Eighty years after Penn researchers helped invent the first general-purpose electronic computer, a new team at the same university has built something that may help replace it. A study published in Physical Review Letters this month describes a way to perform computing operations using light alone — and the energy cost is so low it borders on the absurd.
The team, led by physicist Bo Zhen at the University of Pennsylvania, used a hybrid particle called an exciton-polariton to demonstrate all-optical signal switching, the basic operation that lets a computer make decisions. Their device pulled off the trick using only about 4 femtojoules — four quadrillionths of a joule — far less energy than it takes to briefly power an LED.
Why Electrons Hit a Wall
Every modern computer, from a phone to a data center GPU, moves information by pushing electrons through silicon. Electrons are great at switching, which is why they've been the workhorse of computing since ENIAC's vacuum tubes in 1945. But they carry a charge. That charge meets resistance. Resistance generates heat. And heat — at the densities required by modern AI chips — has quietly become the single largest constraint on how fast computing can grow.
Photons, the particles that make up light, don't have any of those problems. They're chargeless, massless, and they barely interact with anything around them. That's why fiber-optic cables can carry your internet across oceans with almost no loss. But it's also why photons are bad at the one thing computing demands: switching. To get a signal to flip on or off, particles need to interact strongly with their surroundings, and photons just don't.
The Hybrid Trick
Zhen's team got around this by building a particle that's part light, part matter. By coupling photons to electrons trapped in an atomically thin semiconductor, they created exciton-polaritons — quasi-particles that move like light but interact like matter.
"Photons can carry information quickly over long distances with minimal loss, dominating communications technology," said Li He, co-first author of the paper. "But that neutrality means they barely interact with their environment, making them bad at the sort of signal-switching logic that computers depend on."
The polariton solves that. It keeps the speed and efficiency of light while gaining just enough "stickiness" to perform the switching operations a computer needs.
The Energy Number That Matters
Today's photonic AI chips already exist — but they cheat. When light needs to perform a nonlinear "decision" step, the chip typically converts the signal back into electronics, does the math, and converts it back to light. Every conversion burns energy and adds delay, eating into the speed advantage photons should offer.
Penn's polariton device skips that step entirely. The team performed switching with light only — no electronic detour — and recorded an energy cost roughly a thousand times below the efficiency frontier of existing optical AI hardware.
What It Could Unlock
If the technology scales, the implications go well past energy savings on chatbot queries. Photonic chips with native light-based logic could ingest data directly from cameras, sensors, and fiber networks without ever translating the signal into electronics. That matters for autonomous vehicles, medical imaging, scientific instruments, and any AI workload bottlenecked by getting data in and out of the processor.
Scaling, of course, is the catch. The team's demonstration works in the lab on tiny structures. Making polariton devices stable, manufacturable, and integrable with existing chip ecosystems will take years.
But the physics now works. And as AI's appetite for electricity continues to outpace every projection, a generation of chips that compute with photons — and barely warms up doing it — looks less like science fiction and more like a roadmap.
