Penn Researchers Create Light-Matter Chip That Could Slash AI Energy Use

--- title: "Penn Researchers Create Light-Matter Chip That Could Slash AI Energy Use" slug: penn-light-matter-chip-ai-energy category: research story_number: "10" date: 2026-05-24 sources: - name: "Penn Today" url: "https://penntoday.upenn.edu/news/making-light-work-computing" domain: penntoday.upenn.edu - name: "ScienceDaily" url: "https://www.sciencedaily.com/releases/2026/05/260518041341.htm" domain: sciencedaily.com - name: "SciTechDaily" url: "https://scitechdaily.com/light-matter-particles-could-revolutionize-ai-computing/" domain: scitechdaily.com - name: "Physical Review Letters" url: "https://journals.aps.org/prl/abstract/10.1103/gc15-qsvf" domain: journals.aps.org ---

# Penn Researchers Create Light-Matter Chip That Could Slash AI Energy Use

The University of Pennsylvania's physics department has a storied history with computing. Eighty years ago, Penn researchers J. Presper Eckert and John Mauchly unveiled ENIAC, the world's first general-purpose electronic computer, and launched the modern computing era on the back of electrons. Now, a new generation of Penn physicists is betting that electrons have nearly run their course - and that light may be the answer to AI's mounting energy crisis.

A team led by physicist Bo Zhen, the Jin K. Lee Presidential Associate Professor in Penn's Department of Physics and Astronomy, has created a quasiparticle that could make AI chips dramatically more efficient. Publishing their findings April 8 in Physical Review Letters, Zhen and colleagues demonstrated a new class of hybrid light-matter particles - called exciton-polaritons - capable of performing the kind of signal switching that lies at the heart of modern computation. The switching consumed approximately 4 quadrillionths of a joule of energy, an amount so vanishingly small it dwarfs the energy needed to momentarily illuminate a tiny LED bulb. The team says this sets a new benchmark for switching energy in two-dimensional exciton-polariton systems.

The Problem With Electrons - and Photons

Today's AI systems devour electricity at rates that are straining power grids, rattling data center operators, and catching the attention of energy regulators. The culprit, in large part, is the silicon transistor. Electrons carry an electrical charge, which means every time they move through a chip they generate resistance and shed heat. The more complex the calculation, and the more transistors packed onto a chip, the worse that heat problem becomes. For AI workloads that require processing billions of parameters, the energy costs are enormous.

Photons - the massless, chargeless particles that make up light - look like an obvious solution on paper. They already carry the world's data across fiber-optic cables with minimal loss, and they travel at the speed of light. But, as Li He, co-first author of the study and a former postdoctoral researcher in the Zhen Lab, explained: "Because they are charge-neutral and have zero rest mass, photons can carry information quickly over long distances with minimal loss, dominating communications technology. But that neutrality means they barely interact with their environment, making them bad at the sort of signal-switching logic that computers depend on."

In other words, light is fast but passive. Computing requires not just data transmission but decision-making - applying nonlinear activation functions, switching signals, executing logic gates. These operations demand particles that interact with their environment, something photons fundamentally resist doing.

Exciton-Polaritons: Splitting the Difference

The Penn team's approach is elegantly simple in concept even if fiendishly difficult in practice. By coupling photons with excitons - bound pairs of electrons and holes inside an atomically thin semiconductor material - they created hybrid quasiparticles that borrow properties from both worlds. The resulting exciton-polaritons carry light's speed while inheriting matter's ability to interact and switch signals.

In their experiment, light was coupled into a nanoscale cavity and made to interact with an atomically thin semiconductor layer. The resulting quasiparticles could perform all-optical signal switching without the system ever needing to convert light into an electrical current and back again. That continuous conversion - light to electrons, electrons back to light - is exactly what makes current photonic AI chips lose much of their efficiency advantage. Every translation step costs time and energy.

The significance goes beyond raw energy numbers. Many photonic AI chips already exist in research settings and can perform certain straightforward matrix calculations at high speed using light. But when these systems reach nonlinear activation steps - the "decision-making" operations that give neural networks their power - they currently fall back on electronic circuits. Those fallbacks erode the speed and energy benefits that made photonic computing attractive in the first place. Exciton-polaritons could, in principle, eliminate that bottleneck, enabling chips to remain in the optical domain throughout computation.

From Lab Bench to AI Data Center

The research team is careful to frame this as a platform demonstration rather than a finished product. Scaling exciton-polariton systems from nanoscale laboratory cavities to production chips involves substantial engineering challenges. Atomically thin semiconductor materials - in this case, monolayer semiconductors whose properties can be tuned with a gate voltage - are extraordinarily delicate and difficult to manufacture at commercial scale.

Still, the implications if the technology can be scaled are significant. Zhen's group envisions photonic chips that could process information directly from cameras without ever converting optical signals into electronic ones, eliminating the latency and energy penalty that conversion imposes. Applied to AI inference hardware - the chips that run trained models in products like chatbots, image recognition systems, and autonomous vehicles - even a partial shift to optical processing could yield meaningful efficiency gains across an industry that consumed an estimated 1Ð2% of global electricity in 2025, a figure analysts expect to grow sharply through the rest of the decade.

Beyond AI, the research opens a door toward quantum computing on chips. Exciton-polaritons exhibit quantum properties, and the Penn team notes that the same platform might one day support basic quantum logic operations - a potential path toward integrating classical and quantum computation on a single photonic device.

The research was funded by the US Office of Naval Research and the Alfred P. Sloan Foundation. Li He, formerly of the Zhen Lab, is now an assistant professor at Montana State University. Additional co-authors include Zhi Wang and Bumho Kim of Penn's School of Arts & Sciences.

Whether or not exciton-polaritons ultimately power a new generation of AI hardware, the Penn work underscores a broader truth about the AI energy problem: incremental improvements to silicon transistors are no longer sufficient. The industry is running out of runway on the architecture that Eckert and Mauchly first set in motion eight decades ago - and researchers are increasingly convinced that the next leap will have to come from physics, not just engineering.