NeuralStackEighty years after ENIAC helped launch the modern computing era, a team at the University of Pennsylvania is proposing a very different way to power the next wave of machines. Instead of pushing ever more electrons through ever smaller circuits, the researchers have engineered a hybrid light-matter particle that can handle core computing tasks using light alone. Their work targets one of the most stubborn bottlenecks in artificial intelligence hardware: how to do complex processing at high speed without burning through massive amounts of energy.
The project, led by physicist Bo Zhen in Penn’s School of Arts & Sciences, centers on photons, the particles of light that already dominate long-distance communications. Photons move fast, carry information efficiently, and do so with minimal loss because they are charge-neutral and have zero rest mass, as co-first author Li He explains. But that same neutrality makes them notoriously bad at interacting with their surroundings, which is a problem when you need the kind of signal-switching logic that underpins everything from simple Boolean operations to deep neural network layers.
To bridge that gap, the team built a quasiparticle known as an exciton-polariton by tightly coupling photons to electrons in an atomically thin semiconductor. Inside a nanoscale cavity, light interacts with this monolayer material to create particles that inherit light’s speed while gaining matter’s ability to interact. In practical terms, that means these exciton-polaritons can participate in the nonlinear switching operations required for computing, rather than just shuttling signals from one place to another.
That capability is especially important for artificial intelligence systems, which lean heavily on nonlinear “activation” steps that turn raw weighted sums into decisions. Many experimental photonic AI accelerators already use light for certain matrix-heavy calculations, but typically have to convert light back into electronic signals for those nonlinear operations. Each conversion adds latency and wastes energy, eroding the benefits of optical hardware just as AI workloads are exploding in size and complexity.
In their latest work, the Penn researchers demonstrated that exciton-polaritons can enable all-light switching while consuming only about 4 quadrillionths of a joule of energy per operation. That figure is far below the energy needed to briefly power a tiny LED, underscoring how little power is required to toggle these hybrid particles. The team’s devices rely on strongly nonlinear nanocavities and gate-tunable monolayer semiconductors, a combination detailed in a paper published in Physical Review Letters by Zhi Wang, Bumho Kim, Zhen, and He.
Why this matters
If this approach can be scaled and integrated into real-world hardware, it could reshape how AI chips are built and powered. Photonic processors capable of performing both linear and nonlinear steps entirely in the optical domain would no longer need to constantly turn light into electricity and back again, cutting energy losses while preserving speed. That could translate into data centers that run large AI models with much smaller power budgets, and edge devices that process sensor data more efficiently instead of offloading everything to cloud servers.
One of the most intriguing possibilities is photonic chips that process information directly from cameras or other optical inputs without repeated conversions between light and electrons. Instead of capturing light, digitizing it, and then pushing it through conventional silicon, future systems could keep information in the optical realm as it moves from sensing through inference. The same platform could also support basic quantum computing functions, since exciton-polaritons inhabit a regime where quantum behavior and device-level control intersect.
The work also marks a symbolic return to Penn’s roots in computing, from ENIAC’s streams of electrons to nanoscale cavities that sculpt the behavior of light and matter. Zhen, now the Jin K. Lee Presidential Associate Professor in Penn’s Department of Physics and Astronomy, collaborated with He, a former postdoctoral researcher in the lab who is now an assistant professor at Montana State University, along with Wang and Kim from Penn Arts & Sciences. Their research, backed by support from the US Office of Naval Research and the Sloan Foundation, positions exciton-polaritons as a promising building block for the next generation of AI hardware.
Much work remains to turn an elegant laboratory demonstration into a robust platform that can survive the harsh realities of commercial manufacturing and deployment. Engineers will need to scale the nanocavity designs, integrate them with existing chip fabrication processes, and prove that these devices can operate reliably across the wide range of temperatures and workloads modern AI systems demand. But as the industry grapples with the mounting power costs of ever-larger models, Penn’s light-matter approach offers a concrete path toward AI hardware that is not just faster, but fundamentally more efficient.