Light Instead of Silicon: Nanophotonic Chip Changes Computing Physics

On March 10, 2026, a study was published in Nature Communications that could rewrite the current architecture of AI servers. A team from the University of Sydney demonstrated an ultra-compact prototype of a nanophotonic accelerator that performs neural network computations using light.

Modern GPUs have hit thermal and energy limits—electrons in copper conductors are too slow and "hot." The transition to photons (optical computing) allows data to be processed on a picosecond scale at the speed of light. The prototype successfully ran the classification of thousands of medical images. Nanophotonics is the Holy Grail of hardware AI. If the technology can be scaled to commercial foundry lines, data center power consumption will drop by orders of magnitude, and the interconnect bottleneck will disappear forever.

Source: University of Sydney / Nature Communications