Penn Physicists Build a Light-Matter Switch That Could Slash AI Energy to Femtojoules

Physicists at the University of Pennsylvania have demonstrated an all-light switch that uses roughly four quadrillionths of a joule per operation, opening a credible path toward photonic chips that could slash the energy bill of large AI systems. The work was published Monday in Physical Review Letters (Volume 136, Issue 14) by Zhi Wang, Bumho Kim, Bo Zhen, and Li He, with He now based at Montana State University.

The trick is a hybrid quasiparticle called an exciton-polariton, formed when photons couple so strongly to electrons inside an atomically thin semiconductor that the two stop behaving as separate things. The resulting particle inherits light's speed and long-distance reach while gaining matter's ability to interact — the missing ingredient that has long prevented purely optical systems from doing the nonlinear switching that computing requires.

"Photons can carry information quickly over long distances with minimal loss, but they barely interact with their environment," co-first author Li He explained. The Penn team's switching demonstration used energy "far below the energy needed to briefly power a tiny LED," a regime where today's electronic transistors simply cannot operate without prohibitive heat. Removing the conversion penalty between photons and electrons is one of the largest untapped efficiency wins on the table for AI hardware designers.

The most immediate application is photonic chips that ingest data directly from cameras and other optical sensors, processing it without ever round-tripping through electronic signals. The authors also flag basic quantum computing functions as plausible follow-on use cases, given that exciton-polaritons sit at the intersection of optics and condensed-matter physics. Whether the demonstration scales into the dense interconnects needed for a neural-network accelerator is the open question, but the energy numbers are striking enough that groups at Intel, IBM, and well-funded photonic-computing startups will be paying close attention.