AI is hungry for power, and one answer is to stop doing all the math on transistors. Researchers at the University of Tokyo, working with NTT researchers, have shown a way to train and run a neural network on light, using far less computation than conventional methods. The work was announced on November 5, 2026 and published in Nature Computational Science.
The problem: AI that costs too much to run
According to the University of Tokyo press release, training AI models can consume vast amounts of energy, pushing researchers to look for efficiency rather than simply building ever-larger systems. One option is physical computing, which uses the natural behavior of a physical system to process information. Ordinary chips push electrical signals through huge numbers of transistors, which uses energy and produces heat. Light can carry information very quickly and can carry several signals at once, which makes optical computing an attractive platform.
What the team changed
The catch is that standard AI training methods are awkward on physical hardware. They require measuring and recording the system as it evolves, then working backward to decide what to change. The team, led by Associate Professor Kohei Nakajima, combined a training method called SS-DFA (state-skipping training) with a simpler architecture called the single-map framework. Per the university, the method avoids access to most intermediate states and uses the final error plus information about the system to guide learning, which reduces measurement, data storage and computation. The simplified architecture has many units following general rules instead of a complex web of internal connections.
The optical demonstration
In the experiment, information is converted into an optical signal, processed as it travels through a fiber loop, and converted back to an electrical signal for analysis. The system uses a 200-meter optical fiber and a small number of electronic components. The university reports that the simplified design matched, and in some tests outperformed, more complex neural network architectures while needing far less computation. The paper is “Scalable physical deep learning using optical dynamics with state-skipping training and single-map architecture” by Yongbo Zhang, Mitsumasa Nakajima, Katsuma Inoue, Toshikazu Hashimoto, Yasuo Kuniyoshi and Kohei Nakajima (Nature Computational Science).
Why it could matter, and the limits
If the approach scales, the team says it could lower the energy used for computing and cooling in future AI data centers and reduce water use thanks to lower heat output. If miniaturized onto a chip, it could process data inside phones, cameras or vehicles with faster responses and lower energy use. Nakajima is candid about the gaps: the framework has not yet been tested on today’s largest language and multimodal models, and more work is needed on miniaturization, stability and large-scale manufacturing. It has been integrated into several relatively baseline modern neural network architectures, where it improved performance while reducing computation.
Practical takeaways
- Efficiency research is shifting from bigger models to different hardware and learning rules.
- The result is a lab demonstration, not a product; scaling to large models is the open question.
- Watch for follow-up work on chip-integrated optical AI.
Bottom line
Rethinking how AI learns, not just what it runs on, may be the key to making optical hardware practical. This is an early but promising step toward lower-energy AI.