University of Sydney Researchers Build Ultra-Compact Photonic ‘Light Speed’ AI Chip

A team of researchers at the University of Sydney has built a nano photonic chip prototype capable of performing artificial intelligence calculations using light, rather than electricity.

The prototype, developed entirely in-house at the Sydney Nano Hub, uses light particles known as photons to carry out computations at speeds measured in trillionths of a second.

The work, published in Nature Communications, comes as global demand for AI infrastructure continues to accelerate and operators of data centres grapple with the energy and cooling costs of powering conventional silicon-based hardware.

How The Chip Works

Traditional computer chips process information by pushing electrically charged particles, or electrons, through wires. That movement generates heat, which in turn requires cooling systems and draws significant amounts of energy.

The nano photonic chip takes a different approach. Instead of electrons, light passes through nanoscale structures embedded in the chip. Because light can travel through materials without electrical resistance, it does not produce heat in the way conventional electronics do.

As light moves through the nanostructures, the structures themselves perform the calculations automatically. There is no separate processing step required. The nanostructure on the chip occupies an area of tens of micrometres, roughly the width of a human hair.

Those nanostructures are arranged to form a neural network: a system of artificial neurons designed to mimic aspects of how the human brain recognises patterns and completes calculations.

Professor Xiaoke Yi, from the School of Electrical and Computer Engineering and Director of the Photonics Research Group at the University of Sydney, explained that the technology represents a rethinking of how photonics can be applied to computing.

“We’ve re-imagined how the photonics can be used to design new energy efficient and ultrafast computer processing chips,” Professor Yi stated.

“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.”

Testing Against Biomedical Images

To validate the prototype, the research team trained the nanophotonic chip to classify more than 10,000 biomedical images, including breast, chest and abdomen MRI scans.

In both simulations and physical experiments, the nanophotonic neural network achieved classification accuracy of approximately 90 to 99 per cent. The researchers describe the results as demonstrating how AI models can be designed into nanoscale photonic structures that manipulate light to perform the mathematical operations required for machine learning.

The calculations take place on the picosecond timescale, meaning each operation is completed in trillionths of a second. That is essentially the time it takes for light to pass through the nanostructure on the chip.

Energy Consumption In Focus

The research arrives at a time when the energy footprint of AI is under increasing scrutiny. Current data centres rely on large amounts of water and electricity to power and cool the hardware needed to train and run AI models.

The researchers argue that photonic computing offers a potential pathway to AI infrastructure that can scale without proportional increases in power consumption. Because the chip uses light rather than electricity, the team believes it could help lower the overall energy demands of future computing systems.

The distinction matters as tech companies and governments around the world invest heavily in expanding AI capacity. Every new data centre built to support large language models and other AI workloads adds to an already growing demand on power grids and water supplies.

Professor Yi’s team contends the prototype is a step toward addressing that challenge at the hardware level, rather than relying solely on software-based efficiency improvements.

A Decade Of Photonics Research

Photonics, the study of controlling light particles, already underpins a range of everyday technologies including lasers, fibre-optic networks and medical imaging systems.

However, the application of photonics to general-purpose computer processing is a newer area of study, one that has gained urgency as AI workloads have grown.

The Photonics Research Group at the University of Sydney has spent more than a decade investigating how to push the boundaries of what photonic technology can achieve. Previous work by the group has included using photonics to address challenges in wireless communications and in advanced sensing technology capable of detecting and measuring chemical or biological traces in the environment.

PhD student Joel Sved, who played a key role in the design and implementation of the prototype, described the work as a demonstration of how intelligence can be embedded directly into nanoscale photonic structures.

The distinction between this approach and conventional chip design is significant. Rather than building a processor that runs AI software, the team has built a chip where the physical structure of the hardware itself carries out the neural network’s operations as light passes through it.

What Comes Next

Following the testing of the nanophotonic chip prototype, Professor Yi’s team is now working on advancing the technology toward larger-scale photonic neural networks.

Scaling up the approach will involve designing more complex nanostructures capable of handling bigger and more varied datasets, moving beyond the biomedical imaging classification tasks used in the initial validation.

The team’s ambition is to develop photonic hardware that could eventually be deployed in real-world AI systems, offering an alternative to the electronic chips that currently dominate the market.

Whether photonic computing can move from laboratory prototype to commercial product remains an open question. But the University of Sydney researchers believe their work demonstrates the fundamental viability of the approach and provides a foundation for future development.

Last Updated on March 10, 2026 by Nick Ross