Ultra-compact Photonic AI Chip Operates at the Speed of Light

Abstract

Researchers at University of Sydney demonstrate an ultra-compact photonic AI chip that processes information using photons, eliminating heat from electrical resistance. Nanoscale photonic structures form artificial neural networks that perform inference as light passes through them.

Key Contributions

• Inverse-designed nanophotonic structures that act as neural network layers — light propagation through material performs the computation
• Nanostructures measure tens of micrometers (width of a human hair) — orders of magnitude smaller than electronic equivalents
• Processing on picosecond timescale (trillionths of a second)
• No electrical resistance → no heat generation during computation

Results

• 90-99% classification accuracy on biomedical imaging tasks
• Tested on 10,000+ images (breast, chest, abdomen MRI scans)
• Operates at speed of light vs. speed of electrons

Limitations

• Currently demonstrated for inference only, not training
• Scaling to larger networks is ongoing work
• Limited to specific task types (classification)

Significance

• Published in Nature Communications (2026)
• Demonstrates that photonic neural networks can achieve practical accuracy
• Path toward larger-scale photonic neural networks being developed

Source: Ultra-compact photonic AI chip — University of Sydney, Nature Communications 2026