Penn Physicists Create Hybrid Light-Matter Particles That Could Power Faster, Greener AI Chips

Elementary

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.

exciton-polariton: a hybrid quantum particle formed by the combination of a photon of light with an electron-hole pair in a material

femtojoule: an extremely tiny unit of energy equal to one quadrillionth of a joule

semiconductor: a material that can conduct electricity under certain conditions, used to make computer chips

cavity: a very small chamber designed to trap and interact with light

photon: the smallest unit of light, which carries electromagnetic energy

optical: relating to light or the sense of vision

efficiency: using the smallest possible amount of energy or resources to achieve a result

quantum: relating to the smallest possible units of energy or matter in physics

Intermediate

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.

quasi-particle: a concept used in physics to describe collective behavior in matter that acts like a single particle

monolayer: a single layer of atoms, just one atom thick

resonance state: a condition in which a quantum particle oscillates at its natural frequency

logic gate: a basic switching element in a computer that produces an output based on one or more inputs

hybridize: to combine two different things to form a new entity with properties of both

photonic: relating to the use of light particles to transmit or process information

wafer scale: large enough to be manufactured on an industrial semiconductor wafer, the standard production format for chips

terawatt-hour: a unit of electrical energy equal to one trillion watt-hours, used to measure the consumption of large power systems

Advanced

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.

Dennard scaling: the historical principle that as transistors shrink, their power density stays constant; it broke down around 2006 and no longer holds

vacuum Rabi splitting: the energy gap that appears between two quantum states when a photon cavity and a material excitation reach strong coupling

Coulomb-exchange nonlinearity: an optical switching effect in polariton systems arising from electrostatic repulsion between the excitonic components of colliding polaritons

analog-to-digital conversion: the process of transforming a continuous physical signal (such as light) into digital numbers that a processor can use

Q-factor: quality factor; a dimensionless number measuring how well a resonator stores energy relative to energy lost per cycle

non-radiative recombination: a process in a semiconductor where an electron and hole combine and release energy as heat rather than as light, limiting device efficiency

proof of concept: an experimental demonstration that a principle or idea works, without being ready for commercial use

ASIC: Application-Specific Integrated Circuit; a chip custom-designed for a single task such as AI inference or cryptocurrency mining

Beginner

Elementary

Intermediate

Advanced

Penn Physicists Create Hybrid Light-Matter Particles That Could Power Faster, Greener AI Chips

Multiple Choice

True or False

Fill in the Blank