Penn scientists build light-matter particles for ultra-low-power AI

Eighty years after building ENIAC, the world’s first general-purpose electronic computer, physicists at the University of Pennsylvania are trying to pull off another hardware revolution. This time, they want to ditch electrons entirely.

As AI models balloon in size, traditional chips are hitting hard physical limits. Moving electrons creates heat, resistance, and massive power bills for data centers.

Now, a Penn team led by Bo Zhen just demonstrated an all-light switch powered by hybrid particles called exciton-polaritons, publishing their findings in Physical Review Letters.

Giving light the power to interact

Photons are fast and generate practically zero heat because they lack mass and charge. That is why fiber optics dominate modern communications.

Trouble is, photons are too passive. When two beams of light cross, they pass right through each other without interacting. That makes them terrible at building the basic logic gates and switches computers rely on.

To fix this, the Penn researchers embedded an atomically thin semiconductor inside a nanoscale optical cavity. When trapped photons couple with electrons in the material, they form quasiparticles called exciton-polaritons.

Basically, these hybrid particles retain the lightning speed of light while borrowing the interactive qualities of solid matter.

Using this setup, the team achieved all-light switching at an energy scale of just four quadrillionths of a joule. That is an extraordinarily tiny sip of power, far less than what it takes to flick on a tiny LED for a millisecond.

Killing the optical-electrical bottleneck

Current photonic AI chips already use light for certain heavy calculations, like matrix multiplication. But they hit a wall whenever an algorithm needs nonlinear activation or decision-making.

Today, chips solve that by constantly converting optical signals into electrical ones to do the logic, then turning them back into light. These round trips slow down the system and waste enormous amounts of energy.

By enabling switching directly within the optical domain, exciton-polaritons could let future chips run entire neural networks in pure light. In theory, photonic processors could even process raw light straight from a camera sensor without touching electricity.

Why it matters

Do not expect to slot an all-light chip into your workstation next week. This is fundamental physics, not an off-the-shelf product.

Yet for an industry terrified of running out of power, this experiment points to a viable escape hatch. Data center energy demands are already colliding with municipal power grids, and conventional silicon scaling is running out of steam.

We think the clever part here is the hybrid approach. Instead of trying to force reluctant photons to interact with each other, Zhen’s lab essentially gave light a temporary physical body through two-dimensional materials.

The real test will be manufacturability. Aligning atomic layers and nanoscale cavities at commercial scale across millions of components without signal drift is a massive engineering hurdle. Still, proving that all-optical switching can work at ultra-low energy gives the next generation of AI hardware a very real target to chase.