Scientists at the University of Pennsylvania have made an exciting discovery that could change the way computers work. They created tiny particles called exciton-polaritons, which are a mix of light and matter. These particles could one day replace the electrical signals that currently power computer chips.
The most important advantage is energy efficiency. The polaritons can switch signals using only about four femtojoules of energy - an incredibly small amount. This is roughly one billion times less energy than the switches used in modern computer chips.
The material the Penn team used is called molybdenum diselenide, or MoSe2. It is a semiconductor that is only one atom thick. When light enters a tiny cavity containing this material, it combines with the material to form polaritons.
AI data centers currently consume enormous amounts of electricity around the world. If future chips could use light to process information instead of electricity, AI could become far more energy-efficient. The research was published in Physical Review Letters in April 2026.
A team of physicists at the University of Pennsylvania has demonstrated all-optical signal switching using exciton-polaritons - quantum quasi-particles that form when photons couple strongly with electron-hole pairs (excitons) trapped in an atomically thin semiconductor. The result, published in Physical Review Letters on April 8, 2026, is a milestone toward photonic computing that could one day replace silicon transistors in energy-intensive AI hardware.
The Penn team used a nanoscale optical cavity filled with a monolayer of molybdenum diselenide (MoSe2), a two-dimensional semiconductor with exceptionally strong light-matter coupling properties. A laser pulse entering the cavity hybridizes with the material's excitons to form polaritons, which propagate and interact within the cavity. A second, weaker control pulse can shift the polariton population between two resonance states, performing a logic switching operation with an energy cost of approximately four femtojoules (4 x 10 to the negative 15 joules).
That energy figure is roughly one billion times more efficient than a silicon transistor switching at equivalent speed, and well below the threshold of any previous all-optical transistor demonstrated in other material platforms. The key advantage comes from the polariton's dual nature: because it is partially light, it travels fast within the cavity; because it is partially matter, it interacts strongly enough with control pulses to function as a logic gate.
The technology faces significant engineering hurdles before commercialization. Device-quality MoSe2 monolayers are difficult to produce at wafer scale, and the polariton lifetime in the current device is limited to a few picoseconds. However, the Penn team and independent researchers at IBM and imec see a credible path to photonic AI accelerator chips that could slash the energy footprint of data centers - which, globally, are projected to consume more than 1,000 terawatt-hours of electricity annually by 2028.
The University of Pennsylvania polariton paper in Physical Review Letters should be situated within a macro-structural imperative: global data center electricity consumption, currently ~415 TWh per year and growing at roughly 30 percent annually, is propelled predominantly by the AI inference workloads of large language models and diffusion networks. These workloads are throughput-bound, bandwidth-bound, and increasingly power-bound operations that silicon CMOS, now bumping against the physical ceiling of Dennard scaling, can no longer accelerate without proportional energy cost increases. Photonics has long been proposed as the escape vector, but has historically been stymied by the absence of an efficient room-temperature optical nonlinearity - the physical property required to build a logic gate from light alone.
The Penn group's core contribution is a room-temperature optical nonlinearity of sufficient magnitude for practical switching. By placing a MoSe2 monolayer - chosen for its large exciton binding energy (~300 meV, conferring room-temperature stability) and oscillator strength - inside a photonic nanocavity with a Q-factor exceeding 2,000, the team achieved a vacuum Rabi splitting that places the coupled system firmly in the strong-coupling regime. The polariton branches thus formed exhibit Coulomb-exchange-mediated nonlinearity: a four-femtojoule control pulse shifts the polariton population between the lower and upper polariton branches, executing a NOT-equivalent operation. The energy figure is approximately nine orders of magnitude below a 2026-generation Intel 3 gate operating at equivalent switching speed.
The architectural implication is profound if the technology can be scaled. Conventional photonic interconnects are passive data conduits; they shuttle bits between electronic processing nodes but perform no logic. A room-temperature polariton switching array would enable a topologically different AI accelerator: one in which matrix multiplications and nonlinear activation functions are computed entirely in the optical domain, eliminating the analog-to-digital conversion overhead that currently consumes an estimated 20-30 percent of inference energy in state-of-the-art ASICs. The Penn group's theoretical modeling projects that a 10,000-element polariton logic array, if achievable at 4 femtojoules per switch, would consume roughly 0.04 milliwatts at 100 GHz operation - a figure that makes even the most aggressive silicon optimization look profligate.
The honest engineering assessment is that the current device is a proof of concept rather than a product. The MoSe2 monolayer exhibits non-radiative recombination rates that constrain polariton lifetimes to ~2 picoseconds at 295 K, limiting cascaded logic depth to approximately ten gates before signal refreshing is required; chemical vapor deposition of device-grade MoSe2 at 300 mm wafer scale remains a nascent capability; and the planar nanocavity geometry is incompatible with three-dimensional packaging standards used for HBM-stacked AI memory. IBM Research and imec have independently assessed the commercial horizon at five to seven years, with the most likely near-term deployment being edge inference accelerators in energy-constrained environments rather than hyperscale GPU cluster replacements.
Researchers at the University of Pennsylvania have created exciton-polaritons - hybrid particles that fuse a photon's speed with an electron's ability to interact - capable of switching signals using only four femtojoules of energy, roughly one billion times less than a conventional silicon transistor. The breakthrough, published in Physical Review Letters on April 8, 2026, uses an atomically thin layer of molybdenum diselenide inside a nanoscale optical cavity and demonstrates all-optical computing without cryogenic cooling. If scaled successfully, polariton-based photonic chips could dramatically cut the electricity consumption of large AI data centers.
Scientists at a university in the United States made a new kind of tiny particle. This particle is made from light and matter mixed together. Scientists call it an exciton-polariton.
These new particles can do the work that electricity usually does in computers. They can switch signals on and off very fast. And they use much, much less energy to do it.
Computers that use AI need a huge amount of electricity. Electricity costs money and can hurt the environment. If we can use light instead of electricity, AI could use less energy.
The scientists published their discovery in a science journal. Other scientists are excited about this new idea. It could change the way computers work in the future.
1What are the new particles made from?
2What do these particles do inside a computer?
3Where did the scientists work?
4Why is this discovery important for AI?
5What are these new light-matter particles called?
6The new particles are made from light and matter.
7These particles use more energy than normal computer chips.
8AI computers need very little electricity.
9The scientists published their work in a science journal.
10This discovery could help AI use less energy.
11The new light-matter particles are called exciton-_____.
12These particles use much less _____ than normal computer chips.
13The scientists work at the University of _____.