used by conventional digital computers.
This capacity matters because learning in the brain is not based on connections that are simply active or inactive. Connections adjust gradually in response to experience, allowing information to be strengthened, weakened or reorganized. During experiments, the new memristor responded to repeated electrical signals in a similarly gradual manner. The result suggests that future machines might be able to learn while consuming far less energy.
Nevertheless, substantial barriers remain. Manufacturing the device currently requires temperatures of around 700 degrees Celsius, well above the limits of standard chipmaking processes. The researchers are now attempting to lower that temperature without sacrificing the device's performance or reliability.
They must also demonstrate that thousands or millions of these components can operate together effectively. A single device may perform well in a laboratory, but an AI system requires a large and highly coordinated network. Until such a system is built and tested, the possible 70% reduction should be understood as the broader promise of brain-inspired computing, not as an energy saving already achieved by this particular invention.
Even with those limitations, the research points toward an important shift in how engineers approach AI's energy problem. Instead of focusing exclusively on making conventional processors faster, they may need to reconsider the basic architecture of computing. By bringing memory and processing closer together, future hardware could become not only more powerful, but also considerably more efficient.