Artificial intelligence may soon run not just on electricity, but on light.
Researchers at the University of Sydney have developed an ultra-compact artificial intelligence chip that performs calculations using photons rather than electrons — allowing computations to occur at the speed of light while potentially using far less energy than conventional hardware.
The prototype nanophotonic chip, built at the Sydney Nano Hub, represents an early step towards a new class of AI hardware designed to address one of the technology sector’s fastest-growing challenges: the enormous energy demand of modern computing.
As artificial intelligence systems become larger and more widely deployed, the power consumption required to run them is rising rapidly. Today’s data centres already require vast amounts of electricity and water to keep servers cool, and the expansion of generative AI models has intensified concerns about the long-term sustainability of that infrastructure.
Rethinking AI hardware with light
Traditional computer chips process information by moving electrons through microscopic circuits. That process inevitably generates heat and energy loss due to electrical resistance.
The Sydney team’s chip instead performs calculations using light particles, or photons, which can travel through materials without electrical resistance. Because light does not generate heat in the same way electricity does, photonic systems could dramatically reduce the energy needed for certain types of computing tasks.
The research demonstrates how nanoscale photonic structures can be designed to perform the mathematical operations required for machine learning — effectively embedding AI models directly into the physical structure of the chip itself.
As light passes through the nanostructures, the structures automatically perform the required calculations.
That process occurs on the picosecond timescale — trillionths of a second — roughly the time it takes light to pass through the chip’s microscopic components.
Professor Xiaoke Yi from the University of Sydney’s School of Electrical and Computer Engineering, who leads the Photonics Research Group, said the research highlights the potential for photonic technologies to reshape how AI hardware is designed.
“We’ve re-imagined how photonics can be used to design new energy efficient and ultrafast computer processing chips,” Yi said.
“Artificial intelligence is increasingly constrained by the energy consumption. This research performs neural computation using light, enabling faster, more energy-efficient and ultra-compact AI accelerators.”
Tiny structures, powerful calculations
The nanostructure that forms the chip’s neural network is only tens of micrometres across — roughly the width of a human hair. Within that tiny footprint, the photonic structures mimic the behaviour of artificial neurons used in machine-learning systems.
To test the design, researchers trained the chip to classify more than 10,000 biomedical images, including MRI scans of the breast, chest and abdomen.
Both simulations and experimental testing showed classification accuracy between roughly 90% and 99%, demonstrating that the photonic neural network can perform complex pattern-recognition tasks.
The findings, published in Nature Communications, highlight how optical technologies — long used in lasers, fibre-optic communications and medical imaging — could be adapted for a new role in computing.
The race for more efficient AI infrastructure
Photonics has been explored as a computing platform for several years, but interest in the field has accelerated as the energy footprint of artificial intelligence continues to grow.
Data centres supporting large AI models now represent a rapidly expanding share of global electricity demand, prompting governments, technology companies and researchers to search for more efficient hardware architectures.
If photonic processors can eventually be scaled to commercial systems, they could allow AI workloads to run faster while consuming significantly less power than conventional chips.
PhD student Joel Sved, who helped design the prototype, said the work demonstrates how intelligence can be embedded directly into nanoscale photonic structures.
For now, the chip remains a research prototype. But with AI models placing ever greater strain on data centre infrastructure, technologies that can deliver faster processing with dramatically lower energy use are attracting growing attention.
Yi’s team is now working to scale the concept into larger photonic neural networks — a step that could move light-based AI hardware from the lab towards real-world computing systems.